        

Risk Analysis of the NJH Data Set from the Beryllium Machining Facility in Cullman, Alabama  -  CBD and Sensitization
                                        
                                       
1.  Introduction/overview 
	OSHA's risk assessment for beryllium sensitization and Chronic Beryllium Disease (CBD) was developed to provide the Agency with information pertinent to the beryllium rulemaking, including: 
 A characterization of the risks of beryllium sensitization and CBD for workers exposed to airborne beryllium at the level of the preceding time-weighted average permissible exposure limit (TWA PEL); and
 A characterization of the reduction in risk that lowering the TWA PEL would be likely to achieve.
The risk assessment for the beryllium rule incorporates both qualitative and quantitative analyses to characterize these risks, including (1) a review of the published literature on recent beryllium exposure control programs, and (2) analyses of a data set provided by National Jewish Health (NJH) on a population of workers employed at a beryllium machining plant in Cullman, Alabama.  In section VI of the preamble, OSHA presents its final review of the published literature as well as its final analyses of the NJH data set. 
      This document supplements the material presented in section VI regarding OSHA's final prevalence and statistical analyses of the NJH data set.  It contains details regarding the data set itself and OSHA's preliminary and final analyses of it.  OSHA's statistical analysis was updated in response to comments received on the preliminary risk assessment from Dr. Kenny Crump and Ms. Deborah Proctor (Document ID 1660) and NIOSH (Document ID 1725), which are addressed in this document as well.  A preliminary version of OSHA's analyses of the NJH data set was presented in the NPRM and a technical document (Risk Analysis of a Worker Population at a Beryllium Machining Facility, OSHA, 2014, Document ID 1589).   
2.  Background on the Cullman, AL machining plant 
	NJH researchers, with consent and information provided by the facility, compiled a data set containing employee work histories, medical diagnoses, and air sampling results, and provided it to OSHA for analysis.  OSHA's contractor, Eastern Research Group (ERG), gathered additional information: for example, ERG conducted surveys of the Cullman plant (ERG, 2003, Document ID 1876; ERG, 2004a (1877)) and also collected published articles of independent investigations conducted at the plant (Kelleher et al., 2001, Document ID 1363; Madl et al., 2007 (1056); Martyny et al., 2000 (1053); and Newman et al., 2001 (1354)). 
	The Cullman, AL, machining plant is a fabricator of precision machined and processed materials, including beryllium and its alloys, titanium, aluminum, quartz, and glass. The company was founded in Michigan in 1947 and began working with beryllium in the mid-1950s (Cullman Chamber of Commerce, 2003, Document ID OSHA-H005B-2006-0898-0631).  The available information indicates the company moved its operations to Alabama and operated at least two facilities in Cullman, Alabama (Cullman Chamber of Commerce, 2003, Document ID 0631; OSHA, 2014 (1589)).  The current facility has been in operation since 1969 (Martyny et al., 2000, Document ID 1053; Kelleher et al., 2001 (1363); Madl et al., 2007 (1056)).   Machining operations in the old Cullman facility were consolidated into the company's current Cullman location in 1974 (see OSHA, 2014, Document ID 1589). 
	Production operations at the plant include receipt and inspection of materials, machining and polishing, chemical finishing (acid etching), and inspection and quality assurance.  Machining operations include milling (manual and CNC milling), lathe work, lapping, grinding, deburring (grinding with a small hand-operated tool (Martyny et al., 2000, Document ID 1053)), and electrical discharge machining (EDM).  Management and administrative personnel include two groups: those primarily working in the front offices (front administration) and those primarily working on the shop floor (shop management).  Most of the plant's beryllium operations involve machining beryllium metal and high beryllium content composite materials (beryllium metal/beryllium oxide metal composites called E-Metal or E-Material), with occasional machining of beryllium oxide/metal matrix (such as AlBeMet, aluminum beryllium matrix) and beryllium-containing alloys (Madl et al. 2007, Document ID 1056).

Exposure Conditions Over Time       
	Plant conditions over time and source material for this information are summarized in Table 1.  Prior to 1995, exposure controls for machining operations primarily included a low volume/high velocity (LVHV) central exhaust system with operator adjusted exhaust pickups and wet machining methods.  Madl et al. reported that the company prioritized use of engineering and administrative controls to limit workers' beryllium exposures rather than respiratory protection and other PPE (2007, Document ID 1056).  Similarly, OSHA found inconsistent use of work uniforms, gloves, and respiratory protection (ERG, 2004a, Document ID 1877).  
	Prior to the diagnosis of a case of CBD in 1995, the company's exposure sampling and control strategies focused on maintaining compliance with OSHA's preceding TWA PEL of 2 μg/m[3].  After the diagnosis, the facility established an in-house target exposure level of 0.2 μg/m[3], installed change/locker rooms for workers entering the production facility (one for production workers and one for office/administrative workers), eliminated pressurized air hoses and discouraged dry sweeping to reduce re-entrainment of beryllium particles in the air, initiated biennial medical surveillance using the Beryllium Lymphocyte Proliferation Test (BeLPT), and implemented annual beryllium hazard awareness training.
	In 1996, the company instituted a formal policy requiring the use of work uniforms and dedicated work shoes for production workers, eliminated dry sweeping in all departments, and purchased high-efficiency particulate air (HEPA) filter vacuum cleaners for workplace cleanup and decontamination.  The company also initiated major engineering changes that year, including the purchase of a new local exhaust ventilation (LEV) system to exhaust machining operations producing finer particles (i.e., dust as opposed to metal chips).  The facility also began installing mist eliminators for each machine.  Dry machining operations producing chips were exhausted using the existing LVHV exhaust system (ERG, 2004a, Document ID 1877).  In the course of making the ventilation system changes, old ductwork and baghouses were dismantled and new ductwork and air-cleaning devices were installed.  The company also installed Plexiglas enclosures on machining operations, including lapping, deburring, grinding, EDM, and tool and die operations, between 1996 and 1997.  In 1998, LEV was installed in EDM and modified in the lap, deburr, and grind departments. 
	As shown in Table 1, most exposure controls were in place by 2000.  In 2003, the company replaced the LVHV ductwork in the milling and lathe departments (ERG, 2004a, Document ID 1877).  Over time, the facility has built enclosures for operations that consistently produce exposures greater than 0.2 μg/m[3].  In 2004, two surface grinders were fully enclosed and ventilated (ERG, 2004a, Document ID 1877).  That year, the plant also opened a new 20,000-square-foot manufacturing facility to produce the unpolished beryllium mirror segments for the James Webb Space Telescope, the nation's next-generation space telescope and successor to the Hubble.  The new facility houses advanced computer-aided manufacturing and metrology equipment with state-of-the-art ventilation controls, full equipment enclosures, and smooth, easily cleanable surfaces (Marcial and Martyny, 2007, Document ID 1121). 

            Table 1:  Exposure Conditions at the Cullman, AL Plant
                                       
Events
                                     Dates
References

Workplace practices reported to be relatively stable. Exposure control strategy included wet methods and LVHV exhaust system (i.e., Spencer system with baghouses). Inconsistent use of work uniforms by employees. 
                                       
                                 Prior to 1995

ERG, 2004a; 
Kelleher et al., 2001; OSHA, 2014, Document ID 1589

Identification of one worker with CBD (index case).


                                       
                                     1995
                                       

Newman et al., 2001

Established in-house target exposure level of 0.2 μg/m[3] (8-hour TWA).
                                       
                                1995 - Present

Marcial and Martyny, 2007

Installed change/locker room for production workers at rear entrance to facility with a clean side and a dirty side. Installed change room for office/administrative workers entering plant; lab coats and shoe covers required to enter machining area.    
                                       
                                     1995

OSHA, 2014, Document ID 1589

Initiated practice of purchasing new machines with enclosures (i.e., when old, open equipment needs to be replaced, it is replaced with new, enclosed equipment).  
                                       
                                     1995
                                       

Marcial and Martyny, 2007

Initiated annual beryllium hazard awareness training for workers. 
                                       
                                     1995 

Marcial and Martyny, 2007

Removed pressurized air hoses used for blowing beryllium dust off machines and discouraged dry sweeping. 
 
                                       
                                     1995

OSHA, 2014, Document ID 1589

Eliminated dry sweeping in all departments. Purchased approximately 10 high efficiency particulate air filter (HEPA) vacuums for workplace cleanup and decontamination.
                                       
                                     1996

Madl et al., 2007;
Newman et al., 2001; OSHA, 2014, Document ID 1589

Installed 5 Donaldson Torit LEV systems for dust/fume/mist-producing operations. Departments exhausted by Donaldson system include cutter grind (tool and die), E-cell, EDM, flow lines, grind, lapping, and optics. Some Spencer baghouses and ductwork eliminated (old Spencer ductwork dismantled; new Donaldson ductwork installed).  
                                       
                                     1996

OSHA, 2014, Document ID 1589

Began installing Plexiglas enclosures on machining operations; started with deburring.
                                       
                                     1996

Kelleher et al., 2001; OSHA, 2014, Document ID 1589

Installed enclosures on lapping, deburring, grinding, EDM, and tool and die operations.
                                       
                                  1996 - 1997

Madl et al., 2007

Installed mist eliminators on some CNC milling, EDM, grinding, lapping, and cutter grinding (tool and die) processes. 

[Note: Mist eliminators not used for machining operations producing chips. Some machines exhausted to large outdoor mist eliminators that handle multiple pieces of equipment; other machines vented to individual mist eliminators located inside the machine shop.]
                                       
                                  1996 - 1997

Madl et al., 2007;
OSHA, 2014, Document ID 1589

Around 1996, the company implemented a formal policy requiring mandatory use of company uniforms and dedicated work shoes. Policy has been included in employees' manual.

[Note: Employee uniforms have been available at least since 1973 (NIOSH, 1973).]
                                       
                                  1996 - 1997

Kelleher et al., 2001; 
Madl et al., 2007; OSHA, 2014, Document ID 1589

Workers must wash their hands, forearms, and face before breaks/lunch and after the shift. Personal hygiene is included with employee training.
                                       
                              At least since 1996

Marcial and Martyny, 2007;
OSHA, 2014, Document ID 1589

Installed LEV in EDM and updated it in the lap, deburr, and grind departments.
                                       
                                     1998

Kelleher et al, 2001;
Madl et al, 2007

Replaced LEV ductwork in the mill and lathe departments.
                                       
                                     2003

ERG, 2004a

Wearing long sleeves becomes mandatory. 

[Note: Initially workers wore disposable sleeves, now workers wear long sleeve shop coats with elastic cuffs. Glove use not mandatory; however, wearing gloves is emphasized in training and workers are encouraged to wear gloves whenever possible or feasible.]
                                       
                                  2004 - 2005

Marcial and Martyny, 2007; OSHA, 2014, Document ID 1589

All machines have exhaust ventilation and about 65% of machines are also enclosed with either partial or full enclosures to control the escape of machining coolant.
                                       
                                     2004

ERG, 2004a;
OSHA, 2014, Document ID 1589

Installed full LEV enclosures on two surface grinder operations. Surface grinder operator exposure reportedly reduced from an average of 1.6 μg/m[3] to 0.08 μg/m[3] (8-hr TWAs).
                                       
                                     2004

ERG, 2004a

Description of the Cullman, AL, Data Provided by NJH 
      NJH researchers compiled employee work histories, medical diagnoses, and air sampling results for the machining plant, and provided these data to OSHA. 
Work History
      The work-history data set contains records for 348 workers and includes fields for each worker's job(s), start date, time spent at each job, and percentage of work time spent in each job for workers who held two or more jobs concurrently.   
	Machining job titles and departments in the NJH data set include the following: cathode (cath), deburr, E-Cell, EDM, flow lines (flow), gas bearings (gas), glass (non-beryllium machining), grind, HRG cell (non-beryllium machining), James Webb Space Telescope (JWST), lap, lathe, mill, optics, QRoom, specialty cell worker (spec), and tool (non-beryllium machining).  The job title "tool" can also be known as tooling, tool making, tool grinding, tool and die, and cutter grinding (or cgrinding); it refers to workers that repair, maintain, fabricate, and modify machine-cutting tools. Workers can be exposed to beryllium when they perform servicing activities on cutting tools that are contaminated with or contain beryllium.  Non-machining job titles or departments include (but are not limited to) administrator, assembler, custodian, chemical finisher, engineer, estimator, health and safety, ion milling, inspector, maintenance, manager, numerical control (NC) programmer/manufacture programming, product control/inventory control, sales, shipper, and shop management (shop floor).  Table 2 summarizes available information regarding job title descriptions of interest.

