|Year : 2017 | Volume
| Issue : 1 | Page : 2-8
Healthy worker effect phenomenon: Revisited with emphasis on statistical methods – A review
Ritam Chowdhury1, Divyang Shah2, Abhishek R Payal3
1 Department of Biostatistics, Harvard School of Public Health, Boston, Massachusetts; Department of Global Health, Rollins School of Public Health, Emory University, Atlanta, Georgia, USA
2 Health and Medical Services, Larsen and Toubro Limited, Mumbai, Maharashtra, India
3 Massachusetts Eye and Ear Infirmary, Harvard Medical School, Boston, Massachusetts, USA
|Date of Web Publication||13-Dec-2017|
Dr. Ritam Chowdhury
677 Huntington Avenue, Boston, Massachusetts – 02115
Source of Support: None, Conflict of Interest: None
Known since 1885 but studied systematically only in the past four decades, the healthy worker effect (HWE) is a special form of selection bias common to occupational cohort studies. The phenomenon has been under debate for many years with respect to its impact, conceptual approach (confounding, selection bias, or both), and ways to resolve or account for its effect. The effect is not uniform across age groups, gender, race, and types of occupations and nor is it constant over time. Hence, assessing HWE and accounting for it in statistical analyses is complicated and requires sophisticated methods. Here, we review the HWE, factors affecting it, and methods developed so far to deal with it.
Keywords: Causal inference, healthy worker effect, occupational epidemiology, selection bias
|How to cite this article:|
Chowdhury R, Shah D, Payal AR. Healthy worker effect phenomenon: Revisited with emphasis on statistical methods – A review. Indian J Occup Environ Med 2017;21:2-8
|How to cite this URL:|
Chowdhury R, Shah D, Payal AR. Healthy worker effect phenomenon: Revisited with emphasis on statistical methods – A review. Indian J Occup Environ Med [serial online] 2017 [cited 2018 Oct 22];21:2-8. Available from: http://www.ijoem.com/text.asp?2017/21/1/2/220697
| Introduction|| |
“Healthy worker” effect (HWE) is a special type of selection bias, typically seen in observational studies of occupational exposures with improper choice of comparison group (usually general population). This phenomenon has been documented in several studies and methods have been developed to tackle this issue. In this paper, we will review the historical background that led to the discovery of this phenomenon, factors associated with this phenomenon, and summarize epidemiological and statistical methods to deal with this issue.
| History of Healthy Worker Effect|| |
The term HWE was first coined by McMichael  in 1976, who defined it as “the consistent tendency of the actively employed to have a more favorable mortality experience than the population at large.”He found that when standardized mortality ratios (SMRs) are calculated with the general population as reference, the SMR tends to underestimate the mortality experience of the occupational population. He further found that this “effect” did not affect all groups equally even in the same population and recommended that “allowances needed to be made for this unequal effect in different age groups, races, causes for death, elapsed time periods of observation and even different work groups.” Thus, in his paper McMichael not only coined the term but also provided clues to statistical methods that needed to be developed to deal with this issue.
However, this effect had been reported almost a century earlier by William Ogle  in a letter to the Registrar General of Births, Deaths, and Marriages. He observed lower mortality rates in certain occupations requiring hard labor, contrary to what was expected. He speculated accurately that this was due to self-selection by individuals capable of performing these tasks which was related to these observations on mortality. Subsequent studies in the interim period observed the effects of HWE without accounting for the discrepancy. Among them is the famous study by Doll et al. who observed the phenomenon among gas workers. In their landmark paper, Doll et al. estimated the age-adjusted SMR among coal-exposed gas workers versus the general population. They found, contrary to their expectations, that SMR among the high-risk population of gas workers was less than 100 or the exposure to coal was found to be protective in their study, as compared with the general population. However, when they compared the occupationally exposed workers divided into groups as per their exposure status [those with heavy exposure in carbonizing plants (class A), intermittent exposure or exposure to conditions in other gas-producing plants (class B), and such exposure (class C)], in internal comparisons their results were different as they had accounted for HWE. Although their paper pre-dated the paper by McMichael, they had used one of the most common ways to adjust for HWE.
