In 1998, the National Institute for Occupational Safety and Health recommended that exposures to any type of aerosolized metalworking fluids be limited to 0.4 mg/m3 of thoracic particulate mass (particulate matter of diameter <10 μm, PM10) as a time-weighted average concentration.1 This recommendation aimed to reduce respiratory disorders associated with metalworking fluid exposure. Since then, others have argued that a lower limit is needed2,3 and that various fluid types may require distinct limits because of their different compositions3; however, it is unclear what those limits should be. Furthermore, finer particles are now thought to be more relevant for health effects.4 We propose a public health framework that facilitates comparison of the effects of various interventions, here representing possible occupational exposure limits for PM3.5 composed of metalworking fluids. The framework relies on g-estimation to adjust for the healthy worker survivor effect. As a motivating example for this new approach, which is general enough for use in other contexts, we present an application to study the interventions’ effects on mortality from chronic obstructive pulmonary disease (COPD).
COPD is typically characterized by progressive and irreversible limitation of airflow caused by an inflammatory response to noxious particles and gases.5 The recent Global Burden of Disease 2010 report estimated that COPD accounted for 3% of all disability-adjusted life years lost worldwide.6,7 While cigarette smoking is the most common cause, risk of the disease also increases with exposures to occupational and environmental agents, including vapors, gases, dust, and fumes in the workplace.8 Fifteen to 20% of the burden of COPD is attributable to occupational exposures.9,10 Better understanding of exposure–response relationships for workplace hazards and COPD would improve efforts to design effective preventive strategies.
There are several reasons why studies of suspected environmental or occupational causes of COPD are particularly challenging. First, in most occupational studies, information on physician diagnosis of COPD is unavailable. Furthermore, industries associated with COPD risk often involve exposures to a mixture of agents potentially capable of causing airway inflammation.9 Moreover, unlike lung cancer (another multifactorial disease with a long latency), COPD is characterized by an extended period of worsening pulmonary function with increasing respiratory symptoms. This emergence of symptoms before the outcome of interest leads to the healthy worker survivor effect (discussed in detail below), which can obscure the true relationship between occupational exposure and disease.
Aerosolized metalworking fluids used in machining operations have been linked to airway diseases such as asthma11 and hypersensitivity pneumonitis.12 They have also been associated with acute change in pulmonary function over the work shift13–15 and with respiratory symptoms.15–17 Straight (oil-based) fluids contain polycyclic aromatic hydrocarbons attached to PM3.5, which have been shown to cause respiratory health problems.18 Use of this fluid type has declined during the past 100 years because industrial demand for higher speed machining spurred the development of water-based fluids with reduced friction and increased cooling properties. Water-based fluids, both soluble and synthetic, contain emulsifiers, biocides, corrosion inhibitors, and other respiratory irritants.19
Evidence for chronic effects on pulmonary function is less clear; results from two cross-sectional studies of automobile manufacturing workers are somewhat inconsistent.20,21 Analyses comparing a large cohort mortality study of autoworkers with the general population did not find excess risk of nonmalignant respiratory mortality.22,23 No study has reported a relationship between a quantitative measure of metalworking fluid exposure and COPD incidence, prevalence, or mortality.
One possible reason for the lack of studies showing an association of metalworking fluids with COPD, despite its biologic plausibility, is the healthy worker survivor effect.23,24 This bias occurs in longitudinal studies of cumulative exposure when there are shared causes of exposure and outcome, whether the exposure affects the outcome or not. Typically, less-healthy workers reduce their exposure such that the healthiest workers accumulate the most exposure, leading to estimates suggesting that the exposure is only weakly harmful, neutral, or even protective.23 Two of the most common ways for workers to reduce their exposure are to take time off work or to terminate employment. Employment status and time off work are therefore important confounders of the effect of exposure on the outcome, but proper adjustment is not straightforward because these factors change over time and can be affected by prior exposure.