             Table 2. Beryllium Metal Machining Plant -- Job Titles
                                       
                                    Job ID
                                Job Title Code
                                   Comments
                                      11
Lathe
Machining. A machine tool for shaping metal or wood; the workpiece (block of material) turns or spins about a horizontal axis against a fixed single-edge cutting tool. The lathe can perform turning, facing, threading, drilling, and other operations. 
                                      12
Lapping
Machining. An abrasive machining operation where two surfaces are rubbed together with an abrasive between them to remove unwanted material. Lapping may be done by hand or by machine.  
                                      13
Deburring
Machining. Removing burrs from manufactured parts. A burr is a thin ridge or area of roughness produced by cutting or shaping metal. Deburring may be done by hand with hand-deburring tools or with automated deburring machines. 
                                      14
Grinding
Machining. An abrasive machining process to remove material from the workpiece. A grinding machine is a machine tool for producing very fine finishes, making very light cuts, or high precision forms using an abrasive wheel as the cutting device. The simplest grinder is a bench grinder or a hand-held angle grinder for deburring parts or cutting metal with a zip-disc.   
                                      15
Milling
Machining. Milling is the complex shaping of parts (metal or possibly other materials) by removing unneeded material to form the final shape. Milling is done on a milling machine with a rotating multi-edge cutting tool; the workpiece is fixed. Milling machines may be operated manually or under computer numerical control (CNC) and can perform many complex operations such as slot cutting, planing, drilling and threading, rabetting, routing, etc.   
                                      16
NC Programmer
Production support/engineering. Numerical control (NC) programming. Providing manufacturing programming support for numerically controlled machine tools. Job activities could include analyzing engineering drawings for parts, defining and creating programs for machining work centers (e.g., selecting proper cutter and program routines for efficient computer calculation of dimensional and machining requirements for parts), interfacing with shop personnel, and other machine shop activities.  
                                      17
Optics Department
Machining beryllium. Primarily polishing and lapping small parts; some hand operations. [Note: The ERG Site 4 and 9 survey reports indicate that the optics department specializes in making high tolerance mirrors (nickel on a beryllium substrate) and that beryllium parts are also lapped (machined) there.]  
                                      18
HRG Cell
Non-beryllium (glass/quartz) machining. HRG likely refers to hemispherical resonator gyroscope. The James Webb Space Telescope will have HRGs to sense the orientation of the telescope. HRGs are constructed with three pieces of machined quartz. [Note: During the ERG Site 4 survey, beryllium wipe samples were collected in the quartz work area which does not use beryllium but is adjacent to areas that process beryllium-containing parts.]  
                                      19
Tooling
Non-beryllium machining/production support/toolmaking. Tooling refers to workers that repair, maintain, fabricate, and modify machine-cutting tools. Per Kelleher et al. (2001) tool grinders do not machine beryllium. 
                                      20
Clean Room
Non-machining. Refers to an area where precise inspection operations are performed.
                                      21
Assembly
Production support/mechanical assembly. Refers to mechanical assembly of beryllium and non-beryllium parts. Can involve tapping holes and inserting pins. Assembler is a major non-machinist job title per Kelleher et al. (2001).
                                      22
Inspection
Production support/inspection. Inspector is a major non-machinist job title per Kelleher et al. (2001).
                                      23
EDM
Machining/electrical discharge machining (EDM). Machining electrically conductive materials by using precisely controlled sparks that occur between an electrode (cutting tool) and a workpiece in the presence of a dielectric fluid. EDM is a thermal process; material is removed by heat. The material between the electrode and workpiece is heated to the point where the material vaporizes.
                                      24
Chemical Finishing
Production support/chemical finishing or acid etching. Chemical finisher is a major non-machinist job title per Kelleher et al. (2001). 
                                      25
Production Control
Non-machining per Madl et al. (2007); management/administrative. Manager and administrator are major non-machinist job titles per Kelleher et al. (2001). Production control typically refers to the systematic planning, coordinating, and directing of all manufacturing activities and influences to insure having products made on time, of adequate quality, and at reasonable cost. 
                                      26
Maintenance
Non-mailing per Madl et al. (2007). Production support per Kelleher et al. (2001). 
                                      27
Estimating
Non-machining. Production support/engineering. This job title code refers to production estimator. Estimators analyze blueprints, specifications, proposals, and other documentation to prepare time, cost, and labor estimates for products, projects, or services, applying knowledge of specialized methodologies, techniques, principles, or processes. Consults with clients, vendors, or other individuals to discuss and formulate estimates and resolve issues. May specialize according to the particular service performed, type of product manufactured, or phase of work involved. 
                                      28
Engineering
Non-machining. Production support per Kelleher et al. (2001).
                                      29
Manufacture Programming
Production support/engineering. This may be referring to NC programming or something similar.
                                      30
Shop Management
Non-machining per Kelleher et al. (2001). Refers to management and administrative personnel that predominantly work on the shop floor (shop management).
                                      31
Salaried Sales
Non-machining per Kelleher et al. (2001). Production support/sales and administration.
                                      32
Front Administration
Non-machining per Kelleher et al. (2001) and Madl et al. (2007). Refers to management and administrative personnel that predominantly work in the front offices.
                                      33
Flow line
Machining. Primarily machine aluminum spindles; occasionally (infrequently) machine beryllium. The lean manufacturing line where a part is processed along an assembly line. [Note: Specialty cell workers ("Spec") are now categorized under flow lines. During the ERG Site 4 survey, a beryllium wipe sample was collected in the flow line area which is near the grind department and the quartz work area.
                                      34
Health and Safety Personnel
Non-machining. Kelleher et al. (2001) included health and safety with shop management.
                                      35
Specialty Cell/E-Cell/Cathode Cell
Machining. Specialty cell refers primarily to aluminum machining; glass, quartz, and beryllium are also machined in specialty cell. E-cell refers to E-cell lapping department or E-60 area. Cathode refers to cathode grinding (adjacent to lathes). 
                                      36
Contract Worker
Not enough information is provided to determine if this job title code pertains to machining.
                                      37
Gas Bearing
Machining per Madl et al. (2007). Per the ERG Site 4 and Site 9 survey reports, the gas bearings department uses beryllium.
                                      38
Glass
Non-beryllium machining. Machining (e.g., lapping, grinding, milling) optical glasses (infrared lenses). 
                                      55
Laid Off
May have referred to both machinists and non-machinists. During 1999/2000, the plant experienced a downturn in business and laid off 60 to 70 workers.
Sources: Kelleher et al. (2001); Madl et al. (2007); OSHA, 2014, Document ID 1589
Screening results  
      Plant-wide screening for beryllium sensitization using the BeLPT began in May 1995 and was conducted biannually for all employees (Newman et al., 2001, Document ID 1354).  Employees hired or rehired after the initial screening were tested within three months of hire and biannually thereafter.  Thus, employees working at the plant in May 1995 were eligible for follow-up testing in 1997, 1999, 2001, and so on, until they left employment or were found to be sensitized.  Employees receiving their first test in 1996 were eligible for follow-up testing in 1998, 2000, and so on.  In his 2001 publication on the surveillance program, Dr. Lee Newman reported 100 percent participation in follow-up testing among eligible employees (Newman et al., 2001, Document ID 1354).
     	The data set of sensitization screening and CBD diagnoses contains records for 363 workers and includes fields for birth year, year of hire, smoking status, and type of diagnosis.  All 348 workers included in the work-history data set are also represented in the diagnosis data set.  Of the 15 remaining workers in the diagnosis data set, 10 were initially hired in 2004 or 2005, three were first hired between 1997 and 2000, and the initial hire year was not given for two workers.  Seven workers are reported as sensitized only, 16 workers are listed as sensitized and diagnosed with CBD only, and three workers are shown as first sensitized and then diagnosed with CBD.  None of the workers who are absent from the work-history data set are reported as CBD or sensitization cases.
Industrial Hygiene Sampling Data 
      The NJH data set includes industrial hygiene air sampling results for the beryllium metal machining plant from 1980 to 2005.  The exposure data set includes air samples collected by the plant (1980 to 1984 and 1995 to 2005) and NJH (June 1996 to February 1997, and September 1999). 
      A total of 5,082 records comprise the results for 4,370 personal breathing zone samples and 712 area samples.  Sampling data include the sample date, flow rate, sample time, sample volume, sampler type (filter or impactor), sample type (breathing zone or area), and exposure value (in micrograms per cubic meter (μg/m[3])).  Sample times are provided for 2,557 of the personal breathing zone samples.  Sampling times for the plant-collected data are available for 94.8 percent of samples through 2003.  All samples for 2004 and 2005 are full-shift samples (approximately eight hours in duration) (OSHA, 2014, Document ID 1589).  All results in the database represent beryllium concentrations that are based on the actual sampling times (i.e., none of the results have been time-weighted for eight hours).  Table 3 below shows the total number of personal breathing zone samples collected for each process.
                    Table 3. Samples collected by Job Title
                                         
Process Name
                             Number of PBZ Samples
Admin
                                      171
Assy
                                      10
Cath
                                      30
Cgrind
                                      149
Chem
                                      212
Deburr
                                      156
Ecell
                                      118
Edm
                                      127
Eng
                                      258
Flow
                                      184
Gas
                                      186
Glass
                                      65
Grind
                                      148
Health and Safety
                                       6
Insp
                                      337
Jwst
                                      60
Lap
                                      246
Lathe
                                      386
Maint
                                      147
Mill
                                      926
Msupp
                                      219
Optics
                                      89
Pcic
                                      86
Qroom
                                       3
Shop
                                       4
Spec
                                      46
Tool
                                       1

                                       
All processes
                                     4,370

Construction of Job-Exposure Matrix
	Using the industrial hygiene data set described above, ERG developed a job-exposure matrix (JEM) for the beryllium machining facility.  Of the samples in the database for which sample times were recorded, 97.9 percent (2,503 out of 2,557) were personal breathing zone samples with sample times greater than or equal to 400 minutes.  No adjustments were made for sampling time, except in the case of four samples for the "maintenance" process for 1995.  These results show relatively high values and exceptionally short sample times.  The four 1995 maintenance samples were adjusted for an eight-hour sampling time assuming that the maintenance workers received no further beryllium exposure over the rest of their work shift.  
	After reviewing the sample statistics for individual years for different groupings, ERG evaluated the following year groupings to construct the exposure history of workers contained in the work-history database:
Primary Machining Jobs (see Table 4) - 1980 to 1995; 1996 to 1997; 1998 to 1999; 2000 to 2003
Other Jobs (see Table 5) - 1980 to 2003
	The time periods selected for primary machining jobs correspond to stages in the plant's approach to beryllium exposure control described in the previous section.  The first period of exposure sampling, 1980 to 1995, represents a period of relatively minimal control prior to the 1995 discovery of a case of CBD among the plant's workers.  In 1996 and 1997, the company began installing major engineering controls.  By 1999, many such engineering controls had been installed, and exposures were generally lower than in previous periods.  The plant completed installation of most exposure controls by 2000, and exposures were very low throughout the plant in the time period from 2000 to 2005 (mean sample values for most jobs < 0.15 μg/m[3]).  Tables 4 and 5 below present mean exposure values for machining and non-machining job categories using all samples in the data set, by time period.  
                                       
         Table 4. Exposure Values for Machining Job Titles (μg/m[3])
                                       
                                   Job Title
                                  1980 - 1995
                                  1996 - 1997
                                  1998 - 1999
                                  2000 - 2003
                                       
                                    samples
                                     mean
                                    samples
                                     mean
                                    samples
                                     mean
                                    samples
                                     mean
Deburring
                                                                             27
                                                                           1.17
                                                                             20
                                                                           3.30
                                                                              0
NA
                                                                             67
                                                                           0.16
Electrical Discharge Machining
                                                                              2
                                                                           0.06
                                                                              2
                                                                           1.32
                                                                             16
                                                                           0.08
                                                                             63
                                                                           0.12
Grinding 
                                                                             13
                                                                           8.48
                                                                              6
                                                                           0.49
                                                                             15
                                                                           0.24
                                                                             68
                                                                           0.15
Lapping
                                                                              9
                                                                           0.15
                                                                             16
                                                                           0.24
                                                                             42
                                                                           0.21
                                                                            103
                                                                           0.16
Lathe
                                                                             19
                                                                          13.99
                                                                              8
                                                                           1.13
                                                                             40
                                                                           0.17
                                                                            200
                                                                           0.10
Milling
                                                                             43
                                                                           0.64
                                                                             15
                                                                           0.23
                                                                             95
                                                                           0.17
                                                                            434
                                                                           0.12

       Table 5. Exposure Values for Non-Machining Job Titles (μg/m[3])
                                       
                                   Job Title
                                  1980 - 1995
                                  1996 - 1997
                                  1998 - 1999
                                2000  -  20033
                                       
                                    samples
                                     mean
                                    samples
                                     mean
                                    samples
                                     mean
                                    samples
                                     mean
Administration
                                                                              0
                                                                             NA
                                                                              0
                                                                             NA
                                                                             39
                                                                          0.052
                                                                             74
                                                                          0.061
Assembly
                                                                              0
                                                                             NA
                                                                              0
                                                                             NA
                                                                              8
                                                                          0.136
                                                                              2
                                                                          0.051
Cathode
                                                                              0
                                                                             NA
                                                                              0
                                                                             NA
                                                                              0
                                                                             NA
                                                                              9
                                                                          0.156
Cgrind
                                                                              1
                                                                          0.120
                                                                              0
                                                                             NA
                                                                             14
                                                                          0.105
                                                                             76
                                                                          0.112
Chem
                                                                              0
                                                                             NA
                                                                              2
                                                                         12.731
                                                                             21
                                                                          0.277
                                                                             91
                                                                          0.152
Ecell
                                                                              0
                                                                             NA
                                                                             13
                                                                          1.873
                                                                              0
                                                                             NA
                                                                             26
                                                                          0.239
Engineering
                                                                              1
                                                                          0.065
                                                                              0
                                                                             NA
                                                                             49
                                                                          0.069
                                                                            125
                                                                          0.062
Flow Lines
                                                                              0
                                                                             NA
                                                                              0
                                                                             NA
                                                                              0
                                                                             NA
                                                                            113
                                                                          0.083
Gas
                                                                              0
                                                                             NA
                                                                              0
                                                                             NA
                                                                              0
                                                                             NA
                                                                            121
                                                                          0.058
Glass
                                                                              0
                                                                             NA
                                                                              0
                                                                             NA
                                                                              0
                                                                             NA
                                                                             38
                                                                          0.068
Health and Safety
                                                                              0
                                                                             NA
                                                                              0
                                                                             NA
                                                                              0
                                                                             NA
                                                                              5
                                                                          0.076
Inspection
                                                                              0
                                                                             NA
                                                                              0
                                                                             NA
                                                                             32
                                                                          0.101
                                                                            150
                                                                          0.066
Maintenance
                                                                              4
                                                                          1.257
                                                                              1
                                                                          0.160
                                                                             16
                                                                          0.200
                                                                             70
                                                                          0.126
Msupp
                                                                              0
                                                                             NA
                                                                              0
                                                                             NA
                                                                             47
                                                                          0.094
                                                                             68
                                                                          0.081
Optics
                                                                              0
                                                                             NA
                                                                              0
                                                                             NA
                                                                              0
                                                                             NA
                                                                             41
                                                                          0.090
PCIC
                                                                              1
                                                                          0.040
                                                                              0
                                                                             NA
                                                                             13
                                                                          0.071
                                                                             42
                                                                          0.083
Qroom
                                                                              1
                                                                          0.280
                                                                              0
                                                                             NA
                                                                              0
                                                                             NA
                                                                              2
                                                                          0.130
Shop
                                                                              0
                                                                             NA
                                                                              0
                                                                             NA
                                                                              4
                                                                          0.060
                                                                              0
                                                                             NA
Spec
                                                                              3
                                                                          0.247
                                                                              0
                                                                             NA
                                                                             24
                                                                          0.083
                                                                             19
                                                                          0.087
Tool
                                                                              0
                                                                             NA
                                                                              0
                                                                             NA
                                                                              0
                                                                             NA
                                                                              1
                                                                          0.070

      Samples for the early time periods are notably sparse, particularly for some non-machining jobs.  This does not necessarily imply a weakness in exposure characterization for all jobs listed, as some jobs did not appear in workers' job histories during the earlier time periods.  For example, workers with the job "Flow Lines" or "Glass" only worked in those jobs from 2000 onward, and workers with the job "Gas Bearings" only worked in that job from 1999 onward.  However, this is not the case for all such jobs; for example, 26 workers in the data set worked in "Inspection" prior to the first sampling period for that job.  
	ERG identified nine industrial hygiene samples that exceeded four times the 95[th] percentile value for the job title and time period in which they were collected.  These included samples in administration, chemical finishing, deburring, EDM, grinding, lapping, and lathe.  Upon further review of the data set and the processes in use at the plant, ERG judged that five of the nine outlier samples could plausibly have been accurate samples and should be retained in further calculations. ERG considered the remaining four outlier samples to be highly implausible, therefore likely to be sampling errors, and excluded them from further calculations.  The excluded sampling results are:
            Table 6. Erroneous Samples Excluded from the Analyses
                                     Year
                                     Date
                                    Process
                                    Sampler
                                  Sample Type
                                     Flow
                                     Time
                                    Volume
                                   μg/m[3]
                                                                           1997
                                                                      2/18/1997
CHEM
Filter
BZ
 