Although McMichael spoke of the HWE phenomenon and the way to adjust for it by way of internal comparisons, other researchers missed this effect as they continued estimating SMRs by comparing the mortality in occupational cohorts with the general population, leading to erroneous conclusions.
| Understanding Healthy Worker Effect: an Example|| |
The classic paradigm for studies in epidemiology is by way of comparisons of apples to apples and not apples to oranges. HWE is a selection bias resulting from the latter or the improper selection of a comparison group, i.e., the general population. The “general population” is a heterogeneous mixture consisting of both “healthy” people and “unhealthy” people. Those who are not “healthy” such as children, elderly retired people, and ill people are less likely to be employed. Thus, in contrast to the general population, the employed workforce tends to have fewer sick people. In fact, if a healthy person falls ill with say heart disease or stroke or cancer, they will most probably be excluded from employment or have to go on medical leave. However, they are not excluded from the general population and the mortality statistics calculated for the general population. Many occupations, such as fire-fighters, police, and military, have to undergo strenuous physical and endurance examinations to assess their physical health before they are hired to work in such professions. Thus, comparisons of mortality rates between an employed/occupational group and the general population will be biased, as not all in the general population will be at “risk” of being employed. With this in mind, let us see an example.
Suppose exposure to a noxious occupational exposure truly increases the risk of mortality by 20% [risk ratio (RR) = 1.2]. Let us further assume that the general population has an overall risk of death that is 10% higher than the occupational group. Using the general population as the reference group in this case would lead to an underestimation of the risk ratio (RR = 1.1).
Thus, any observational study of workers or among occupational cohorts could potentially face this problem to a varying degree. Most studies find an average underestimation of about 25% of the association of a noxious stimulus on mortality as a result of HWE.,,,,,,,,,,,,,,, This underestimation of risk implies that being exposed to the noxious stimulus was protective. ,,,, HWE may not be complete and may only partially mask the excess mortality  and in rare cases may not be protective.
So far we have spoken of HWE in terms of improper selection of the reference group. In addition, epidemiologists have classified HWE as selection bias, confounding, or both [Figure 1].
Selection bias: HWE can be conceptualized as a form of selection bias because the comparison group is “systematically different in characteristics” from the exposed group of workers, who had to meet a certain criteria (their occupation) in order to be selected [Figure 1]. Most occupational studies will exclude workers who quit or leave after a short period of service, further compounding the issue. The ones who leave have systematically different characteristics than the ones who stay in the occupation. This self-selection into the occupation and the heterogeneous nature of the comparison group are reasons why HWE is considered to be a form of selection bias [Figure 2].
Confounding: The reasons for HWE being considered a confounder may have less substantial reasoning as one may view HWE due to inappropriate selection of a comparison/control group and hence selection bias as opposed to confounding [Figure 3]. However, we define a confounding factor as a predictor of both the exposure and outcome and yet is not in the causal pathway. As discussed above, workers in professions are healthier and able-bodied. If they fall ill or as they age, they leave the workforce. Thus, if we consider the health status to be a predictor of exposure (or being able to work in an industry) and we know that the health status is a predictor of mortality, we can now consider HWE to be a confounder. Later we will develop this concept more and show how HWE can be considered to be a time-varying confounder as the health status declines over the years and more specialized methods are needed to take this into account.
Both: HWE can also be thought of as a combination of a selection bias and confounding, as it satisfies the criteria for both. There is self-selection based on health status leading to a confounded association between the exposed group and mortality. This may underestimate the effect of exposure to noxious occupational stimuli on subsequent mortality when compared to the general population.
In any case, HWE needs to be factored in any observational study of occupational cohorts. It comprised several processes, namely: healthy worker survivor effect, healthy hire effect, and decline in health in time since hire.,
| Components of Healthy Worker Effect|| |
This discrepancy in the health status can be a result of multiple factors [Figure 4].