We will examine the effects of lowering the limit for metalworking fluid exposure to various levels on COPD mortality in the United Autoworkers–General Motors cohort, accounting for the healthy worker survivor effect. As a tool to understand healthy worker survivor bias in this context, we present Figure 1 as a causal diagram representing our occupational study. The exposure E is PM3.5 composed of metalworking fluids (straight oil-based, soluble, or synthetic), the outcome D is COPD mortality, and our confounders L and W represent time off work and employment status. The unmeasured variable U is health status, and subscripts index the time intervals; this is a simplified diagram showing only two time intervals and omitting baseline covariates. (For introductions to causal diagrams, see Greenland et al25 and Hernán et al.26)
Results of our analysis will be biased if we do not adjust for L and W because there are unblocked noncausal paths from exposures at times 1 and 2 to COPD death passing through the shared cause U. However, if we use a traditional regression model to adjust for L and W, we introduce a noncausal association between the earlier exposures and the outcome through the colliders L and W. Figure 1 illustrates that even when adjustment is necessary, traditional regression methods cannot adjust for time-varying confounders affected by prior exposure without creating bias.25,26
Therefore, we must use a method that specifically addresses this problem, such as one of those proposed by Robins.27–32 We implement g-estimation of a structural accelerated failure time model. This method has not been applied widely,28–35 and its first application to an occupational study was quite recent.36 In that article, we reanalyzed the autoworkers mortality study using g-estimation to address healthy worker survivor bias in an analysis of duration of exposure and mortality from selected causes, including COPD.36 The positive finding for duration of exposure to straight fluids and COPD mortality provides the strongest evidence to date that metalworking fluids may cause COPD. The absence of a quantitative exposure metric in that report provides the starting point for the present study.
To examine the potential effects of metalworking fluids on COPD mortality, we have combined a quantitative approach to exposure with g-estimation to address healthy worker survivor bias. Rather than estimating the etiologic strength of effect, we framed our research question as an evaluation of a series of interventions. This novel approach was designed to determine how much the burden of COPD mortality could have been reduced in the autoworkers cohort by lowering the occupational exposure limit for PM3.5 composed of metalworking fluids.
The cohort consists of workers at three General Motors plants in Michigan, USA, who were hired between 1938 and 1982 and stayed at least 3 years.37 For our analysis, follow-up began 3 years after hire and continued regardless of employment status until the earliest of three possible ends of follow-up: death, the end of 1994, or age 95 years. Death from COPD was defined by International Classification of Disease, revision 9 codes 490–496.
Exposure and time off work were determined by combining employment records with a job-exposure matrix developed in an extensive exposure assessment.37,38 Exposures to three types of metalworking fluids (straight, soluble, and synthetic) were measured. In separate analyses, each fluid type was analyzed as the exposure of interest, with the other two treated as potential confounders.
The approach we propose entails creating a series of binary exposure variables, each equal to 0 if quantitative exposure is below a specified cutoff value and 1 if exposure is above the cutoff. The analyses assume that the effect is constant within each category, regardless of the actual level of exposure. Each binary exposure variable corresponds to an intervention in which annual average daily exposure never exceeds the cutoff. G-estimation is then run separately for each binary exposure variable, and the results are transformed to an effect measure that allows us to examine the potential effects of a series of such interventions.
For each binary exposure variable, our structural accelerated failure time model relates the observed survival time and the observed duration of exposure above the cutoff to the counterfactual survival time that would have been observed if never exposed (above the cutoff). This relationship is quantified by an unknown coefficient to be estimated (see eAppendix; http://links.lww.com/EDE/A774).