 
                                                                           2567
                                                                          24.93
                                                                           1996
                                                                      7/25/1996
DEBURR
Filter
BZ
 
 
                                                                           1326
                                                                          41.48
                                                                           1981
                                                                      8/12/1981
GRIND
Filter
BZ
1.88
                                                                            435
                                                                          817.8
                                                                          73.37
                                                                           1995
                                                                      5/10/1995
LATHE
Filter
BZ
2
                                                                             10
 
                                                                         250.00

                                     Year
                                     Date
                                    Process
                                    Sampler
                                  Sample Type
                                     Flow
                                     Time
                                    Volume
                                   μg/m[3]
                                                                           1997
                                                                      2/18/1997
CHEM
Filter
BZ
 
 
                                                                           2567
                                                                          24.93
                                                                           1996
                                                                      7/25/1996
DEBURR
Filter
BZ
 
 
                                                                           1326
                                                                          41.48
                                                                           1981
                                                                      8/12/1981
GRIND
Filter
BZ
1.88
                                                                            435
                                                                          817.8
                                                                          73.37
                                                                           1995
                                                                      5/10/1995
LATHE
Filter
BZ
2
                                                                             10
 
                                                                         250.00

				Following removal of these sampling results, the mean exposure sample values for these jobs were considerably reduced.  The final mean values were 0.53 for chemical finishing (1996-1997), 1.29 μg/m[3] for deburring (1996-1997), 3.07 for grinding (1980  -  1995), and 0.88 μg/m[3] for lathe (1980-1995), . 
      Reviewing the revised sample statistics for individual years for different groupings, OSHA noted that exposures in the 1996 to 1997 period for some machining jobs were equivalent to, or even higher than, exposure levels recording during the 1980 to 1995 period.  During 1996 to 1997, major engineering controls were being installed, but exposure levels were not yet consistently reduced.  The Agency therefore used the following year groupings to construct the exposure history of workers contained in the database who performed primary machining jobs: 1980 to 1997; 1998 to 1999; and 2000 to 2003.  Exposure values from the time period 1980 to 1997 were used to estimate exposures for machining jobs prior to 1980.  In general, the use of data collected in later time periods to characterize exposures in earlier years may lead to some underestimation of early exposures.  Exposure estimates are most reliable for the post-2000 period and, to a lesser extent, the 1998 to 1999 period.
	Some jobs -- including N/C Programming, HRG Cell, Manufacture Programming, and Salaried Sales -- did not have associated sample values.  Few (< 10) workers in the data set were employed in these jobs.  Based on information on the nature of each job (see Table 2), ERG used samples collected in Engineering to characterize exposures in N/C Programming and Manufacture Programming; samples in Front Administration to characterize Salaried Sales; and samples in Optics to characterize HRG Cell (OSHA, 2014, Document ID 1589).  

Development of Exposure Estimates for Workers
	The work-history data set included entries for 348 workers.  No job title was provided for eight workers in the data set, including three long-term workers (employed between 14 and 20 years), one shorter-term worker (5 years), and four workers with incomplete dates of employment.  None of these workers were found to be beryllium sensitized or diagnosed with CBD, and all eight workers were excluded from the analyses.  Seven of the workers in the data set were hired before 1969, the year reported in the published literature as the Cullman plant's first year of operations.  These workers, none of whom were found to be sensitized or diagnosed with CBD, were also excluded from the analysis.  An additional 14 workers were reported to have no exposure time.  These records may represent individuals who went through the hiring process but did not actually spend time at the plant, or they may reflect errors in the data set.  They were excluded from the analysis of sensitization and CBD prevalence in exposure quartiles described below, as it was not OSHA's intention to include unexposed individuals in the lowest quartile of exposure.  After these 29 exclusions, a total of 319 workers remained in the data set.
	The initial work-history data set includes values for the total time (including fractions of years) workers spent on each job recorded in the work-history database.  For confidentiality reasons, however, the exact initial start date (i.e., month and day) was not provided.  ERG imputed the amount of time in each year spent on a job by first assuming that an equal length of time was spent during the initial and final years of a job, and that full years were spent during the intervening years.  For example, for a work-history entry of 3 years in milling between 1997 and 2000, ERG assumed the worker was employed as a miller for 0.5 years in 1997, for 2 years in 1998 to 1999, and for 0.5 years in 2000.  
	Cumulative exposure was calculated as , where ei is the arithmetic mean sample level for job (i) (in the appropriate time period) and ti is the time spent in job (i).  Cumulative exposure was divided by total length of employment to estimate each worker's lifetime-weighted (LTW) average exposure.  For the prevalence and logistic regression analyses discussed in the next section, exposures were estimated at the end of the reported work history for each worker who was not beryllium sensitized or diagnosed with CBD.  For workers with beryllium sensitization or CBD, exposure estimates excluded exposures following diagnosis.  For three workers who were identified as sensitized and later diagnosed with CBD, exposure estimates excluded exposures after they were found to be sensitized.  For the proportional hazards regression analyses, exposures were estimated at multiple times throughout a worker's follow-up.  This exposure construction is described in more detail below.
	Workers employed for long time periods in jobs with low-level exposures tend to have low average and cumulative exposures due to the way these measures are constructed (i.e., incorporating the worker's entire work history).  It has been suggested that higher-level exposures, such as those encountered in machining jobs, may be more relevant to development of sensitization or CBD than exposures experienced in administrative or other low-exposure work (Madl et al., 2007, Document ID 1056).  To explore this possibility, ERG constructed a third type of exposure estimate.  This measure reflects the exposure level associated with the highest-exposure job (HEJ) and the JEM time period during which the worker was employed in that job.  For example, a worker employed in deburring for one year in 1989 and in administration for 10 years from 1990 to 2000 would be assigned the exposure level associated with deburring in the time period 1980 to 1997.  This exposure estimate (HEJ), the cumulative exposure estimate, and the average exposure were used in a prevalence analysis and in statistical analyses.

3. Prevalence of Beryllium Sensitization and CBD

	In the database provided to OSHA, seven workers were reported as sensitized only, 16 workers were listed as sensitized and diagnosed with CBD upon initial clinical evaluation, and three workers were shown as first sensitized only and later diagnosed with CBD.  Tables 7 to 9 below present the prevalence of sensitization and CBD cases across several categories of TWA, cumulative, and HEJ exposure.  Exposure values were grouped by quartile.  Because all workers with CBD are also sensitized, the columns "Total" and "Total %" refer to all sensitized workers in the data set, including those with CBD and those not diagnosed with CBD.  
 Table 7. Prevalence of Sensitization and CBD by LTW Average Exposure Quartile
                                               LTW Average Exposure  (μg/m[3])
                                                                     Group Size
                                                                Sensitized only
                                                                            CBD
                                                                          Total
                                                                        Total %
                                                                          CBD %
                                                                  0.0  -  0.080
                                                                             91
                                                                              1
                                                                              1
                                                                              2
                                                                           2.2%
                                                                           1.0%
                                                                   0.081 - 0.18
                                                                             73
                                                                              2
                                                                              4
                                                                              6
                                                                           8.2%
                                                                           5.5%
                                                                    0.19 - 0.51
                                                                             77
                                                                              0
                                                                              6
                                                                              6
                                                                           7.8%
                                                                           7.8%
                                                                    0.51 - 2.15
                                                                             78
                                                                              4
                                                                              8
                                                                             12
                                                                          15.4%
                                                                          10.3%
                                                                          Total
                                                                            319
                                                                              7
                                                                             19
                                                                             26
                                                                           8.2%
                                                                           6.0%

 Table 8. Prevalence of Sensitization and CBD by Cumulative Exposure Quartile
                                               Cumulative Exposure (μg/m3-yrs)
                                                                     Group Size
                                                                Sensitized only
                                                                            CBD
                                                                          Total
                                                                        Total %
                                                                          CBD %
                                                                  0.0  -  0.147
                                                                             81
                                                                              2
                                                                              2
                                                                              4
                                                                           4.9%
                                                                           2.5%
                                                                0.148  -  1.467
                                                                             79
                                                                              0
                                                                              2
                                                                              2
                                                                           2.5%
                                                                           2.5%
                                                                1.468  -  7.008
                                                                             79
                                                                              3
                                                                              8
                                                                             11
                                                                          13.9%
                                                                           8.0%
                                                                7.009  -  61.86
                                                                             80
                                                                              2
                                                                              7
                                                                              9
                                                                          11.3%
                                                                           8.8%
                                                                          Total
                                                                            319
                                                                              7
                                                                             19
                                                                             26
                                                                           8.2%
                                                                           6.0%

     Table 9. Prevalence of Sensitization and CBD by HEJ Exposure Quartile
                                                        HEJ Exposure (μg/m[3])
                                                                     Group Size
                                                                Sensitized only
                                                                            CBD
                                                                          Total
                                                                        Total %
                                                                          CBD %
                                                                    0.0 - 0.086
                                                                             86
                                                                              1
                                                                              0
                                                                              1
                                                                           1.2%
                                                                           0.0%
                                                                  0.091 - 0.214
                                                                             81
                                                                              1
                                                                              6
                                                                              7
                                                                           8.6%
                                                                           7.4%
                                                                  0.387 - 0.691
                                                                             76
                                                                              2
                                                                              9
                                                                             11
                                                                          14.5%
                                                                          11.8%
                                                                  0.954 - 2.213
                                                                             76
                                                                              3
                                                                              4
                                                                              7
                                                                           9.2%
                                                                           5.3%
                                                                          Total
                                                                            319
                                                                              7
                                                                             19
                                                                             26
                                                                           8.2%
                                                                           6.0%

	Table 7 shows increasing prevalence of total sensitization and CBD with increasing LTW average exposure.  The lowest prevalence of sensitization and CBD was observed among workers with LTW average exposure levels < 0.08 μg/m[3], where two sensitized workers were found (2.2%), including one case of CBD (1.0%).  The sensitized worker without CBD had worked at the facility as an inspector since 1972.  Because this job was believed to have very low exposures, it was not sampled prior to 1998.  Estimates of exposures in this job are based on data from 1998 to 2003.  It is possible that exposures earlier in this worker's employment history were higher than reflected in his estimated average exposure.  The worker diagnosed with CBD had been hired in 1996 in production control, and had an estimated average exposure of 0.08 μg/m[3].  He was diagnosed with CBD in 1997.  Based on samples from the 1998 to 1999 period and one sample from the 1980 to 1995 period (0.04 μg/m[3]), this job appeared to have very low exposures (1998 to 1999 mean = 0.07 μg/m[3], 95[th] percentile = 0.15 μg/m[3], max = 0.16 μg/m[3]).
      The second quartile of LTW average exposure (0.081 - 0.18 μg/m[3]) shows a marked rise in overall prevalence of beryllium-related health effects, with six workers sensitized (8.2%), of whom four were diagnosed with CBD (5.5%).  Three of the workers with CBD and one worker who was sensitized only had been hired before the mid-1990s, in periods with sparse exposure sampling.  The remaining two workers were hired in the 1998 to 1999 period and the 2000 to 2003 period, respectively, and had well-characterized LTW average exposures of 0.12 μg/m[3] and 0.18 μg/m[3], respectively.  Among six sensitized workers in the third quartile (0.19 - 0.51 μg/m[3]), all were diagnosed with CBD (7.8%).  Another increase in prevalence is seen from the third to the fourth quartile, with 12 cases of sensitization (15.4%), including 8 diagnosed with CBD (10.3%).
	When workers' HEJ exposures are considered, the exposure-response pattern is similar to that for LTW average exposure in the lower quartiles.  The lowest prevalence is observed in the first quartile (0.0 - 0.086 μg/m[3]), with increasing prevalence from first to second and second to third exposure quartiles.  The prevalence of sensitization and CBD in the top quartile (0.954 - 2.213 μg/m[3]) decreases relative to the third, with levels similar to the overall prevalence in the data set.
	The quartile analysis of cumulative exposure also shows generally, but not uniformly, increasing prevalence of sensitization and CBD with increasing exposure.  As shown in Table 8, the lowest prevalences of CBD and sensitization are in the first two quartiles of cumulative exposure (0.0 - 0.147 μg/m[3]-years, 0.148 - 1.467 μg/m[3]-years).  However, the prevalence of sensitization and CBD did not increase from the first to the second quartile.  The upper bound of the first quartile, 0.147 μg/m[3]-yrs, is approximately the cumulative exposure that a worker would have if exposed to beryllium at a level of 0.03 μg/m[3] for a working lifetime of 45 years; 0.15 μg/m[3] for 10 years; or 0.3 μg/m[3] for five years.
	A sharp increase in prevalence of CBD and total sensitization occurs between the second and third quartile (1.468 - 7.008 μg/m[3]-years) of cumulative beryllium exposure.  Cumulative exposures in the third quartile would be experienced by a worker exposed to levels between 0.03 and 0.16 μg/m3 for 45 years; 0.15 and 0.7 μg/m3 for 10 years; or 0.3 and 1.4 μg/m3 for five years.  Roughly similar levels of both sensitization and CBD occur in the highest quartile (7.009 - 61.86 μg/m[3]-years).  

4. Preliminary Statistical Analysis
	A statistical modeling analysis of the NJH data set was conducted under contract with Dr. Roslyn Stone of the University of Pittsburgh Graduate School of Public Heath, Department of Biostatistics.  Dr. Stone used statistical modeling approaches to characterize risk of sensitization and CBD among workers exposed to the preceding TWA PEL, the final TWA PEL, and each of the alternate TWA PELs that were considered for the revised beryllium rule.  Asymptotic logistic regression models provided an initial review of the data set.  This was followed up with a more rigorous statistical analysis using a proportional hazards model to develop quantitative risk estimates for sensitization and CBD at the preceding TWA PEL, the final TWA PEL, and the alternate TWA PELs considered by OSHA.  