Healthy worker survivor effect
- Healthy hire: Employees in certain professions, such as military, fire-fighters, and shipyard workers to name a few, need to be able-bodied. Hence, people with poor health, disabilities, and chronic conditions such as heart disease and asthma  will be excluded at the time of hire due to failure to complete physical endurance tests. Further, hiring of workers may be affected by personal habits and physical conditioning such as weight, alcohol intake, smoking, or gender depending on the situation ,,,
- Time since hire or survival effect: Self-selection is a major contributor to HWE.,, Healthy workers tend to remain in the workforce. However, over time, the health status of workers drops and they leave the workforce. Those afflicted with chronic diseases, alcoholism, or other health issues may either change jobs frequently or may be forced to retire early. The ones who remain over time are the healthiest of those who started. Thus, dropouts and incomplete follow-up of certain workers have no association with the exposure of interest but rather a function of their baseline health status
- Advantages of working: Workers tend to remain employed not only based on their health status but also as a result of their improved access to healthcare, regular screening for disease, early treatment of conditions, and physical exercise. Each of these in turn maintains their health status.
Wearing off of healthy worker effect
Based on the above, one may assume that if one works in an industry with no noxious exposure then their overall mortality experience would be better. However, HWE is not constant and depends on several factors  [Figure 2] and [Figure 5].
- Time since hire/Duration of employment: HWE wears off over time., Studies have tried to estimate when the effect actually wears off without any degree of success. The effect does wear off with age. For example, a 20-year-old fire-fighter is much healthier than a 35-year-old fire-fighter. So after about 15–20 years of being in an occupation, the HWE gradually decreases and the health status of the worker approaches closer to that of the general population from which he/she was drawn. One could consider comparing a 25-year-old fire-fighter with a 25-year-old office worker, as both are in professional fields and apparently healthy. However, the health status of the fire-fighter would be much higher as he would need to clear physical endurance tests that the office worker would never have to consider. Thus, one can think of the health status declining over time and the consequent risk of mortality increasing until it reaches that of the general population and plateaus., Further as duration of employment increases, unhealthier people will quit or move to different work and so the HWE will intensify 
- Age at hire: Younger people tend to be healthier than older people. Older workers trying to enter a field with rigorous physical requirements will have a very low probability of being selected
- Choice of comparison group: One can imagine that professions such as law enforcement officers may have to satisfy physical requirements initially at the time of hire, while others such as fire-fighters have to satisfy these requirements on a more regular basis. Hence comparing fire-fighters to law enforcement personnel might also lead to biased results ,
- Gender: A stronger HWE is seen in women as compared to men ,,
- Race: Different effects are seen in different races. Black men had increased the risk of ischemic heart disease compared to white men, with increased exposure to at least one type of metalworking fluid in a prospective study by Costello et al
- Socioeconomic status: Blue-collar workers demonstrate a stronger overall HWE as their work tends to be more physically demanding than white-collar professionals.,
Thus, the HWE is not constant across age groups, gender, race, and workers nor is it constant over time. Hence, assessing HWE and accounting for it in statistical analyses is complicated and requires sophisticated methods.
| Statistical Methods for Dealing With Healthy Worker Effect|| |
As per the counterfactual definition of cause, the best comparison group for a specific population at a time point is the same population at the same time point without the exposure of interest  – quite impossible as that would involve time travel. Many approaches have been developed to minimize HWE as we can never truly neutralize this effect. Here we explore some of these methods and their potential drawbacks.