Unlike conventional models, g-estimation can adjust correctly for time-varying confounders affected by prior exposure. In our example, employment status and leaves of absence are variables that, in standard analyses, would result in healthy worker survivor bias whether or not they were included in the regression models predicting the outcome. In contrast, g-estimation adjusts for confounders by including them in a model predicting the exposure. All studies, regardless of statistical method, require the assumption that we have measured all confounders (without error). This means that within strata of the measured confounders, the survival time if never exposed must be statistically independent of observed exposure. G-estimation uses optimization methods to estimate the value of the coefficient that achieves this statistical independence. For more details, see eAppendix (http://links.lww.com/EDE/A774) and references.32,33,36
For each person who experienced the outcome of interest (“case”), the coefficient estimate for a cutoff is transformed into an estimate of that person’s survival time under the corresponding intervention (see eAppendix for details; http://links.lww.com/EDE/A774). The difference between this counterfactual survival time and the observed survival time represents the estimated number of years of that person’s life that would have been saved under the intervention. The sum (over all cases) of these years of life saved represents an estimate of the burden of disease attributable to the failure to enforce such an occupational exposure limit in this cohort. Unexposed workers would have been unaffected by any of the interventions and therefore contribute zero to the total years of life that could have been saved (among the cases) by enforcing a lower occupational exposure limit. (However, they do contribute person-time to the analysis in the g-estimation step.)
Because this effect measure compares observed outcomes (which are the same for all cutoffs) with counterfactual outcomes under each intervention, we may compare results from various cutoffs. Furthermore, this metric is based on survival time, which is the quantity modeled, so it does not require the additional assumptions needed to convert our results to hazard ratios.
Quantitative annual exposures to each fluid type were transformed into several binary variables, each equal to 0 if the exposure was less than or equal to a specific cutoff value, and 1 otherwise. The cutoffs were selected based on the exposure distributions: a maximum cutoff was chosen near the 75th percentile of the nonzero annual exposures for each fluid type. At least 7 cutoffs were assessed for each fluid type, including a cutoff of 0 to represent a ban. For example, we estimated the total years of life that would have been saved among the cases if annual average daily exposure to straight fluids had never been permitted to exceed a limit of 0.15 mg/m3. The analysis was repeated for 0.1 mg/m3 and for limits ranging from 0 to 0.05 mg/m3 in intervals of 0.01. For soluble fluids, cutoffs ranged from 0 to 0.3 mg/m3 in intervals of 0.05 mg/m3. For synthetic fluids, cutoffs included 0.05 mg/m3 and values from 0 to 0.025 mg/m3 in intervals of 0.005.
Our analysis included the following covariates in the pooled logistic models for exposure: current age in years (linear and quadratic terms), race (white or black), sex, plant (I, II, or III), an indicator for calendar year (before or after 1970, when exposures dramatically decreased),37 prior exposure history (whether or not exposed above the cutoff in previous years), exposures to the other two fluid types (whether or not exposed in previous years), and intermittent time off work (percent of each year as a linear term). G-estimation also requires adjusting for a variable representing each person’s maximum observable survival time (equal to the length of time from cohort entry to the end of 1994 or to the worker’s 95th birthday, whichever occurs sooner; this quantity is known at cohort entry). As exposure never occurs in those who are not at work, this model was fitted only on the actively employed person-time, thus controlling for employment status by conditioning on it. However, outcomes are not censored at termination of employment: the survival times in the structural accelerated failure time model do include post-employment time.39
We adjusted for loss to follow-up (<4%) and censoring by death from competing risks (<24%) using inverse-probability-of-censoring weights (see eAppendix; http://links.lww.com/EDE/A774).33 Weights ranged from 0.88 to 1.45.
Confidence intervals were obtained by running 200 bootstraps. For each cutoff, if all bootstraps produced an estimate, then the 2.5th and 97.5th percentiles were chosen as the lower and upper limits of the 95% confidence interval. If at least one sample, but fewer than 5% of them, failed to produce estimates within the search interval, then the minimum and maximum estimates were chosen as the lower and upper limits to construct a conservative 95% confidence interval. See eAppendix for further details (http://links.lww.com/EDE/A774).
Analyses were conducted in SAS 9.3 (SAS Institute Inc., Cary, NC). The study involved reanalysis of existing data and was approved by the UC Berkeley Committee for the Protection of Human Subjects.
Table 1 presents the demographic characteristics of the cohort, which was mostly male and white, with average employment duration of about 8 years less than the average follow-up length.