Logistic Regression Models

	As a preliminary look at exposure-response patterns within the data set, logistic regression models were fit to two health endpoints: (1) CBD and (2) sensitization (including sensitized individuals with and without CBD).  Since only seven workers were reported as sensitized without CBD, this endpoint was not included in the statistical modeling.  Predictors tested in the models included cumulative and time-weighted average exposure measures, as well as total duration of exposure.  For average and HEJ exposure measures, models were also fit with exposure categories: 0.00 - < 0.05, 0.05 - < 0.1, 0.1 - < 0.2, 0.2 - < 0.5, 0.5 - < 1.0, 1.0 - < 2.0, and > 2.0.  A categorical analysis was also used for years of exposure: 0 - <1, 1 - < 10, 10 - < 20, and > 20 years.  The results of the categorical analyses were similar to those using continuous exposure values, and are not reported.     
	The results of the logistic regression showed a significant (p < 0.05) positive association between sensitization and cumulative exposure to beryllium, total years of exposure, and the average exposure weighted by years spent at each job, as shown in Tables 10 to 13 below.  HEJ exposure was not as strongly related to sensitization in this model.
      Table 10. Logistic Regression: Cumulative Exposure, Sensitization Endpoint

                                  Coefficient
                                  Std. Error
                                       z
                                   P>|z|
                            95% confidence interval
                                                      Exposure (μg/m[3]-years)
                                     0.03
                                     0.015
                                     2.15
                                     0.03
                                 0.003 - 0.63
                                                                       constant
                                     -2.69
                                     0.247
                                    -10.86
                                   <0.001
                                 -3.17 - -2.20
                                                                 Log-likelihood
                                    -88.72




                                                                               





                                       
  Table 11. Logistic Regression: LTW Average Exposure, Sensitization Endpoint

                                  Coefficient
                                  Std. Error
                                       Z
                                   P>|z|
                            95% confidence interval
                                                            Exposure (μg/m[3])
                                     0.92
                                     0.372
                                     2.46
                                     0.01
                                  0.19 - 1.65
                                                                       constant
                                     -2.85
                                     0.284
                                    -10.03
                                   <0.001
                                 -3.40 - -2.29
                                                                 Log-likelihood
                                    -88.06




                                       
      Table 12. Logistic Regression: HEJ Exposure, Sensitization Endpoint

                                  Coefficient
                                  Std. Error
                                       Z
                                   P>|z|
                            95% confidence interval
                                                            Exposure (μg/m[3])
                                     0.48
                                     0.279
                                     1.72
                                     0.09
                                 -0.07 - 1.03
                                                                       constant
                                     -2.78
                                     0.283
                                     -9.83
                                   <0.001
                                 -3.34 - -2.23
                                                                 Log-likelihood
                                    -90.48
                                       
                                       
                                       
                                       
                                       
                                       
   Table 13. Logistic Regression: Exposure Duration, Sensitization Endpoint
                                       

                                  Coefficient
                                  Std. Error
                                       Z
                                   P>|z|
                            95% confidence interval
                                                               Exposure (years)
                                     0.04
                                     0.019
                                     2.12
                                     0.03
                                 0.003 - 0.08
                                                                       constant
                                     -3.03
                                     0.357
                                     -8.49
                                   <0.001
                                 -3.73 - -2.33
                                                                 Log-likelihood
                                    -89.59
                                       
                                       
                                       
                                       
                                       
	As shown in Tables 14 to 17 below, risk of CBD also increased with exposure duration (total years of exposure, p = 0.02), and with cumulative beryllium exposure (p = 0.04).  Unlike beryllium sensitization, CBD was only weakly associated with average exposure and HEJ exposure in the logistic model (p = 0.12 and p = 0.39, respectively).  

       Table 14. Logistic Regression: Cumulative Exposure, CBD Endpoint

                                  Coefficient
                                  Std. Error
                                       Z
                                   P>|z|
                            95% confidence interval
                                                      Exposure (μg/m[3]-years)
                                     0.03
                                     0.02
                                     2.03
                                     0.04
                                 0.001 - 0.07
                                                                       Constant
                                     -3.04
                                     0.29
                                    -10.59
                                   <0.001
                                 -3.60 - -2.48
                                                                 Log-likelihood
                                    -70.76




                                       
       Table 15. Logistic Regression: LTW Average Exposure, CBD Endpoint

                                  Coefficient
                                  Std. Error
                                       Z
                                   P>|z|
                            95% confidence interval
                                                            Exposure (μg/m[3])
                                     0.69
                                     0.44
                                     1.57
                                     0.12
                                 -0.17 - 1.56
                                                                       constant
                                     -3.08
                                     0.32
                                     -9.55
                                   <0.001
                                 -3.71 - -2.44
                                                                 Log-likelihood
                                    -71.38
                                       
                                       
                                       
                                       
                                       

           Table 16. Logistic Regression: HEJ Exposure, CBD Endpoint

                                  Coefficient
                                  Std. Error
                                       Z
                                   P>|z|
                            95% confidence interval
                                                            Exposure (μg/m[3])
                                     0.29
                                     0.34
                                     0.86
                                     0.39
                                 -0.37 - 0.96
                                                                       constant
                                     -3.00
                                     0.32
                                     -9.37
                                   <0.001
                                 -3.62 - -2.37
                                                                 Log-likelihood
                                    -72.92
                                       
                                       
                                       
                                       
                                       
        Table 17. Logistic Regression: Exposure Duration, CBD Endpoint

                                  Coefficient
                                  Std. Error
                                       Z
                                   P>|z|
                            95% confidence interval
                                                               Exposure (years)
                                     0.05
                                     0.02
                                     2.33
                                     0.020
                                  0.01 - 0.10
                                                                       constant
                                     -3.56
                                     0.44
                                     -8.08
                                   <0.001
                                 -4.42 - -2.69
                                                                 Log-likelihood
                                    -70.52
                                       
                                       
                                       
                                       

	Dr. Stone performed the logistic regressions as a simple first analysis of the data.  Although relatively simple to interpret, standard logistic regression has substantial limitations in the present setting and is not a fully appropriate approach for analysis of this data set.  Logistic regression models the probability of an event in a fixed period of time.  As such, logistic regression cannot account for essential features of these data, specifically, differential time at risk and changing exposures over time. OSHA therefore did not use the results of the logistic regression to create quantitative risk estimates for beryllium sensitization and CBD.  Dr. Stone developed a proportional hazards model, a more appropriate model form for this data set, as a basis for the Agency's preliminary quantitative risk estimates.

Discrete Time Proportional Hazards Analysis
	Dr. Stone performed an exposure-response analysis of the NJH data set using a discrete time proportional hazards analysis (DTPHA), which was presented in section VI of the proposed rule and in a technical document (OSHA, 2014, Document ID 1589).  The DTPHA is an extension of logistic regression that allows for time-dependent exposures and differential time at risk.  It accounts for the fact that individuals in the data set are followed for different amounts of time, and that their exposures change over time.  The DTPHA is a discrete-time analog to the Cox proportional hazards model for continuous time, and was considered to be appropriate by Dr. Stone because diagnosis dates are known only to the nearest year.  In addition, where logistic regression yields odds ratios, the proportional hazards model provides hazard ratios (HRs).
	To perform this analysis, ERG constructed exposure files in Excel with cumulative and average exposures for each worker in the data set in each year that a case of sensitization or CBD was identified.  Separate files were constructed for two outcomes, sensitization (26 cases, including all cases of CBD) and diagnosed CBD (19 cases).  This analysis focuses on cumulative and long-term (eight-hour) time-weighted average (LTWA) exposure measures, as well as duration of exposure (in years).  All exposures were treated as continuous variables. A negative value for cumulative exposure for one worker, the result of an error in the data set, was set to 0.  

      The time-specific risk sets were defined as follows.  Time-dependent exposures were estimated for each worker "at risk."  In the file constructed for the sensitization exposure-response analysis, workers were considered to be at risk from the time of hire at the plant until the end of their employment or their identification as a sensitized worker, whichever came earlier.  Thus, workers found to have sensitization or CBD were considered to be at risk until the reported time of sensitization, or the time of diagnosed CBD for those CBD cases who were identified as sensitized in the same round of testing that they were diagnosed with CBD.  Those three workers who were first identified as sensitized only and later diagnosed with CBD were considered to be at risk until the reported time of sensitization.  The "sensitization" outcome was coded as missing in years following sensitization, so such observations would be dropped from the analysis. 
	Similarly, in the file constructed for the CBD exposure-response analysis, workers were considered to be at risk from the time of hire at the plant until the end of their employment or their CBD diagnosis.  Workers diagnosed with CBD were considered to be at risk until the reported time of diagnosis. Those who were first identified as sensitized only and later diagnosed with CBD were considered to be at risk until the diagnosis of CBD.  Workers who were not diagnosed with CBD were considered at risk until their last date of observation (the last year of employment from their work history).  The "CBD" outcome was coded as missing in years following diagnosis, so such observations would be dropped from the analysis. 
	For each year that a given worker is at risk for an outcome, ERG created a vector of variables denoting the case status and the values of the exposure variables for the worker at that time.  Data were converted from Excel to Stata for statistical analysis using Stat Transfer, version 8, and read into Stata SE version 10. 
	The DTPHA is the discrete time counterpart of the Cox model (Document ID 1660, p. 9).    Instead of the continuous baseline hazard that is estimated indirectly in the Cox model, the DTPHA explicitly estimates the baseline hazard using dummy variables for time.  For sensitization models, dummy variables were generated for each of the years in which a sensitized worker was identified.  For CBD models, there were dummy variables for each of the years in which a worker was diagnosed with CBD.
The DTPHA model is:
                      log(-log (1-pit))) = β0t + β1Xit
where pit is the probability of sensitization or CBD at time t, β0t is the time-specific intercept, β1 is the exposure coefficient, and Xit is the exposure for worker i at time t.  The time-specific intercept β0t was modeled as a constant plus dummy variables for the non-reference diagnosis year.  The exponentiated β1 coefficient in the proportional hazards model is interpreted as the HR per unit change in X (for continuous exposures) or the HR comparing level X to the reference level (for categorical exposures).  Exposures were treated as continuous variables for this analysis.  These models were fit using maximum likelihood estimation implemented in standard generalized linear model software.  The assumed linearity was checked using fractional polynomials. 
	In the course of preparing the data set for preliminary statistical analysis, ERG discovered that regular upkeep of the work-history records was maintained only through 2003.  Beginning in 2004, maintenance of the work-history records became uneven, and ERG found several errors in the post-2003 portion of work-history records (see OSHA, 2014, Document ID 1589).  Dr. Stone therefore performed analyses truncating exposure, work history, and screening history at 2002 (the last year before 2004 that any cases were identified).  As five workers with sensitization and CBD were identified in 2004 and 2005, the truncated analysis included fewer cases of sensitization and CBD than the analysis using post-2003 data. 

Preliminary Models
	The exposure-response models resulting from the discrete proportional hazards analyses, based on the data set including cases diagnosed in 2004 and 2005, are presented in Tables 18 to 25 below.  All coefficients used in the models are displayed, including the exposure coefficient, the model constant, and additional coefficients for year of diagnosis fit in the proportional hazards modeling procedure.  
                Table 18. Cumulative Exposure and Sensitization
                                       
                                   Variable
                                  Coefficient
                            95% confidence interval
                                    P-value
                                              Cumulative Exposure (μg/m3 -yrs)
                                                                           0.03
                                -0.003 to 0.06
                                                                           0.08
                                                                       Constant
                                                                          -3.43
                                -4.20 to -2.66
                                                                      <0.001
                                                                           1996
                                                                          -1.49
                                 -3.04 to 0.06
                                                                           0.06
                                                                           1997
                                                                          -0.30
                                 -1.31 to 0.72
                                                                           0.57
                                                                           1998
                                                                          -1.57
                                -3.12 to -0.02
                                                                           0.05
                                                                           1999
                                                                          -1.58
                                -3.13 to -0.03
                                                                           0.05
                                                                           2004
                                                                          -1.77
                                -3.33 to -0.22
                                                                           0.03
                                                                           2005
                                                                          -1.37
                                -2.70 to -0.03
                                                                           0.04
                                       
               Table 19. LTW Average Exposure and Sensitization
                                   Variable
                                  Coefficient
                            95% confidence interval
                                    P-value
                                                    Average Exposure (μg/m[3])
                                                                           0.44
                                 -0.18 to 1.05
                                                                           0.16
                                                                       Constant
                                                                          -3.47
                                -4.31 to -2.63
                                                                      <0.001
                                                                           1996
                                                                          -1.49
                                 -3.04 to 0.06
                                                                           0.06
                                                                           1997
                                                                          -0.30
                                 -1.32 to 0.72
                                                                           0.56
                                                                           1998
                                                                          -1.56
                                -3.11 to -0.004
                                                                           0.05
                                                                           1999
                                                                          -1.55
                                -3.11 to 0.002
                                                                           0.05
                                                                           2004
                                                                          -1.73
                                -3.30 to -0.16
                                                                           0.03
                                                                           2005
                                                                          -1.32
                                 -2.66 to 0.03
                                                                           0.06
                                       
                   Table 20. HEJ Exposure and Sensitization
                                       
                                   Variable
                                  Coefficient
                            95% confidence interval
                                    P-value
                                                        HEJ Exposure (μg/m[3])
                                                                           0.20
                                 -0.34 to 0.75
                                                                           0.47
                                                                       Constant
                                                                          -3.33
                                -4.15 to -2.51
                                                                      <0.001
                                                                           1996
                                                                          -1.50
                                 -3.05 to 0.05
                                                                           0.06
                                                                           1997
                                                                          -0.32
                                 -1.33 to 0.70
                                                                           0.54
                                                                           1998
                                                                          -1.60
                                -3.15 to -0.05
                                                                           0.04
                                                                           1999
                                                                          -1.62
                                -3.17 to -0.06
                                                                           0.04
                                                                           2004
                                                                          -1.82
                                -3.38 to -0.27
                                                                           0.02
                                                                           2005
                                                                          -1.41
                                -2.74 to -0.08
                                                                           0.04

                 Table 21. Exposure Duration and Sensitization

                                   Variable
                                  Coefficient
                            95% confidence interval
                                    P-value
                                                      Exposure Duration (years)
                                                                           0.04
                                -0.009 to 0.08
                                                                           0.11
                                                                       Constant
                                                                          -3.68
                                -4.63 to -2.72
                                                                      <0.001
                                                                           1996
                                                                          -1.47
                                 -3.02 to 0.08
                                                                           0.06
                                                                           1997
                                                                          -0.29
                                 -1.30 to 0.73
                                                                           0.58
                                                                           1998
                                                                          -1.58
                                -3.13 to -0.03
                                                                           0.05
                                                                           1999
                                                                          -1.62
                                -3.17 to -0.07
                                                                           0.04
                                                                           2004
                                                                          -1.84
                                -3.39 to -0.29
                                                                           0.02
                                                                           2005
                                                                          -1.42
                                -2.75 to -0.10
                                                                           0.04

	As shown in Tables 18 to 21 above, relative risk of beryllium sensitization increased with cumulative exposure, LTW average exposure, HEJ exposure, and exposure duration, but the relationships were not statistically significant.  
	The models for the CBD endpoint (Tables 22 to 25 below) using data through 2005 also showed positive, but not statistically significant relationships between the relative risk of CBD and cumulative exposure, LTW average exposure, and HEJ exposure.  A significant positive association was observed between duration of exposure and CBD (p = 0.04).  
                                       
                                       
                                       
                                       
                                       