- Broadening exposure group/Latency: Many occupational cohort studies simply use the general population as the reference population.,,,,,, They mention the HWE but make different assumptions about the effect or extent of effect this has on their analysis results. For example, one can include the experience of every individual who worked in that industry or factory. Thus, the exposed group would include the unhealthy people and their experiences would be closer to that of the general population. Another approach when using the general population as the reference group is to allow for the health status of the workers to deteriorate by allowing a period of latency of time, e.g., 1 year, 5 years, and thus reduce the disparity in the health status of the two groups.,,,, Restricting to survivors, however, reduces the generalizability of the results. No conclusions could be drawn to those working for shorter periods
- Internal comparisons: The most common method is the use of internal comparisons or comparisons among the occupational cohort with differing levels of exposure (high versus low exposure), which has been used by multiple studies.,,,,,,,,, A drawback of this approach is that workers with lower levels of exposure may be at a lower health status than those exposed to high levels. For example, office workers in a factory versus factory floor workers. Both groups may need to achieve a certain level of health status to work in the occupation; however, the health status of factory floor worker may be higher than that of the office worker (and so is the exposure status). Another drawback of this approach is that it may not be likely that an occupational exposure may pose a gradient effect on health, i.e., low exposure and high exposure may have comparable effects
- External worker group comparisons: Another approach is to use a comparable group in another industry with similar health status but without the exposure of interest as the reference population.,, This approach can result in a comparable reference group. However, this new reference group may have exposure to other noxious stimuli and hence this approach is not as popular
Now if we consider HWE to be a form of confounding we can adjust for the confounder and get a more accurate estimate of exposure effect on outcome
- Adjusting: As with other confounders, one could also adjust for employment status by stratifying. This approach makes the assumption that HWE is independent of the exposure disease pathway. However, if this assumption is violated and HWE is also an intermediate in addition to being a confounder, then the results are biased., The primary reason for this is that the reasons for people leaving a job/occupation are not random with regard to health status and so would be related to their mortality. In addition, this determines future health status (i.e., health status varies with time) and depends on past exposure and occupational history. These issues led to the development of methods for time-dependent covariates
Earlier we had mentioned that HWE can be considered to be a special form of confounding. Also HWE is not constant over time and may wear off. Thus, one could think of HWE as a form of time-varying confounding by health status of individuals. This line of thinking has led to the development of specialized techniques discussed in brief below
- Marginal structural models (MSMs): MSMs using inverse probability weighting can also be used to estimate risk accurately in occupational cohorts to deal with time-dependent confounding. This method assigns weights to each subject based on their probability of being exposed or their exposure history. This creates a sort of pseudo-population – the whole unexposed and exposed population (which is, of course, twice as large as the original population). However, these methods are easier than G-estimation methods as they are more intuitive to understand and easier to perform compared to G-estimation. These models do assume that exposure is a possibility for all combinations of predictors that are included in a model
- Other methods for time-dependent covariates: One of epidemiology's greatest researchers, Robins, has developed multiple techniques for this issue, including the G-null test, the Monte-Carlo G-computation algorithm, and the G-estimation methods ,,,
- G-null test: Here we match cases to noncases based on past history of exposure and employment status. Thus, we compare those who develop the disease to those who do not, but share the same employment history and exposures till a particular point in time. Then we can examine the exposures at that time point for systematic differences. No comparison is made between those who are currently employed and those who are not employed. Since this technique employs matching, we only use the discordant pair to make inferences. However, this does not yield an estimate of the magnitude of the causal association between exposure and death
- Monte-Carlo G-computation Algorithm: This technique was developed to overcome the drawback of the G-null test.,,,, However, this technique is difficult to implement , and has been largely replaced by newer methods developed by Robins 
- G-estimation methods: These methods, using stricter matching on past exposure and occupational history, yield an estimate of the effect of exposure on the change in life expectancy. Robins does this by modeling log odds of exposure versus no exposure at each time point t from the time of being hired as a function of past occupational and exposure history, and all other relevant covariates and the time of death. Using a continuous time t makes it impractical, so he recommends using intervals of 6-month durations for his method. Other intervals are also possible. He further defines a counterfactual experience, wherein the person was not exposed, calculates an “expected” or counterfactual time of death, and relates this to the observed time of death. The main underlying concept is that the counterfactual time of death reflects the individual's underlying health status
These methods do assume that only those with the same prior exposure history can be compared. Further, their accuracy depends on how well past exposures are captured. Though these methods seem counterintuitive, as they do not seem to relate traditional measures of exposure with mortality, they are very effective in reducing the bias seen as a result of HWE in traditional studies.
Further models and methods are being described in the literature, including structural nested failure time models among others.,
| Conclusion|| |
HWE is a form of selection bias or confounding that results from improper choice of a reference group. Multiple methods have been put forth to overcome this issue. However, more work is needed to develop methods that can overcome the biases introduced by HWE.
Financial support and sponsorship
Conflicts of interest
There are no conflicts of interest.
| References|| |
McMichael AJ. Standardized mortality ratios and the “healthy worker effect”: Scratching beneath the surface. J Occup Med 1976;18:165-8.
Ogle W. Letter to the Registrar-General on the mortality in the registration districts of England and Wales during the ten years 1871-80; 1885.