Table 2 presents summary statistics on the distributions of the three metalworking fluid exposures. All three distributions were skewed with long tails to the right. Of the 308 COPD cases, 129 were never exposed to straight fluids, 37 were never exposed to soluble fluids, and 206 were never exposed to synthetic fluids. Only 28 were never exposed to any of the fluids.
Figure 2 shows the total years of life saved among COPD cases resulting from lowering the occupational exposure limit for annual average daily exposure to PM3.5 composed of straight fluids to selected levels (cutoffs). The greatest effect is achieved by preventing any exposure; as the limit increases, the effect diminishes, following a monotonic trend.
Figure 3 shows the corresponding effects of enforcing an occupational exposure limit to regulate soluble fluids. These effects are higher than those estimated for the other two fluid types, and there is a downward trend as the limit increases. Interventions to prohibit exposures above 0.2 mg/m3 are estimated to offer no benefit.
Figure 4 displays the effects of enforcing an occupational exposure limit to regulate synthetic fluids. The greatest effect occurs under the intervention in which no worker is ever exposed to synthetic fluids. The downward trend is similar to those observed for the other two fluids, although power was low.
Figure 5 shows estimates for all three fluids using common scales on the x- and y-axes so that effects of enforcing occupational exposure limits on the three fluid types can be compared visually.
G-estimation of a structural accelerated failure time model, with its control for healthy worker survivor bias, suggests that under interventions to limit occupational exposure to straight, soluble, or synthetic metalworking fluids, the estimated number of years of life saved increases as the occupational exposure limit decreases toward 0, with a convincing trend. There is no evidence supporting a threshold under which exposure is safe with regard to COPD.
The most beneficial intervention would be a ban on soluble fluids, which seems to have the potential to save an estimated 1550 years of life among the people who died of COPD. Note that this is due largely to the fact that nearly all of the cases had experienced at least some exposure to soluble fluids. These fluids are composed of 30–85% of oil diluted with water, as well as a variety of chemicals, biocides, and other additives.1 This heterogeneity may help explain the wide confidence intervals for the limits over 0.
Power to assess the effects of interventions on synthetic fluids was limited by the fact that only about one third (102) of the COPD cases were ever exposed to synthetic fluids. These fluids were not in widespread use during much of the study period and merit further study.
No previous studies using traditional analysis have examined COPD in relation to occupational exposure to quantitative measures of metalworking fluids. However, a relationship between metalworking fluids and COPD is indirectly supported by cross-sectional studies of several respiratory outcomes. Various measures of pulmonary function have been modestly associated with straight fluids in a subset of the autoworkers cohort21 and with soluble fluids in another study of autoworkers.20 Workers exposed to straight and synthetic fluids also had increased risk of respiratory symptoms relative to assembly workers without direct exposure.17,21 Synthetic metalworking fluid exposure in the 2 years before diagnosis was associated with adult-onset asthma.11 In a Finnish study of machine workers, chronic bronchitis was associated with higher levels of metalworking fluids in the general workshop air and with longer duration of machining work. However, evidence was less convincing for air in the breathing zone of machine workers, where one might expect stronger relationships with symptoms.16 One possible explanation for the lack of stronger evidence is the healthy worker survivor effect.
The strongest evidence for an effect of metalworking fluid exposure on COPD is provided by Chevrier et al,36 who report a striking difference between the null results from standard regression models and the positive findings from g-estimation in this autoworkers cohort, suggesting that the healthy worker survivor effect was biasing the standard results. G-estimation results indicated that a hypothetical 5-year duration of exposure reduced COPD survival times to 90% of unexposed survival time (95% confidence interval = 83–96%); the analysis used a binary exposure variable equivalent to the one we used for an occupational exposure limit of 0. The present article expands on that work by estimating a more policy-relevant effect measure and by taking into account (albeit relatively crudely) the quantitative exposure.
When exposures above and below a series of binary cutoffs are analyzed using standard statistical methods, estimates from different cutoffs cannot be compared because they do not share a reference group. Using g-estimation within our public health framework avoids that problem by comparing outcomes under a series of interventions with the observed outcomes, which do not depend on the cutoff. A major strength of the approach we introduce here is that the effect estimates obtained for various occupational exposure limits were therefore comparable with one another.