                                       
                    Table 22.  Cumulative Exposure and CBD
                                       
Variable
                                  Coefficient
                            95% confidence interval
                                    P-value
                                              Cumulative Exposure (μg/m3 -yrs)
                                                                           0.03
                                 -0.01 to 0.06
                                                                           0.09
                                                                       Constant
                                                                          -3.75
                                -4.64 to -2.86
                                                                      <0.001
                                                                           1997
                                                                          -0.59
                                 -1.86 to 0.67
                                                                           0.36
                                                                           1998
                                                                          -2.01
                                 -4.13 to 0.10
                                                                           0.06
                                                                           1999
                                                                          -0.63
                                 -1.90 to 0.63
                                                                           0.33
                                                                           2002
                                                                          -2.13
                                -4.25 to -0.01
                                                                           0.05
                                                                           2004
                                                                          -2.19
                                -4.31 to -0.07
                                                                           0.04
                                                                           2005
                                                                          -1.49
                                 -3.10 to 0.11
                                                                           0.07
                                       
                    Table 23. LTW Average Exposure and CBD
                                       
                                   Variable
                                  Coefficient
                            95% confidence interval
                                    P-value
                                               Average Exposure (μg/m[3] -yrs)
                                                                           0.19
                                 -0.61 to 0.99
                                                                           0.58
                                                                       Constant
                                                                          -3.59
                                -4.54 to -2.63
                                                                      <0.001
                                                                           1997
                                                                          -0.61
                                 -1.88 to 0.66
                                                                           0.34
                                                                           1998
                                                                          -2.03
                                 -4.15 to 0.11
                                                                           0.09
                                                                           1999
                                                                          -0.65
                                 -1.92 to 0.62
                                                                           0.32
                                                                           2002
                                                                          -2.17
                                -4.30 to -0.04
                                                                           0.05
                                                                           2004
                                                                          -2.23
                                -4.37 to -0.10
                                                                           0.04
                                                                           2005
                                                                          -1.54
                                 -3.15 to 0.03
                                                                           0.08

                        Table 24. HEJ Exposure and CBD
                                   Variable
                                  Coefficient
                            95% confidence interval
                                    P-value
                                                   HEJ Exposure (μg/m[3] -yrs)
                                                                           0.02
                                 -0.67 to 0.71
                                                                           0.96
                                                                       Constant
                                                                          -3.48
                                 -4.42 to 2.55
                                                                      <0.001
                                                                           1997
                                                                          -0.62
                                 -1.88 to 0.65
                                                                           0.34
                                                                           1998
                                                                          -2.05
                                 -4.16 to 0.07
                                                                           0.06
                                                                           1999
                                                                          -0.68
                                 -1.95 to 0.59
                                                                           0.29
                                                                           2002
                                                                          -2.21
                                -4.33 to -0.09
                                                                           0.04
                                                                           2004
                                                                          -2.28
                                -4.40 to -0.16
                                                                           0.04
                                                                           2005
                                                                          -1.58
                                 -3.19 to 0.02
                                                                           0.05
                                       
                                       
                                       
                     Table 25.   Exposure Duration and CBD
                                   Variable
                                  Coefficient
                            95% confidence interval
                                    P-value
                                                      Exposure Duration (years)
                                                                           0.06
                                -0.001 to -0.11
                                                                           0.04
                                                                       Constant
                                                                          -4.24
                                -5.39 to -3.10
                                                                      <0.001
                                                                           1997
                                                                          -0.57
                                 -1.84 to 0.69
                                                                           0.38
                                                                           1998
                                                                          -2.01
                                 -4.13 to 0.11
                                                                           0.06
                                                                           1999
                                                                          -0.67
                                 -1.94 to 0.59
                                                                           0.30
                                                                           2002
                                                                          -2.23
                                -4.34 to -0.11
                                                                           0.04
                                                                           2004
                                                                          -2.27
                                -4.38 to -0.15
                                                                           0.04
                                                                           2005
                                                                          -1.57
                                 -3.17 to 0.03
                                                                           0.06

	As discussed previously, ERG discovered several errors in the last years of work history (2004-2005) in the NJH data set.  To test whether excluding these years would affect the model results, Dr. Stone conducted an analysis on a subcohort of workers hired prior to 2003, truncating work histories and screening histories at the end of 2002.  This analysis showed that truncating the analysis at 2002 appeared to affect models that use beryllium sensitization as the endpoint, as shown in Tables 26 to 29 below.  The truncated data set was smaller than the full data set and included fewer cases of sensitization.  It might therefore be expected to yield a less statistically significant exposure coefficient than the full data set.  However, the p-values associated with the exposure coefficients were lower in the truncated data set than in the full data set for the cumulative, LTW average, and HEJ exposure measures.  Thus, truncating the data set at 2002, before the point when work histories were no longer reliably updated, may have reduced the amount of error in the Cullman workers' exposure estimates.
	Similar to the full cohort analysis, relative risk of sensitization increased with cumulative exposure, LTW average exposure, HEJ exposure, and exposure duration in the analysis of the subcohort.  The relationship was borderline statistically significant (p = 0.05) in the cumulative exposure analysis truncated in 2002.  In the other models, the relationship was not statistically significant.  However, the model using average exposure and truncated at 2002 was suggestive of a relationship between average exposure and relative risk of sensitization (p = 0.09).  The association was much weaker for exposure duration (p = 0.31), consistent with the expected biological action of an immune hypersensitivity response where onset is believed to be more dependent on the concentration of the sensitizing agent at the target site rather than the number of years of occupational exposure.  The association was also much weaker for HEJ exposure (p = 0.3).
               Table 26.  Cumulative Exposure and Sensitization
                                   Variable
                                  Coefficient
                            95% confidence interval
                                    P-value
                                              Cumulative Exposure (μg/m3 -yrs)
                                                                          0.031
                                -0.001 to 0.063
                                                                           0.05
                                                                       Constant
                                                                          -3.48
                                -4.27 to -2.69
                                                                      <0.001
                                                                           1996
                                                                          -1.49
                                 -3.04 to 0.06
                                                                           0.06
                                                                           1997
                                                                          -0.29
                                 -1.31 to 0.72
                                                                           0.57
                                                                           1998
                                                                          -1.56
                                -3.11 to -0.01
                                                                           0.05
                                                                           1999
                                                                          -1.57
                                -3.12 to -0.02
                                                                           0.05
   
                 Table 27.  Average Exposure and Sensitization

                                   Variable
                                  Coefficient
                            95% confidence interval
                                    P-value
                                                    Average Exposure (μg/m[3])
                                                                           0.54
                                 -0.09 to 1.17
                                                                           0.09
                                                                       constant
                                                                          -3.55
                                -4.42 to -2.69
                                                                      <0.001
                                                                           1996
                                                                          -1.48
                                 -3.03 to 0.07
                                                                           0.06
                                                                           1997
                                                                          -0.29
                                 -1.31 to 0.72
                                                                           0.57
                                                                           1998
                                                                          -1.54
                                 -3.09 to 0.01
                                                                           0.05
                                                                           1999
                                                                          -1.53
                                 -3.08 to 0.03
                                                                           0.05

                   Table 28.  HEJ Exposure and Sensitization
                                   Variable
                                  Coefficient
                            95% confidence interval
                                    P-value
                                                        HEJ Exposure (μg/m[3])
                                                                           0.31
                                 -0.27 to 0.88
                                                                           0.30
                                                                       constant
                                                                          -3.42
                                -4.27 to -2.56
                                                                      <0.001
                                                                           1996
                                                                          -1.49
                                 -3.04 to 0.06
                                                                           0.06
                                                                           1997
                                                                          -0.31
                                 -1.33 to 0.70
                                                                           0.55
                                                                           1998
                                                                          -1.59
                                -3.14 to -0.04
                                                                           0.05
                                                                           1999
                                                                          -1.60
                                -3.15 to -0.05
                                                                           0.04
                                       
                                       
                Table 29.  Exposure Duration and Sensitization
                                   Variable
                                  Coefficient
                            95% confidence interval
                                    P-value
                                                      Exposure Duration (years)
                                                                           0.03
                                 -0.03 to 0.08
                                                                           0.31
                                                                       Constant
                                                                          -3.55
                                -4.57 to -2.53
                                                                      <0.001
                                                                           1996
                                                                          -1.48
                                 -3.03 to 0.70
                                                                           0.06
                                                                           1997
                                                                          -0.30
                                 -1.31 to 0.72
                                                                           0.57
                                                                           1998
                                                                          -1.59
                                -3.14 to -0.04
                                                                           0.05
                                                                           1999
                                                                          -1.62
                                -3.17 to -0.72
                                                                           0.04

	Truncating the analysis at 2002 had less of an impact on models for the CBD endpoint, as shown in Tables 30 to 33 below.  This analysis showed positive, but not statistically significant, relationships between relative risk of CBD and cumulative exposure, average exposure, HEJ exposure, and exposure duration. Similar to the initial analysis, the model using cumulative exposure was suggestive of a relationship in the analysis truncated at 2002 (p = 0.09), whereas average and HEJ exposures were not significantly related to CBD risk (p > 0.5 for all models).  

                    Table 30.   Cumulative Exposure and CBD
                                                                       Variable
                                                                    Coefficient
                                                        95% confidence interval
                                                                        P-value
                                              Cumulative Exposure (μg/m3 -yrs)
                                                                           0.03
                                                                   0.00 to 0.07
                                                                           0.09
                                                                       Constant
                                                                          -3.77
                                                                 -4.67 to -2.86
                                                                      <0.001
                                                                           1997
                                                                          -0.59
                                                                  -1.86 to 0.68
                                                                           0.36
                                                                           1998
                                                                          -2.01
                                                                  -4.13 to 0.11
                                                                           0.06
                                                                           1999
                                                                          -0.63
                                                                  -1.90 to 0.64
                                                                           0.33
                                                                           2002
                                                                          -2.13
                                                                 -4.25 to -0.01
                                                                           0.05

                                       
                                       
                   Table 31.   LTW Average Exposure and CBD
                                                                       Variable
                                                                    Coefficient
                                                        95% confidence interval
                                                                        P-value
                                                LTW Average Exposure (μg/m[3])
                                                                           0.24
                                 -0.59 to 1.06
                                                                           0.58
                                                                       constant
                                                                          -3.62
                                -4.60 to -2.64
                                                                      <0.001
                                                                           1997
                                                                          -0.61
                                 -1.87 to 0.66
                                                                           0.35
                                                                           1998
                                                                          -2.02
                                 -4.14 to 0.10
                                                                           0.06
                                                                           1999
                                                                          -0.64
                                 -1.92 to 0.63
                                                                           0.32
                                                                           2002
                                                                          -2.15
                                -4.28 to -0.02
                                                                           0.05

                       Table 32.   HEJ Exposure and CBD
                                                                       Variable
                                                                    Coefficient
                                                        95% confidence interval
                                                                        P-value
                                                        HEJ Exposure (μg/m[3])
                                                                           0.03
                                 -0.70 to 0.77
                                                                           0.93
                                                                       Constant
                                                                          -3.49
                                -4.45 to -2.53
                                                                      <0.001
                                                                           1997
                                                                          -0.62
                                 -1.88 to 0.65
                                                                           0.34
                                                                           1998
                                                                          -2.05
                                 -4.16 to 0.07
                                                                           0.06
                                                                           1999
                                                                          -0.68
                                 -1.94 to 0.59
                                                                           0.30
                                                                           2002
                                                                          -2.21
                                -4.33 to -0.09
                                                                           0.04
                                       
                     Table 33.  Exposure Duration and CBD
                                                                       Variable
                                                                    Coefficient
                                                        95% confidence interval
                                                                        P-value
                                                        Exposure Duration (yrs)
                                                                          0.052
                                                                  -0.01 to 0.11
                                                                           0.10
                                                                       constant
                                                                          -4.18
                                                                 -5.40 to -2.96
                                                                     < 0.001
                                                                           1997
                                                                          -0.53
                                                                  -1.84 to 0.69
                                                                           0.38
                                                                           1998
                                                                          -2.01
                                                                  -4.13 to 0.11
                                                                           0.06
                                                                           1999
                                                                          -0.67
                                                                  -1.94 to 0.60
                                                                           0.30
                                                                           2002
                                                                          -2.22
                                                                 -4.34 to -0.10
                                                                           0.04

	In the models described in Tables 26 to 33, cumulative exposure and exposure duration appeared to be more closely related to relative risk of both sensitization and CBD than LTW average exposure or HEJ exposure.  For CBD, this is not surprising since the disease is believed to respond to the accumulated burden of beryllium retained in the lung over time.  Lung burden would be expected to primarily track number of years exposed, cumulative exposure, and, to a lesser extent, long-term average exposure.  The model results are consistent with this biological expectation.
	On the other hand, beryllium sensitization is an immune hypersensitivity response that is expected to be primarily dependent on beryllium concentration at the target site, rather than accumulated lung burden over an extended period of time.  The more frequently the critical target concentration is achieved, the greater the risk of becoming sensitized.  This biological mode of action is consistent with the sensitization model results that show a stronger statistical association with cumulative and LTW average exposure than with number of years exposed.    
	OSHA estimated risk using the models presented in Table 26 and Table 30 (cumulative exposure, data truncated at 2002) for the preceding TWA PEL of 2 μg/m[3], the new TWA PEL of 0.2  μg/m[3], and the other TWA PELs that OSHA considered: 0.5 μg/m[3], 0.2 μg/m[3], and 0.1 μg/m[3].  To estimate risk of sensitization and CBD from a working lifetime of exposure, the Agency assumed 45 years of exposure at each level, for total cumulative exposures of 90, 45, 22.5, 9, and 4.5 μg/m[3]-years (corresponding to 45 years of occupational exposure at 2 μg/m[3], 1 μg/m[3], 0.5 μg/m[3], 0.2 μg/m[3], and 0.1 μg/m[3], respectively).  However, most workers in the data set were exposed to beryllium for shorter time periods.  There is, therefore, more uncertainty associated with risk estimates for a working lifetime of exposure than with risk estimates for shorter exposure durations.  In addition to risk estimates for 45 years of exposure, OSHA estimated risk associated with five, ten, and twenty years of exposure at each level.  The Agency used the models based on cumulative exposure in the data set truncated at 2002 to derive risk estimates because the cumulative exposure metric showed the most consistent association with sensitization and CBD in these models.  
	Risk estimates for sensitization and CBD were calculated using the equation:
                                       
where pit is the probability that worker i is found to have the health outcome of interest (sensitization or CBD) at time t, β0t is the time-specific intercept, β1 is the exposure coefficient, and Xit is the cumulative exposure for worker i at time t. The time-specific intercept β0t is modeled as a constant plus dummy variables for the non-reference diagnosis year.  Thus, the model provides time-specific estimates of risk for a worker with any given exposure level, varying across the years in which cases were diagnosed.  
	Table 34 below presents a range of sensitization risk estimates based on all time-specific intercepts, where the upper end of the range is derived using β0,1995 = -3.48 and the lower end is derived using β0,1998 = -5.08 (the time-specific intercepts for 1995 and 1999, respectively).  Each estimate represents the number of sensitized workers the model predicts in a group of 1,000 workers exposed at the specified level, duration, and time.  Thus, for example, 56 workers would be expected to be sensitized among 1,000 occupationally exposed at 2 μg/m[3] for 10 years from 1985 to 1995.  Similarly, Table 35 presents a range of CBD risk estimates based on all time-specific intercepts, where the upper end of the range is derived using β0,1995 = -3.77 and the lower end is derived using β0,2002 = -5.90 (the time-specific intercepts for 1995 and 2002, respectively).
                                   Table 34
Expected Cases of Sensitization per 1,000 Workers Exposed at Preceding, Final, and Alternate TWA PELs; Cumulative Exposure Metric
                                       