Fox AJ, Collier PF. Low mortality rates in industrial cohort studies due to selection for work and survival in the industry. Br J Prev Soc Med 1976;30:225-30.
Doll R, Fisher RE, Gammon EJ, Gunn W, Hughes GO, Tyrer FH, et al
. Mortality of gasworkers with special reference to cancers of the lung and bladder, chronic bronchitis, and pneumoconiosis. Br J Ind Med 1965;22:1-12.
Rinsky RA, Zumwalde RD, Waxweiler RJ, Murray WE Jr, Bierbaum PJ, Landrigan PJ, et al
. Cancer mortality at a Naval Nuclear Shipyard. Lancet 1981;1:231-5.
Vinni K, Hakama M. Healthy worker effect in the total Finnish population. Br J Ind Med 1980;37:180-4.
Ames RG, Trent B. Mobility of diesel versus non-diesel coal miners: Some evidence on the healthy worker effect. Br J Ind Med 1984;41:197-202.
Dahl E. High mortality in lower salaried Norwegian men: The healthy worker effect? J Epidemiol Community Health 1993;47:192-4.
Zock JP, Heederik D, Doekes G. Evaluation of chronic respiratory effects in the potato processing industry: Indications of a healthy worker effect? Occup Environ Med 1998;55:823-7.
Iwatsubo Y, Matrat M, Brochard P, Ameille J, Choudat D, Conso F, et al
. Healthy worker effect and changes in respiratory symptoms and lung function in hairdressing apprentices. Occup Environ Med 2003;60:831-40.
Agerbo E. Effect of psychiatric illness and labour market status on suicide: A healthy worker effect? J Epidemiol Community Health 2005;59:598-602.
Fornalski KW, Dobrzyński L. The Healthy Worker Effect and Nuclear Industry Workers. Dose Response 2010;8:125-47.
Naimi AI, Cole SR, Westreich DJ, Richardson DB. A comparison of methods to estimate the hazard ratio under conditions of time-varying confounding and nonpositivity. Epidemiology 2011;22:718-23.
Thygesen LC, Hyidtfeldt UA, Mikkelsen S, Bronnum-Hansen H. Quantification of the healthy worker effect: A nationwide cohort study among electricians in Denmark. BMC Public Health 2011;11:571.
Nishikitani M, Nakao M, Tsurugano S, Yano E. The possible absence of a healthy-worker effect: A cross-sectional survey among educated Japanese women. BMJ Open 2012;2.
Naimi AI, Cole SR, Hudgens MB, Brookhart MA, Richardson DB. Assessing the component associations of the healthy worker survivor bias: Occupational asbestos exposure and lung cancer mortality. Ann Epidemiol 2013;23:334-41.
Chowdhury R. A Large Cohort Linkage Study of Lead Exposure for Mortality & End Stage Renal Disease. Emory University; 2013.
Chowdhury R, Darrow L, McClellan W, Sarnar S, Steenland K. Incident ESRD among participants in a lead surveillance program. Am J Kidney Dis 2014;64:25-31.
Chowdhury R, Sarnat SE, Darrow L, McClellan W, Steenland K. Mortality among participants in a lead surveillance program. Environ Res 2014;132:100-04.
Chowdhury R, McClellan W, Sarnat S, Mukhopadhyay A, Darrow L, Steenland K. Survival Patterns of Lead-Exposed Workers With End-Stage Renal Disease From Adult Blood Lead Epidemiology and Surveillance Program. Am J Med Sci 2015;349:222-7.
Steenland K, Chowdhury R. Renal disease incidence among 58 000 male workers with blood lead measurements. Occup Environ Med 2014;71:A45-A45.
Solan S, Wallenstein S, Shapiro M, Teitelbaum SL, Stevenson L, Kochman A, et al
. Cancer Incidence in World Trade Center Rescue and Recovery Workers, 2001-2008. Environ Health Perspect 2013;121:699-704.
McNamee R. Confounding and confounders. Occup Environ Med 2003;60:227-34.
Arrighi HM, Hertz-Picciotto I. The evolving concept of the healthy worker survivor effect. Epidemiology 1994;5:189-96.