The true value of the effect measure estimated (years of life that would have been saved by enforcing an occupational exposure limit) depends on (1) the incidence of the outcome in the study population during follow-up, (2) the distribution of exposure among the cases, and (3) the shape of the exposure–response curve. Thus, the analysis does not address a purely etiologic question, but goes a step further, taking the novel approach of estimating the burden of disease that could have been avoided by enforcing a lower exposure limit in the cohort under study. To our knowledge, no previous application of g-estimation has reported estimates of this effect measure.
Our analysis is based on several assumptions, violations of which could affect the validity of our results. First, as in all studies, the analysis depends on the assumption that all relevant confounders have been measured (without error). We were unable to take into account job transfers to lower exposed jobs or the use of protective equipment (such as respirators) that might help protect those in highly exposed jobs. There may therefore be some residual confounding because of these, although the bias would likely be downward. In addition, smoking data were unavailable, so we cannot rule out confounding by this important risk factor for COPD. However, a quantitative evaluation of the bias potentially introduced by uncontrolled confounding by alcohol and tobacco use in this cohort indicated that it would probably not exceed 20%.40 In fact, the nonexposed workers smoked more than the exposed in this cohort,17 so confounding by smoking would probably bias the effect estimates toward the null.
Furthermore, we assume that all of our models are correct. This is a particularly strong assumption for the structural accelerated failure time model, which assumes that all those who died of COPD would have died of COPD under every possible exposure scenario; exposure merely affects the time at which COPD death occurs. This modeling assumption may be unrealistic.
Positivity, or experimental treatment assignment, refers to the assumption that all nonempty strata of covariates include both exposed and unexposed people. This assumption is structurally violated because workers who are no longer actively employed cannot be exposed. This is why marginal structural models using inverse probability of treatment weighting cannot be used for this study. However, g-estimation of our structural model does not require this assumption.41
We also assume consistency: that each person’s counterfactual survival time under their observed exposure is equal to their observed survival time. This holds when the exposure is subject to well-defined interventions. For example, ways for a workplace to comply with an occupational exposure limit include introducing better ventilation systems, enforcing the use of protective equipment, and replacing fluids with other types. We assume that the effects of such different interventions are equivalent if they correspond to the same limit.
A related assumption is that, for each cutoff, exposure being above or below that limit is the relevant exposure classification and that a worker’s actual annual average daily exposure level within those categories is not relevant. By studying the effects of interventions based on several cutoffs, we hope to have minimized any bias caused by violations of this exposure classification assumption. (A related technical caveat is briefly described in the eAppendix; http://links.lww.com/EDE/A774.)
Many occupational cohort studies aim to answer an etiologic question using quantitative exposure data to estimate an exposure–response relationship. Instead, we chose a novel framework that focuses on the more policy-relevant issue of envisioning possible interventions and estimating their effects. G-estimation of a structural accelerated failure time model is well suited to this type of question and has the added benefit of avoiding healthy worker survivor bias. Such a bias is likely to be particularly relevant in a study of COPD, the symptoms of which might cause many workers with early stages of this disease to reduce their exposures by leaving work or taking time off. Our application of this method indicates that lowering occupational exposure limits as much as possible for straight, soluble, or synthetic fluids could have saved considerable years of life for those who died of COPD. Furthermore, interventions to reduce exposure to soluble fluids would have affected more workers and had a greater effect than intervening on the other fluids.
The public health framework we introduce in this article could be applied in other (non-occupational) contexts. Its estimates of the effects of possible interventions are easy to interpret and could help guide policymaker decisions.
1. Department of Health and Human Services PHS, Centers for Disease Control and Prevention, National Institute for Occupational Safety and Health (NIOSH). . Occupational exposure to metalworking fluids, March 1998. Report No. 98–116.