                               Exposure Duration

                                    5 years
                                   10 years
                                   20 years
                                   45 years
                           Exposure Level (μg/m[3])
Cumulative (μg/m[3]-yrs)
                                  cases/1000
                                 μg/m[3]-yrs
                                  cases/1000
                                 μg/m[3]-yrs
                                  cases/1000
                                 μg/m[3]-yrs
                                  cases/1000
                                                                            2.0
                                     10.0
                                  8.4 - 41.1
                                     20.0
                                 11.5  -  55.7
                                     40.0
                                21.3  -  101.0
                                     90.0
                                96.3  -  394.4
                                                                            1.0
                                      5.0
                                  7.2 - 35.3
                                     10.0
                                  8.4 - 41.1
                                     20.0
                                 11.5  -  55.7
                                     45.0
                                24.8  -  116.9
                                                                            0.5
                                      2.5
                                  6.7 - 32.7
                                      5.0
                                  7.2 - 35.3
                                     10.0
                                  8.4 - 41.1
                                     22.5
                                  12.4 - 60.0
                                                                            0.2
                                      1.0
                                  6.4 - 31.3
                                      2.0
                                  6.6 - 32.2
                                      4.0
                                  7.0 - 34.3
                                      9.0
                                 8.2  -  39.9
                                                                            0.1
                                      0.5
                                  6.3 - 30.8
                                      1.0
                                  6.4 - 31.3
                                      2.0
                                  6.6 - 32.2
                                      4.5
                                  7.1 - 34.8
                                       
                                       
                                   Table 35
Expected Cases of CBD per 1,000 Workers Exposed at Preceding, Final, and Alternate TWA PELs;
                           Cumulative Exposure Metric
                                       

                               Exposure Duration

                                    5 years
                                   10 years
                                   20 years
                                   45 years
                           Exposure Level (μg/m[3])
Cumulative (μg/m[3]-yrs)
                                  cases/1000
                                 μg/m[3]-yrs
                                  cases/1000
                                 μg/m[3]-yrs
                                  cases/1000
                                 μg/m[3]-yrs
                                  cases/1000
                                                                            2.0
                                     10.0
                                  3.7 - 30.9
                                     20.0
                                 5.1  -  41.9
                                     40.0
                                 9.4  -  76.6
                                     90.0
                                43.6  -  312.9
                                                                            1.0
                                      5.0
                                  3.2 - 26.6
                                     10.0
                                 3.7  -  30.9
                                     20.0
                                  5.1 - 41.9
                                     45.0
                                 11.0  -  88.8
                                                                            0.5
                                      2.5
                                  3.0 - 24.6
                                      5.0
                                 3.2  -  26.6
                                     10.0
                                  3.7 - 30.9
                                     22.5
                                 5.5  -  45.2
                                                                            0.2
                                      1.0
                                  2.8 - 23.5
                                      2.0
                                 2.9  -  24.2
                                      4.0
                                  3.1 - 25.8
                                      9.0
                                 3.6  -  30.0
                                                                            0.1
                                      0.5
                                  2.8 - 23.1
                                      1.0
                                 2.8  -  23.5
                                      2.0
                                  2.9 - 24.2
                                      4.5
                                  3.1 - 26.2

	This preliminary statistical modeling analysis predicted substantial risks of both sensitization and CBD at the preceding PEL of 2 μg/m[3], with particularly high risks predicted for a 45-year working lifetime of exposure.  The predicted risk is much reduced, but still significant, at the final TWA PEL of 0.2 and the action level of 0.1 μg/m[3]. 

Comments Received on the Preliminary Statistical Analysis
      As discussed in the NPRM, OSHA did not rely on this preliminary statistical analysis for its significance of risk determination or to develop its benefits analysis due to limitations in the analysis, primarily constraints related to the size of the study.  The Agency relied on its analyses of the epidemiological literature and the prevalence analysis of the NJH data for its preliminary significance of risk determination, and on the prevalence analysis for its preliminary estimate of benefits.  Although OSHA did not rely on the results of the preliminary statistical analysis for its findings, the Agency presented the DTPHA in the NPRM in order to inform the public of its results, explain its limitations, and solicit public comment on the Agency's approach.  After considering comments on this preliminary model, OSHA instructed its contractor to change the statistical analysis to address technical concerns and to incorporate suggestions from Crump and Proctor, as well as NIOSH (Document ID 1660; 1725).  The comments OSHA received on the preliminary statistical analysis, OSHA's response to these comments, and the methods and results of the final statistical analysis are presented here. 
      Dr. Kenny Crump and Ms. Deborah Proctor submitted comments on OSHA's preliminary risk assessment (Document ID 1660).  Crump and Proctor agreed with OSHA's review of the epidemiological literature and the prevalence analysis presented previously in this section.  They stated that they "agree with OSHA's conclusion that there is a significant risk (> 1/1000 risk of CBD) at the [then] current PEL, and that risk is reduced at the [then] proposed PEL (0.2 μg/m[3]) in combination with stringent measures (ancillary provisions) to reduce worker's exposures. This finding is evident based on the available literature...and the prevalence data presented for the Cullman facility" (Document ID 1660, p. 2).  Crump and Proctor also provided a detailed evaluation of the statistical analysis of the NJH data presented in the NPRM, including a critique of OSHA's modeling approach and interpretation and suggestions for alternate analyses. 
      One of the concerns raised by Crump and Proctor was with OSHA's interpretation of the Discrete Time Proportional Hazards Analysis (DTPHA):  "OSHA misinterpreted the output of the DTPHA as a cumulative probability of effect (CBD or BeS [sensitization]), whereas in reality it represents the condition[al] probability of effect at a specific time, given no previous effect" (Document ID 1660, p. 1).  That is, they believed that OSHA interpreted the results of the DTPHA to reflect the total predicted cases of CBD or sensitization, rather than the number of newly identified cases at any given time.  
      Crump and Proctor's criticism likely reflects a misunderstanding that arose from the way OSHA described the results of the proportional hazards analysis in the NPRM.  There, OSHA stated: 
      [t]he models provide time-specific point estimates of risk for a worker with any given exposure level.... Each estimate represents the number of sensitized workers the model predicts in a group of 1000 workers at risk during the given year with an exposure history at the specified level and duration.  For example, in the exposure scenario where 1000 workers are occupationally exposed to 2 μg/m[3] for 10 years in 1995, the model predicts that about 56 (55.7) workers would be sensitized that year.  The model for CBD predicts that about 42 (41.9) workers would be diagnosed with CBD that year (80 FR 47639 (8/7/15)).

In the NPRM's description, OSHA intended the phrase "the number of sensitized workers" to mean the number of sensitized workers identified in the facility's screening program.  Similarly, OSHA intended "[number] workers would be sensitized that year" to mean the number of workers who would be identified as sensitized in screening that year.  The phrase OSHA used for the CBD model results ("[number] workers would be diagnosed with CBD that year") better conveys OSHA's intended meaning.  In the discussion of modeling results presented here, OSHA is careful to refer to cases of sensitization identified and cases of CBD diagnosed in any given year.  These phrases mean cases of sensitized newly identified and cases of CBD newly diagnosed, which OSHA believes will be evident to readers because this is the usual meaning of "cases identified" or "cases diagnosed" in a given year.  Thus, OSHA's interpretation of the proportional hazards model results is consistent with Crump and Proctor's description of "the condition[al] probability of effect at a specific time, given no previous effect" (Document ID 1660, p. 1), and OSHA has clarified its description of model results accordingly. 
      Crump and Proctor also commented that the type of model used by OSHA in the preliminary risk analysis (the DTPHA) is used as a substitute for a different method, the Cox continuous proportional hazard model, "to make computation feasible when analyzing a large data set with a large number of tied times."  Because the NJH data set is small, they argued, it is not necessary to use a DTPHA in place of a Cox proportional hazards model (Document ID 1660, p. 9).  OSHA has reviewed the NJH data set, and agrees with Crump and Proctor that the computations required for the Cox analysis are feasible.  OSHA concluded that the Cox model is a more appropriate model for analyzing this data set than the DTPHA that was used in the preliminary analysis. Thus, Dr. Stone conducted the reanalysis of the NJH data using the Cox proportional hazards model.  
      Crump and Proctor further commented that the DTPHA may not be appropriate to analyze the NJH data set, because one of the assumptions of the DTPHA model is that "the health status of all cohort members is fully known at every discrete time applied in the modeling" (that is, the sensitization or CBD status of every worker in the data set is known at each time that some cases of sensitization or CBD are reported) (Document ID 1660, p. 5).  The NJH data set does not report results for every worker at every testing time because, while annual testing was conducted in the facility beginning in 1995, individual workers were only tested every other year.  Crump and Proctor also commented that the DTPHA cannot account for irregularity in the length of time between rounds of testing (Document ID 1660, p. 9).  They are referring to the fact that testing for beryllium sensitization and CBD began in 1995 and was conducted every year thereafter.  However, by 1995, many workers in the data set had been employed at the Cullman facility for years.  The first interval between time points in the model (1969 to 1995) was therefore much longer than subsequent intervals.  Crump and Proctor recommend that OSHA use the screening protocol and date of hire for each worker to determine the most likely years of testing for each worker and use both the last year a subject was known to be free of sensitization (or CBD) and the first year he or she was known to have sensitization (or CBD) in the modeling procedure to calculate his or her contribution to model likelihood (Document ID 1660, p. 10).
      OSHA agrees that the screening protocol can be used to develop reasonable assumptions about the years of testing for each worker in the data set.  For the reanalysis, Dr. Stone used the screening protocol to determine years in which each worker in the data set was scheduled to receive a BeLPT and removed workers who were not scheduled to be tested in a given year from the analysis for that year.  This procedure addresses Crump and Proctor's concern that the sensitization or CBD status of every worker in the data set is not known at each time that some cases of sensitization or CBD are reported (Document ID 1660, p. 5).  However, OSHA does not agree that Crump and Proctor's suggested approach can adequately address the issue they raised regarding the difference between the first interval (prior to medical testing) and subsequent testing intervals. Because workers in the data set had been exposed to beryllium for many years before the first round of medical screening in 1995, the results of medical testing from this year represent the pre-existing (prevalent) cases of sensitization and CBD in a population with long-term exposures.  That is, there is no information on when in the interval before 1995 these cases developed.  In contrast, the results of testing from years following 1995 represent new (incident) cases arising in the population.  
      The method suggested by Crump and Proctor cannot adequately address the fact that undetected cases of sensitization and CBD had accrued for up to 26 years between 1969 (the year the Cullman facility opened) and the first screening in 1995.  Due to the length of this interval, Crump and Proctor's suggested statistical method cannot adequately approximate the time that cases of sensitization and CBD arose within this span of time.  For this reason, instead of using Crump and Proctor's proposed method, OSHA has excluded the cases of sensitization and CBD that were found in the 1995 screening from the reanalysis.  This step greatly reduces uncertainty in the timing of case development in the analysis.  After excluding the cases identified in 1995, the remaining cases in the data set were identified within two years of their development, since workers were screened every two years beginning in 1995.  
      Crump and Proctor stated that OSHA should have provided the statistical uncertainty around the risk estimates for beryllium sensitization and CBD derived from the DTPHA (Document ID 1660, pp. 1-2).  OSHA notes that Tables VI-18a and VI-18b in Chapter VI of the NPRM (80 FR 47640-41) report ranges of risk based on the 95% confidence interval around the exposure coefficient; however, these intervals did not include statistical uncertainty in the "baseline" risk due to issues in the estimation of baseline risk in OSHA's model.  As pointed out by NIOSH in its comment, the meaning of these ranges was clearly described in the NPRM (Document ID 1725, p. 6).  For the re-analyses that Dr. Stone conducted to address other comments and suggestions raised by Crump and Proctor, the Agency presents 95% confidence intervals consistent with Crump and Proctor's suggestion.  Since the Cox proportional hazard models do not estimate baseline risk, the 95% confidence intervals based on the exposure coefficient uncertainty fully represent the statistical uncertainty in OSHA's risk estimates for beryllium sensitization and CBD.
      Crump and Proctor commented that "OSHA's implementation of the DTPHA estimated yearly background levels of [beryllium] effects that were implausibly variable and implausibly high" (Document ID 1660, p. 2).  In addition, "[t]he estimated ranges of risks from posited [beryllium] exposures include the `background' risks estimated by OSHA's model, and consequently a sizable fraction of OSHA's estimated exposure-related risk is not due to [beryllium] exposure" (Document ID 1660, p. 2).  Crump and Proctor further commented that "a separate background response was estimated for each year included in the analysis... [A] more appropriate way of handling background rates in the DTPHA would be to make them independent of time, and bounded to insure reasonable values that are consistent with the literature," that is, no more than 1 percent, the rate of confirmed sensitization among new employees in Donovan et al. (2007) (Document ID 1660, p. 6).
      In response to Crump and Proctor's comments, NIOSH noted that it is difficult to judge the plausibility of baseline rates based on currently available information, and that the yearly baseline rates estimated in the DTPHA model for sensitization were not much different from Crump and Proctor's suggestion of 1 percent as a fixed baseline (Document ID 1725, p. 6).  NIOSH stated that a near zero baseline rate for sensitization and CBD is expected, and suggested that OSHA could use a fixed low intercept (the statistical model term representing baseline risk) such as 0.5 percent in its statistical model (Document ID 1725, p. 6).
      OSHA agrees with Crump and Proctor that the baseline risks estimated in the DTPHA model were implausibly variable and high.  However, OSHA does not agree with their suggestion to subtract baseline risks from OSHA's risk estimates, because the "baseline" risks in the statistical model do not reflect a "background" (i.e., unrelated to beryllium exposure) level of sensitization or CBD.  As explained in Section V, Health Effects, beryllium sensitization and CBD occur only as a result of exposure to beryllium.  As a result, beryllium sensitization and CBD do not have a "background" rate unrelated to beryllium exposure.  Thus, both the "baseline" rates of beryllium sensitization and CBD estimated by the proportional hazards models and the additional cases predicted at various exposure levels reflect cases acquired from beryllium exposure, and OSHA intentionally reported the combined number of cases in its risk estimates.  Although OSHA understands that subtracting "baseline" from total risk is standard practice when a true baseline risk exists, OSHA does not believe that subtracting the model's "baseline" from the total number of predicted sensitization and CBD cases was necessarily appropriate in reporting the results of the preliminary risk assessment for beryllium.  OSHA therefore disagrees with Crump and Proctor's criticism and their conclusion that the Agency erred in presenting the total number of predicted cases of sensitization and CBD in its preliminary risk assessment.  OSHA does, however, find that the suggestion offered by Crump and Proctor and NIOSH to fix a low intercept is reasonable.  The reanalysis presented later in this section therefore estimates risk using two choices of "background" risk for sensitization, 1 percent and 0.5 percent, and a "background" rate of 0.5 percent for CBD, based on Crump, Proctor, and NIOSH's recommendations. 
      Crump and Proctor additionally expressed concern that the results of the fractional polynomial analysis OSHA conducted to check for nonlinearities in the DTPHA models (which did not find evidence of such nonlinearities) might have been affected by the issues of baseline rate estimation in the DTPHA discussed above (Document ID 1660, pp. 7-8).  Crump and Proctor further speculated that, due to genetic susceptibility factors that can influence individuals' risk of sensitization or CBD, workers' risk might increase sharply with increasing exposure at lower exposures and then approach a plateau at higher exposures.  They presented graphs based on the prevalence of sensitization and CBD in the NJH data set that appear to show a sharp increase in risk of both around the level of 5 μg/m[3]-years cumulative exposure (Document ID 1660, pp. 8-9).  To address Crump and Proctor's concern and confirm that a non-linear model is still not appropriate following changes made for the reanalysis, Dr. Stone performed a new fractional polynomial analysis to check for possible nonlinearities in the reanalysis, which incorporates fixed baseline rates as recommended by Crump and Proctor. 
      Crump and Proctor also criticized OSHA for excluding three years of data (2003 to 2005) from some models in the DTPHA due to uncertainty in workers' job histories (Document ID 1660, p 11).  They pointed out that workers' exposure estimates from the early years of the Cullman facility's operation are at least as uncertain (due to lack of sampling data) as those from 2003 to 2005; therefore, they said, it does not make sense to include data from the early years but not from 2003 to 2005.  They recommended that OSHA consider including data from the years 2003 to 2005, and also perform an analysis excluding workers hired before 1980 from the data to explore the effects of uncertainty in exposure estimates for the early years of the plant on the model results.  OSHA agrees and has incorporated both of these suggestions into the reanalysis, described below.
	Crump and Proctor further recommended that OSHA develop a statistical model of beryllium sensitization risk using an exposure metric designed to capture beryllium concentration.  In particular, they suggested that OSHA construct an additional exposure metric, such as duration of exposure above a specified air concentration or cumulative exposure above a specified air concentration (Document ID 1660, pp. 11-12).  However, OSHA does not have the data necessary to conduct a meaningful analysis along these lines.  The exposure data in the NJH data set reflect long-term averages (exposure levels averaged over a year or more); they do not show how much time any given worker was exposed above any particular concentration.  Therefore, OSHA was unable to develop an analysis of this type.  
      Finally, Crump and Proctor suggested that OSHA should conduct analyses that "assume two subpopulations with differing sensitivities to [beryllium], and posit distinct hazards for each subpopulation" to conduct a "formal test of the hypothesis of a sensitive subpopulation" (Document ID 1660, pp. 10-11).  OSHA does not believe such an analysis would benefit the beryllium risk analysis.  The procedure recommended by Crump and Proctor does not account for the complexity of genetic risk factors related to beryllium exposure (which exist in a variety of forms) or the lack of well-established information about their effects (see Section V, Health Effects, for a discussion of genetic risk factors for beryllium-related health effects).  It is therefore likely to generate misleading results.  The genetic risk factors of individuals in the NJH data set are unknown and the data set is too small to support a model-fitting procedure to effectively test for the presence of multiple sensitive subpopulations.  Furthermore, the relative hazards for each genetic subgroup are not well defined and, in Crump and Proctor's suggested procedure, would have to be assumed by OSHA.  For all of these reasons, OSHA believes that risk estimates resulting from the suggested analysis would be unreliable and potentially misleading.  OSHA therefore chose not to conduct this analysis. 
       