Arrighi HM, Hertz-Picciotto I. Controlling the healthy worker survivor effect: An example of arsenic exposure and respiratory cancer. Occup Environ Med 1996;53:455-62.
Le Moual N, Kauffmann F, Eisen EA, Kennedy SM. The Healthy Worker Effect in Asthma: Work may cause asthma, but asthma may also influence work. Am J Respir Crit Care Med 2008;177:4-10.
Shah D. Healthy worker effect phenomenon. Indian J Occup Environ Med 2009;13:77-9.
] [Full text]
Monson RR. Observations on the healthy worker effect. J Occup Med 1986;28:425-33.
Naimi AI, Cole SR, Hudgens MB, Brookhart MA, Richardson DB. Assessing the component associations of the healthy worker survivor bias: Occupational asbestos exposure and lung cancer mortality. Ann Epidemiol 2013;23:334-41.
Symons JM, Kreckmann K, Le H, Starks S. Components of the healthy worker effect with quantification for different referent comparisons. Occup Environ Med 2014;71:A119.
Lea CS, Hertz-Picciotto I, Andersen A, Chang-Claude J, Olsen JH, Pesatori AC, et al
. Gender differences in the healthy worker effect among synthetic vitreous fiber workers. Am J Epidemiol 1999;150:1099-106.
Baillargeon J, Wilkinson G, Rudkin L, Baillargeon G, Ray L. Characteristics of the healthy worker effect: A comparison of male and female occupational cohorts. J Occup Environ Med 1998;40:368-73.
Costello S, Picciotto S, Rekhopf DJ, Eisen EA. Social disparities in heart disease risk and survivor bias among autoworkers: An examination based on survival models and g-estimation. Occup Environ Med 2015;72:138-44.
Parascandola M, Weed DL. Causation in epidemiology. J Epidemiol Community Health 2001;55:905-12.
Bolumar F, Porta M. Epidemiologic methods: Beyond clinical medicine, beyond epidemiology. Eur J Epidemiol 2004;19:733-5.
Roberts MH, Sim MR, Black O, Smith P. Occupational injury risk among ambulance officers and paramedics compared with other healthcare workers in Victoria, Australia: Analysis of workers' compensation claims from 2003 to 2012. Occup Environ Med 2015;72:489-95.
Robins J. A new approach to causal inference in mortality studies with a sustained exposure period-application to control of the healthy worker survivor effect. Mathematical Modelling 1986;7:1393-512.
Robins J. A graphical approach to the identification and estimation of causal parameters in mortality studies with sustained exposure periods. J Chronic Dis 1987;40:139S-61S.
Neophytou AM, Costello S, Brown DM, Picciotto S, Noth EM, Hammond SK, et al
. Marginal structural models in occupational epidemiology: Application in a study of ischemic heart disease incidence and PM2.5 in the US aluminum industry. Am J Epidemiol 2014;180:608-15.
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;3:319-36.
Robins JM. Addendum to “a new approach to causal inference in mortality studies with a sustained exposure period-application to control of the healthy worker survivor effect”. Comput Math Appl 1987;14:923-45.
Snowden JM, Rose S, Mortimer KM. Implementation of G-Computation on a Simulated Data Set: Demonstration of a Causal Inference Technique. Am J Epidemiol 2011;173:731-8.
Vansteelandt S, Keiding N. Invited commentary: G-computation--lost in translation? Am J Epidemiol 2011;173:739-42.
Rose S, Snowden JM, Mortimer KM. Mortimer, Response to invited commentary. Rose et al
. respond to “G-computation and standardization in epidemiology”. Am J Epidemiol 2011;173:743-4.
Naimi AI, Cole SR, Hudgens MB, Richardson DB. Estimating the effect of cumulative occupational asbestos exposure on time to lung cancer mortality: Using structural nested failure-time models to account for healthy-worker survivor bias. Epidemiology 2014;25:246-54.
Joffe MM. Structural Nested Models, G-Estimation, and the healthy worker effect: The promise (Mostly Unrealized) and the pitfalls. Epidemiology 2012;23:220-2.
[Figure 1], [Figure 2], [Figure 3], [Figure 4], [Figure 5]