2. Mirer FE. New evidence on the health hazards and control of metalworking fluids since completion of the OSHA advisory committee report. Am J Ind Med. 2010;53:792–801
3. Park D. The occupational exposure limit for fluid aerosol generated in metalworking operations: limitations and recommendations. Saf Health Work. 2012;3:1–10
4. Brown JS, Gordon T, Price O, Asgharian B. Thoracic and respirable particle definitions for human health risk assessment. Part Fibre Toxicol. 2013;10:12
5. . Global Initiative for Chronic Obstructive Lung Disease (GOLD). Global strategy for the diagnosis, management, and prevention of COPD; http://www.goldcopd.org
. 2014; p. 19.
6. Lozano R, Naghavi M, Foreman K, et al. Global and regional mortality from 235 causes of death for 20 age groups in 1990 and 2010: a systematic analysis for the Global Burden of Disease Study 2010. Lancet. 2012;380:2095–2128
7. Murray CJ, Vos T, Lozano R, et al. Disability-adjusted life years (DALYs) for 291 diseases and injuries in 21 regions, 1990-2010: a systematic analysis for the Global Burden of Disease Study 2010. Lancet. 2012;380:2197–2223
8. Eisner MD, Anthonisen N, Coultas D, et al.Committee on Nonsmoking COPD, Environmental and Occupational Health Assembly. An official American Thoracic Society public policy statement: novel risk factors and the global burden of chronic obstructive pulmonary disease. Am J Respir Crit Care Med. 2010;182:693–718
9. Balmes J, Becklake M, Blanc P, et al.Environmental and Occupational Health Assembly, American Thoracic Society. American Thoracic Society statement: occupational contribution to the burden of airway disease. Am J Respir Crit Care Med. 2003;167:787–797
10. Blanc PD, Torén K. Occupation in chronic obstructive pulmonary disease and chronic bronchitis: an update. Int J Tuberc Lung Dis. 2007;11:251–257
11. Eisen EA, Holcroft CA, Greaves IA, Wegman DH, Woskie SR, Monson RR. A strategy to reduce healthy worker effect in a cross-sectional study of asthma and metalworking fluids. Am J Ind Med. 1997;31:671–677
12. Kreiss K, Cox-Ganser J. Metalworking fluid-associated hypersensitivity pneumonitis: a workshop summary. Am J Ind Med. 1997;32:423–432
13. Kennedy SM, Greaves IA, Kriebel D, Eisen EA, Smith TJ, Woskie SR. Acute pulmonary responses among automobile workers exposed to aerosols of machining fluids. Am J Ind Med. 1989;15:627–641
14. Robins T, Seixas N, Franzblau A, et al. Acute respiratory effects on workers exposed to metalworking fluid aerosols in an automotive transmission plant. Am J Ind Med. 1997;31:510–524
15. Kennedy SM, Chan-Yeung M, Teschke K, Karlen B. Change in airway responsiveness among apprentices exposed to metalworking fluids. Am J Respir Crit Care Med. 1999;159:87–93
16. Jaakkola MS, Suuronen K, Luukkonen R, et al. Respiratory symptoms and conditions related to occupational exposures in machine shops. Scand J Work Environ Health. 2009;35:64–73
17. Greaves IA, Eisen EA, Smith TJ, et al. Respiratory health of automobile workers exposed to metal-working fluid aerosols: respiratory symptoms. Am J Ind Med. 1997;32:450–459
18. Gale SL, Noth EM, Mann J, Balmes J, Hammond SK, Tager IB. Polycyclic aromatic hydrocarbon exposure and wheeze in a cohort of children with asthma in Fresno, CA. J Expo Sci Environ Epidemiol. 2012;22:386–392
19. Byers JP Metalworking Fluids. 20062nd ed London, UK Taylor & Francis
20. Kriebel D, Sama SR, Woskie S, et al. A field investigation of the acute respiratory effects of metal working fluids. I. Effects of aerosol exposures. Am J Ind Med. 1997;31:756–766
21. Eisen EA, Smith TJ, Kriebel D, et al. Respiratory health of automobile workers and exposures to metal-working fluid aerosols: lung spirometry. Am J Ind Med. 2001;39:443–453