5. OSHA's Reanalysis of the NJH data set	
	
      As noted above, Dr. Stone reanalyzed the NJH data set in order to address concerns raised by Crump and Proctor (Document ID 1660).  This section summarizes the methods and results of the reanalysis.  Where OSHA's approach in the reanalysis differed from the approach taken in the preliminary statistical analysis, the reasons for OSHA's choices were discussed above and are briefly summarized here.
      OSHA's reanalysis uses a Cox proportional hazards model instead of the DTPHA used in the preliminary statistical analysis.  Like the DTPHA, a Cox proportional hazards model can accommodate time-dependent data; however, the Cox model has the advantage that it does not estimate different "baseline" rates of sensitization and CBD for different years.  The form of the Cox model is:  hZ(t) = h0(t) exp(Zβ), where  h0(t) is the baseline hazard at time t (i.e., Z=0), Z is a vector of possibly time-dependent predictors, and β is a vector of the corresponding coefficients, estimated using partial likelihood.  The exp(Zβ) term represents the HR.  Time-specific risk sets were constructed to accommodate the time-dependent exposures.  P-values were based on likelihood ratio tests (LRTs), with p-values <0.05 considered to be statistically significant.  As in the preliminary statistical analysis, Dr. Stone used fractional polynomials to check for possible nonlinearities in the exposure-response models, and checked the effects of age and smoking habits using the data on birth year and smoking (current, former, never) provided in the NJH data set.  
      Data on workers' estimated exposures and health outcomes through 2005 were included in the reanalysis.   The 1995 risk set (i.e., analysis of cases of sensitization and CBD identified in 1995) was excluded from all models in the reanalysis so as not to analyze long-standing (prevalent) cases of sensitization and CBD together with newly arising (incident) cases of sensitization and CBD.  In addition, Dr. Stone performed an analysis excluding workers hired before 1980 to examine the possible effects of uncertainty in exposures prior to the start of exposure sampling in 1980.
      Dr. Stone used the testing protocols provided in the Newman et al. study on the Cullman population to determine years in which each employee was scheduled to be tested, and excluded employees from the analysis for years in which they were not scheduled to be tested (Newman et al., 2001, Document ID 1354).  These models are labeled "Accounting for scheduled testing" in Tables 36 to 39 below.  Dr. Stone also fit models in the reanalysis using the previous approach for comparison.  The previous approach included all employees with no previously recorded sensitization or CBD in all time points of the analysis, without regard to whether they were likely to have been tested in any given year.  These models are labeled "Not accounting for scheduled testing" in Tables 36 to 39.  
      Tables 36 to 39 present modeling results of analyses using cumulative exposure and LTW average exposure as predictors.  As shown in these tables, LTW average exposure was more strongly associated with beryllium sensitization and cumulative exposure was more strongly associated with CBD.  OSHA therefore used the final model for LTW average exposure to estimate risk of sensitization at the preceding TWA PEL, the final TWA PEL, and several alternate TWA PELs the Agency considered.  Similarly, OSHA used the final model for cumulative exposure to estimate risk of CBD at the preceding, final, and alternate TWA PELs.  In calculating these risks, OSHA applied a small, fixed estimate of "baseline" risk (i.e., risk of sensitization or CBD among persons with no known exposure to beryllium), using rates suggested by Dr. Crump and Ms. Proctor and NIOSH (1% and 0.5% for sensitization and 0.5% for CBD) (Document ID 1660;1725). 
      OSHA's final risk estimates for sensitization and CBD, generated using the Cox proportional hazards models shown in Tables 37 and 38, are presented in Table 40 and Table 41, respectively, along with 95% confidence intervals based on the statistical uncertainty in the exposure coefficient for each model.  Since the Cox models do not estimate baseline risk and other variables were not included in the final models, the 95% confidence intervals based on the exposure coefficient uncertainty fully represent the statistical uncertainty in OSHA's risk estimates for beryllium sensitization and CBD. 
      
      
Results of the Beryllium Sensitization Analysis
      OSHA's reanalysis for beryllium sensitization is based on 18 cases.  The eight cases of sensitization identified in 1995 were excluded from the reanalysis because there is no information on when in the interval before 1995 these cases developed, as discussed previously.  As shown in Table 37, the HR for sensitization increased significantly with increasing LTW average exposure, without or with accounting for scheduled testing (HR=3.06 and HR = 2.92, respectively; p = 0.001 for both; note that numbers are rounded to the second decimal place). 
       As shown in Table 36 below, cumulative exposure was also a statistically significant predictor for beryllium sensitization, although it was not as strongly related to sensitization as LTW average exposure (HR=1.04, 95% CI 1.01  -  1.07 not accounting for testing; HR=1.04, 95% CI 1.00  -  1.07 accounting for testing).   Across most models in the reanalysis, p-values and standard errors were slightly larger when scheduled testing was taken into account, due to the smaller risk sets when employees not scheduled to be tested in a given year were excluded from that year.  
Table 36: Cox Proportional Hazards Model:  Cumulative Exposure and Sensitization
                                                                       Variable
                                                                   Hazard Ratio
                                                        95% confidence interval
                                                                        P-value
Not accounting for scheduled testing:
                                                                               
                                       
                                                                               
Cumulative Exposure (μg/m[3]-yrs)
                                                                           1.04
                                 1.01  -  1.07
                                                                           0.02
Model fit: LRT 3.94; Wald p=0.048
                                                                               
                                       
                                                                               
Accounting for scheduled testing:
                                                                               
                                       
                                                                               
                                             Cumulative Exposure (μg/m[3]-yrs)
                                                                           1.04
                                 1.00  -  1.07
                                                                           0.03
Model fit: LRT=3.73; Wald p=0.054
                                                                               
                                       
                                                                               
      
      
      
      
      Table 37: Cox Proportional Hazards Model:  LTW Average Exposure and Sensitization
                                                                       Variable
                                                                   Hazard Ratio
                                                        95% confidence interval
                                                                        P-value
Not accounting for scheduled testing:
                                                                               
                                       
                                                                               
                                                LTW Average Exposure (μg/m[3])
                                                                           3.06
                                 1.58  -  5.92
                                                                          0.001
Model fit: LRT= 8.50; Wald p=0.004
                                                                               
                                       
                                                                               
Accounting for scheduled testing:
                                                                               
                                       
                                                                               
                                               LTW Average Exposure (μg/m[3]) 
                                                                           2.92
                                 1.51  -  5.66
                                                                          0.001
Model fit: LRT=7.92; Wald p=0.005
                                                                               
                                       
                                                                               
      
      None of the analyses Dr. Stone performed to check for nonlinearities in exposure-response or effects of smoking, age, or hire prior to 1980 substantially impacted the results of the analyses for beryllium sensitization.  In tests for potential nonlinearities, Dr. Stone did not find evidence that a nonlinear exposure-response model would fit the NJH data set for beryllium sensitization better than the linear exposure-response used in the main analyses for either cumulative exposure or LTW average exposure.  Models that included smoking status (current, former, never) or age as a variable yielded HRs and confidence intervals virtually identical to those in models that did not consider smoking status.  For example, in the model that used the LTW average exposure metric and accounted for scheduled testing and age, the average exposure HR was 3.06 (95% CI 1.57 - 6.01).  In the model that used the LTW average exposure metric and accounted for scheduled testing and smoking the average exposure HR was 3.04 (95% CI 1.56 - 5.94).  When workers who were hired before 1980 were removed from the analysis (as suggested by Crump and Proctor in order to examine the possible effects of uncertainty in exposures prior to the start of exposure sampling in 1980), the LTW average exposure HR increased slightly to 3.15 (95% CI 1.39 - 7.21) when scheduled testing was accounted for, and to 3.23 (95% CI 1.41 - 7.45) when scheduled testing was not accounted for. 
      
Results of the CBD Analysis
      OSHA's reanalysis for CBD is based on 13 cases, six fewer than the number of cases in the analysis performed for the NPRM.  Six CBD cases had been diagnosed in 1995 and were excluded in the reanalysis.  As shown in Table 38, the HR for CBD increased significantly with increasing cumulative exposure, without or with accounting for scheduled testing (HR=1.04 for both; p=0.02 for both). 
      LTW average exposure was not as strong a predictor for CBD as cumulative exposure.  The HR for CBD increased somewhat with increasing LTW average exposure, with or without accounting for the scheduled testing, although this increase was not statistically significant at the 0.05 level (HR=2.34 not accounting for testing; HR=2.25 accounting for testing).   
    Table 38: Cox Proportional Hazards Model:  Cumulative Exposure and CBD
                                                                       Variable
                                                                   Hazard Ratio
                                                        95% confidence interval
                                                                        P-value
Not accounting for scheduled testing:
                                                                               
                                       
                                                                               
Cumulative Exposure (μg/m[3]-yrs)
                                                                           1.04
                                  1.01 - 1.08
                                                                           0.02
Model fit: LRT 3.97; Wald p=0.046
                                                                               
                                       
                                                                               
Accounting for scheduled testing:
                                                                               
                                       
                                                                               
Cumulative Exposure (μg/m[3]-yrs)
                                                                           1.04
                                  1.01 - 1.08
                                                                           0.02
Model fit: LRT=3.89; Wald p=0.049
                                                                               
                                       
                                                                               
                                       
    Table 39: Cox Proportional Hazards Model:  LTW Average Exposure and CBD
                                                                       Variable
                                                                   Hazard Ratio
                                                        95% confidence interval
                                                                        P-value
Not accounting for scheduled testing:
                                                                               
                                       
                                                                               
LTW Average Exposure (μg/m[3])
                                                                           2.34
                                  0.98 - 5.57
                                                                           0.06
Model fit: LRT= 2.94; Wald p=0.09
                                                                               
                                       
                                                                               
Accounting for scheduled testing:
                                                                               
                                       
                                                                               
                       LTW Average Exposure (μg/m[3]) 
                                                                           2.25
                                  0.94 - 5.35
                                                                           0.07
Model fit: LRT=2.71; Wald p=0.10
                                                                               
                                       
                                                                               
	
      In tests for potential nonlinearities, Dr. Stone did not find evidence that a nonlinear exposure-response model would fit the NJH data set for CBD better than the linear exposure-response used in the main analyses for either cumulative exposure or LTW average exposure.  Models that included smoking status (current, former, never) or age as a variable yielded HRs and confidence intervals virtually identical to those in models that did not consider smoking status or age.  For example, in the model that used the cumulative exposure metric and accounted for scheduled testing and smoking, the cumulative exposure HR was 1.04 (95% CI 1.01 - 1.08), which was identical to the results achieved when not accounting for smoking status or age.  In the model that used the cumulative exposure metric and accounted for scheduled testing and age, the cumulative exposure HR was 1.05 (95% CI 1.01 - 1.08).
	When workers who were hired before 1980 were removed from the analysis, eight cases of CBD remained in the data set.  The cumulative exposure HR dropped to slightly below one and was not statistically significant (HR 0.96, 95% CI 0.81 - 1.13, p = 0.6).  This result may partially reflect the latency required for CBD to develop.  As discussed in Section V, Health Effects, CBD can develop quickly (e.g., within a few years) or slowly (e.g., over decades); thus, excluding workers hired before 1980 means that some members of the study population would not have had sufficient time to develop CBD from their exposure.  
      The reduced time for CBD to develop, in combination with genetic risk factors, may explain the lack of significant exposure-response observed in this analysis.  It has been theorized that certain genetic traits may greatly increase both risk of CBD from low beryllium exposures and speed of progression from sensitization to CBD (see Section V, Health Effects).  If this is the case, some workers in the NJH data set would be expected to follow an exposure-response pattern of gradually increasing risk with gradually increasing exposure, and some would be expected to develop CBD from virtually any exposure.  OSHA finds some support for this explanation in the patterns of exposure and onset of disease among the eight cases of CBD diagnosed in workers hired after 1980.  Three had both rapid onset of CBD (1 year after hire) and very low cumulative exposures (0.115 to 0.138 μg/m[3]-years), suggesting membership in a genetically sensitive subpopulation.  Four cases of CBD developed in workers hired between 1981 and 1986 who were diagnosed with CBD between 16 and 19 years after hire, suggesting they were not members of a genetically sensitive subgroup.  These four had much higher cumulative exposures (1.5 to 8.7 μg/m[3]-years). 
      The higher the proportion of such a sensitive subpopulation among cases, the more difficult it is for a statistical model to detect an exposure-response relationship in the population as a whole.  By excluding workers hired before 1980, the analysis specifically excludes several workers with CBD that took many years to develop.  In contrast, using the pre-1980 hire cutoff does not exclude any cases of rapidly developed CBD.  The shortened development time for CBD in the models that exclude workers hired prior to 1980 thus may selectively increase the proportion of cases that occurred in genetically sensitive individuals, which could explain why a statistically significant exposure-response was observed in the main models but not in this analysis.
      
Risk Estimates for Beryllium Sensitization and CBD Based on OSHA's Reanalysis

	 OSHA used the Cox proportional hazards models, accounting for scheduled testing, to derive quantitative risk estimates for sensitization and CBD.  The LTW average exposure metric was used to derive estimates for sensitization because it showed the strongest association with sensitization in the reanalysis.  The cumulative exposure metric showed the strongest association with CBD in the reanalysis, and was therefore used to derive estimates for CBD. 
      As discussed previously, OSHA chose to use low fixed rates of "background" sensitization and CBD in its derivation of risk estimates from the Cox proportional hazards models using the Cullman data.  Table 40 summarizes the risk estimates for sensitization and the corresponding 95% confidence intervals using two different fixed "background" rates of sensitization, 1 percent and 0.5 percent.  Table 41 summarizes the risk estimates for CBD and the corresponding 95% confidence intervals using a fixed "background" rate of CBD of 0.5 percent.   The corresponding interval is based on the uncertainty in the exposure coefficient (i.e., the predicted values based on the 95% confidence limits for the exposure coefficient).  Since the Cox proportional hazards model does not estimate a baseline risk, this 95% interval fully represents statistical uncertainty in the risk estimates.


Table 40:  Predicted Cases of Sensitization per 1,000 Workers Exposed at Preceding, New, and Alternative PELs Based on Cox Proportional Hazards Model, LTW Average Exposure Metric, with Corresponding Interval based on the Uncertainty in the Exposure Coefficient. 1% and 0.5% baselines.  


                           Exposure Level (μg/m[3])
                      Estimated cases/1000, .5% baseline
                                    95% CI
                       Estimated cases/1000, 1% baseline
                                    95% CI
                                                                            2.0
                                                                               
                                     42.75
                                       
                                11.4  -  160.34
                                       
                                     85.49
                                       
                               22.79  -  320.69
                                       
                                                                            1.0
                                                                               
                                     14.62
                                       
                                7.55  -  28.31
                                       
                                     29.24
                                       
                                15.10  -  56.63
                                       
                                                                            0.5
                                                                               
                                     8.55
                                       
                                6.14  -  11.90
                                       
                                     17.10
                                       
                                12.29  -  23.80
                                       
                                                                            0.2
                                                                               
                                     6.20
                                       
                                 5.43  -  7.07
                                       
                                     12.39
                                       
                                10.86  -  14.15
                                       
                                                                            0.1
                                                                               
                                     5.57
                                       
                                 5.21  -  5.95
                                       
                                     11.13
                                       
                                10.42  -  11.89
                                       
      

      
Table 41:  Predicted Cases of CBD per 1,000 Workers Exposed at Preceding, New, and Alternative PELs Based on Cox Proportional Hazards Model, Cumulative Exposure Metric, with Corresponding Interval based on the Uncertainty in the Exposure Coefficient. 0.5% baseline.  


                               Exposure Duration

                                    5 years
                                   10 years
                                   20 years
                                   45 years
                           Exposure Level (μg/m[3])
Cumulative (μg/m[3]-yrs)
                             Estimated cases/1000
                                    95% CI
                                 μg/m[3]-yrs
                             Estimated cases/1000
                                    95% CI 
                                 μg/m[3]-yrs
                             Estimated cases/1000
                                    95% CI
                                 μg/m[3]-yrs
                             Estimated cases/1000
                                    95% CI
                                                                            2.0
                                     10.0
                                     7.55
                                  5.34-10.67
                                     20.0
                                     11.39
                                  5.70-22.78
                                     40.0
                                     25.97
                                  6.5-103.76
                                     90.0
                                    203.60
                                 9.02-4595.67
                                                                            1.0
                                      5.0
                                     6.14
                                   5.17-7.30
                                     10.0
                                     7.55
                                  5.34-10.67
                                     20.0
                                     11.39
                                  5.70-22.78
                                     45.0
                                     31.91
                                  6.72-151.59
                                                                            0.5
                                      2.5
                                     5.54
                                   5.08-6.04
                                      5.0
                                     6.14
                                   5.17-7.30
                                     10.0
                                     7.55
                                  5.34-10.67
                                     22.5
                                     12.63
                                  5.79-27.53
                                                                            0.2
                                      1.0
                                     5.21
                                   5.03-5.39
                                      2.0
                                     5.43
                                  5.07 - 5.82
                                      4.0
                                      5.9
                                   5.13-6.77
                                      9.0
                                     7.24
                                   5.30-9.89
                                                                            0.1
                                      0.5
                                      5.1
                                   5.02-5.19
                                      1.0
                                     5.21
                                   5.03-5.39
                                      2.0
                                     5.43
                                  5.07 - 5.82
                                      4.5
                                     6.02
                                   5.15-7.03
      
      The Cox proportional hazards model, used with the fixed "baseline" rates of 0.5 percent and 1 percent, predicted risks of sensitization totaling 43 and 86 cases per 1,000 workers, respectively, or 4.3 and 8.6 percent, at the preceding PEL of 2 μg/m[3].  The predicted risk of CBD is 204 cases per 1,000 workers, or 20.4 percent, at the preceding PEL of 2 μg/m[3], assuming 45 years of exposure (cumulative exposure of 90 μg/m[3]-yr).  OSHA notes that this level of cumulative exposure (90 μg/m[3]-yr) is substantially higher than the highest level of cumulative exposure of any worker in the NJH data set (61.86 μg/m[3]-yr); thus, the risk estimates for lower cumulative exposures may be more reliable (for example, 26 cases of CBD per 1,000 at 40 μg/m[3]-yr, which equals 20 years of exposure at the preceding TWA PEL).  The predicted risks of sensitization at the new PEL of 0.2 μg/m[3] are substantially lower, at 6 and 12 cases per 1,000 for the baselines of 0.5 percent and 1.0 percent, respectively.  The predicted risk of CBD is also substantially lower at the new PEL of 0.2 μg/m[3] (9 μg/m[3]-year), at 7 cases per 1,000 assuming 45 years of exposure.
      Due to limitations in the Cox analysis, including the small size of the data set, relatively limited exposure data from the plant's early years, constraints related to the size of the study, and limited follow-up time on many workers, OSHA must interpret the model-based risk estimates presented in Tables 40 and 41 with caution.  The Cullman study population is a relatively small group and can support only limited statistical analysis.  For example, its size precludes inclusion of multiple covariates in the exposure-response models and a two-stage exposure-response analysis to model both sensitization and the subsequent development of CBD within the subpopulation of sensitized workers cannot be performed.  The limited size of the NJH data set is characteristic of studies on beryllium-exposed workers in modern, low-exposure environments, which are typically small-scale processing plants (up to several hundred workers, with up to 20-30 cases).  However, these more recent studies, including Cullman, also have important strengths: they include workers hired after the institution of stringent exposure controls, and have extensive exposure sampling using full-shift personal lapel samples.  In contrast, older studies of larger populations tend to have higher exposures, less exposure data, and exposure data collected in short-term samples or outside of workers' breathing zones.   
      Another limitation of the Cullman data set, which is common to recent low-exposure studies, is the short follow-up time available for many of the workers.  While in some cases CBD has been known to develop in short periods (< 2 years), it more typically develops over a longer time period. Sensitization typically occurs in a shorter time frame, but new cases of sensitization have also been observed in workers exposed to beryllium for many years.  Because the data set is limited to individuals still working at the plant, the NJH data set cannot capture CBD occurring among workers who retire or leave the plant.  OSHA thus presumes that the data set does not fully represent the risk of sensitization and, in particular, CBD among workers exposed to beryllium at this facility. The short follow-up time is likely a significant source of uncertainty in the statistical analysis, which could lead to underestimation of risk in this population.
      Particle size, particle surface area, and beryllium compound solubility are believed to be important factors influencing the risk of sensitization and CBD among beryllium-exposed workers.  The workers at the Cullman machining plant were primarily handling poorly-soluble beryllium compounds, such as beryllium metal and beryllium metal/beryllium oxide composites; however, quantitative information on exposure to poorly-soluble and soluble forms of beryllium was not available.  Particle size distributions from a limited number of airborne beryllium samples collected just after the 1996 installation of engineering controls indicate worker exposure to a substantial proportion of respirable particulates (Kelleher et al., 2001, Document ID 1363).  There was no available particle size data for the 1980 to 1995 period, prior to installation of engineering controls, when total beryllium mass exposure levels were greatest. Particle size data was also lacking from 1998 to 2003 when additional control measures were in place and total beryllium mass exposures were lowest.  For these reasons, OSHA was not able to quantitatively account for the influence of particle size and solubility in developing the risk estimates based on the NJH data set.  However, it is not unreasonable to expect the sensitization and CBD experienced by this cohort to generally reflect the risk from exposure to beryllium that is relatively insoluble and contains a substantial proportion of respirable particles. 
      Additional uncertainty is introduced when extrapolating the quantitative estimates presented above to operations that process beryllium compounds with different solubility and particle characteristics than those encountered at the Cullman machining plant.  OSHA does not have sufficient information to quantitatively assess the degree to which the risks of beryllium sensitization and CBD based on the NJH data may be impacted in workplaces where different beryllium forms and processes are used.  However, this uncertainty does not alter OSHA's conclusions, based on the literature review and the prevalence analysis of the NJH data set, that there is significant risk at the preceding TWA PEL, new TWA PEL and the action level of 0.1 μg/m[3] (see the preamble to the final rule at Section VII, Significance of Risk).  The studies reviewed provide clear evidence of sensitization and CBD risk among workers exposed to a number of beryllium forms during different processes, such as beryllium machining, beryllium-copper alloy production, and beryllium ceramics production.  The Agency has determined that all of these forms of beryllium exposure create risks of sensitization and CBD among beryllium-exposed workers.
6. Conclusions
	OSHA investigated the exposure-response relationship for beryllium sensitization and CBD by analyzing a data set that NJH provided on workers at a beryllium machining facility in Cullman, Alabama.  OSHA's analysis of this data set showed positive exposure-response trends using multiple analytical approaches, including examination of beryllium sensitization and CBD prevalence by exposure categories and a Cox proportional hazards modeling approach.  
	The prevalence analysis shows that cases of beryllium sensitization and disease occur at average levels of exposure below the preceding TWA PEL and at the final TWA PEL (see Table 7).  The lowest prevalence of beryllium sensitization (2.2%) and CBD (1%) was observed among workers with LTW average exposure levels in the lowest quartile of exposure (< 0.08 μg/m[3]).  In the quartiles of LTW average exposure closest to the preceding PEL of 2 μg/m[3], the prevalence of beryllium sensitization and CBD was high.  Of workers with LTW average exposures between 0.51 μg/m[3] and 2.15 μg/m[3], 15.4 percent were sensitized to beryllium, about two-thirds of whom were diagnosed with CBD (10.3% of workers in the quartile).  High prevalences of beryllium sensitization (> 7.8%) and CBD (> 5.5%) were observed among workers in the second and third quartiles of exposure as well (exposures between 0.08 μg/m3  and 0.51 μg/m[3]).  
	OSHA's final statistical analysis of the NJH data set found increasing risk of sensitization with both cumulative exposure and average exposure.  OSHA also found an increasing risk of CBD with cumulative exposure and number of years exposed, but did not find a statistically significant relationship between CBD and average exposure.  OSHA used the LTW average model results to estimate HRs and risk of sensitization at the preceding TWA PEL of 2 μg/m3 and the new TWA PEL of 0.2 μg/m[3] and below and used the cumulative exposure model results to estimate HRs and risk of CBD at the same exposure levels.  For the CBD model, OSHA also estimated risk for various time periods of exposure, from 5 years to a working lifetime of 45 years.
      The Cox proportional hazards model, used with the fixed "baseline" rates of 0.5 percent and 1 percent, predicted risks of sensitization totaling 43 and 86 cases per 1,000 workers, respectively, or 4.3 and 8.6 percent, at the preceding PEL of 2 μg/m[3].  The predicted risks of sensitization at the new PEL of 0.2 μg/m[3] are substantially lower, at 6 and 12 cases per 1,000 for the baselines of 0.5 percent and 1.0 percent, respectively.  The predicted risk of CBD is 204 cases per 1,000 workers, or 20.4 percent, at the preceding PEL of 2 μg/m[3], assuming 45 years of exposure (cumulative exposure of 90 μg/m[3]-yr), and 26 cases per 1,000 workers at the preceding PEL, assuming 20 years of exposure (cumulative exposure of 40 μg/m[3]-yr).  The predicted risk of CBD is also substantially lower at the new PEL of 0.2 μg/m[3] (9 μg/m[3]-year), at 6 cases per 1,000, assuming 45 years of exposure and using a baseline of 0.5 percent.  Because the statistical models used to derive these values were based on a small data set, OSHA believes these risk estimates must be viewed with caution.  Nevertheless, the statistical exposure-response models support the qualitative conclusions of the literature analysis discussed in the preamble to the final rule (Section VII, Significance of Risk) and the prevalence analysis of the NJH data that significant risk of sensitization and CBD exists at the preceding PEL of 2.0 μg/m[3], and that this risk can be lessened by lowering workers' exposure levels.  