22. Eisen EA, Tolbert PE, Monson RR, Smith TJ. Mortality studies of machining fluid exposure in the automobile industry I: a standardized mortality ratio analysis. Am J Ind Med. 1992;22:809–824
23. Eisen EA, Picciotto S, Robins JMEl-Shaarawi AH, Piegorsch WW. Healthy worker effect in occupational studies. In: Encyclopedia of Environmetrics. 20122nd ed Chichester, UK John Wiley & Sons:1269–1272
24. Radon K, Goldberg M, Becklake M. Healthy worker effect in cohort studies on chronic bronchitis. Scand J Work Environ Health. 2002;28:328–332
25. Greenland S, Pearl J, Robins JM. Causal diagrams for epidemiologic research. Epidemiology. 1999;10:37–48
26. Hernán MA, Hernández-Díaz S, Robins JM. A structural approach to selection bias. Epidemiology. 2004;15:615–625
27. Robins JM. A new approach to causal inference in mortality studies with sustained exposure periods—application to control of the healthy worker survivor effect. Mathematical Modelling. 1986;7:1393–1512
28. Robins J. Estimation of the time-dependent accelerated failure time model in the presence of confounding factors. Biometrika. 1992;79:321–334
29. Robins J, Tsiatis AA. Semiparametric estimation of an accelerated failure time model with time-dependent covariates. Biometrika. 1992;79:311–319
30. Robins JM. Correcting for non-compliance in randomized trials using structural nested mean models. Commun Stat Theory. 1994;23:2379–2412
31. Robins JMAndersen PK, Keiding N. Structural nested failure time models. Survival Analysis. 1998 Chichester, UK John Wiley & Sons:4372–4389
32. Robins JM, Blevins D, Ritter G, Wulfsohn M. G-estimation of the effect of prophylaxis therapy for Pneumocystis carinii pneumonia on the survival of AIDS patients. Epidemiology. 1992;3:319–336
33. Hernán MA, Cole SR, Margolick J, Cohen M, Robins JM. Structural accelerated failure time models for survival analysis in studies with time-varying treatments. Pharmacoepidemiol Drug Saf. 2005;14:477–491
34. Joffe MM, Yang WP, Feldman H. G-estimation and artificial censoring: problems, challenges, and applications. Biometrics. 2012;68:275–286
35. Shinozaki T, Matsuyama Y, Iimuro S, et al. Effective prevention of cardiovascular disease and diabetes-related events with atorvastatin in Japanese elderly patients with type 2 diabetes mellitus: adjusting for treatment changes using a marginal structural proportional hazards model and a rank-preserving structural failure time model. Geriatr Gerontol Int. 2012;12(Suppl 1):88–102
36. Chevrier J, Picciotto S, Eisen EA. A comparison of standard methods with g-estimation of accelerated failure-time models to address the healthy-worker survivor effect: application in a cohort of autoworkers exposed to metalworking fluids. Epidemiology. 2012;23:212–219
37. Hallock MF, Smith TJ, Woskie SR, Hammond SK. Estimation of historical exposures to machining fluids in the automotive industry. Am J Ind Med. 1994;26:621–634
38. Woskie SR, Virji MA, Hallock M, Smith TJ, Hammond SK. Summary of the findings from the exposure assessments for metalworking fluid mortality and morbidity studies. Appl Occup Environ Hyg. 2003;18:855–864
39. Picciotto S, Brown DM, Chevrier J, Eisen EA. Healthy worker survivor bias: implications of truncating follow-up at employment termination. Occup Environ Med. 2013;70:736–742
40. Kriebel D, Zeka A, Eisen EA, Wegman DH. Quantitative evaluation of the effects of uncontrolled confounding by alcohol and tobacco in occupational cancer studies. Int J Epidemiol. 2004;33:1040–1045
41. Robins JM, Hernán MAFitzmaurice G, Davidian M, Verbeke G, Molenberghs G. Estimation of the causal effects of time-varying exposures. Longitudinal Data Analysis. 2009 New York Chapman & Hall/CRC:553–599