REPORT TO THE WORKERS COMPENSATION BOARD ON THE HEALTHY WORKER EFFECT

REPORT TO THE WORKERS COMPENSATION BOARD ON THE HEALTHY WORKER EFFECT

Industrial Disease Standards Panel (ODP)
IDSP Report No. 3
Toronto, Ontario
June, 1988

The Industrial Disease Standards Panel is a Schedule 1 Agency of the Government of Ontario attached to the Ministry of Labour. The function of the Panel, as defined in Section 86p of the Workers' Compensation Act of Ontario, is as follows:

(a) to investigate possible industrial diseases;

(b) to make findings as to whether a probable connection exists between a disease and an industrial process, trade or occupation in Ontario;

to advise on eligibility rules regarding compensation for claims respecting industrial diseases.

The Panel is required by statute to report its findings to the Workers' Compensation Board of Ontario.

Additional copies of this publication are available by writing:

Industrial Disease Standards Panel
10 King Street East, 7th Floor
Toronto, Ontario M5C IC3
(416) 965-5056

ISBN 0-7729-4335-4

REPORT TO THE WORKERS' COMPENSATION BOARD
ON
THE HEALTHY WORKER EFFECT

TABLE OF CONTENTS

1. LETTER OF TRANSMITTAL

2. REPORT OF THE INDUSTRIAL DISEASE STANDARDS PANEL
ON THE HEALTHY WORKER EFFECT

APPENDIX A: CONTRIBUTED PAPERS ON THE HEALTHY WORKER EFFECT

APPENDIX B: EVIDENTIARY BASE FOR THE REPORT ON THE HEALTHY WORKER EFFECT

REPORT OF THE INDUSTRIAL DISEASE STANDARDS PANEL ON THE HEALTHY WORKER EFFECT

June 21, 1988

MEMORANDUM TO:		WORKERS' COMPENSATION BOARD
FROM:		INDUSTRIAL DISEASE STANDARDS PANEL
RE:		REPORT ON THE HEALTHY WORKER EFFECT

1.0 ISSUES:

In a letter dated June 12, 1986, the Workers Compensation Board requested that the Industrial Disease Standards Panel comprehensively review the issue of the Healthy Worker Effect (HWE) and develop appropriate recommendations for the interpretation of epidemiological studies the results for which will, in turn, be employed to develop adjudicative criteria for the compensation of industrial disease. In particular, the Board requested responses to the following questions:

1. Should the WCB take the HWE into account in evaluating the epidemiological data found in mortality and morbidity studies?

2. If the answer to question 1. is yes, then

a) What type of correction factor should typically be employed to address this potential source of bias?

b) Are there any sorts of mortality or morbidity outcomes (e.g. cancer) in respect of which this correction factor should not apply?

2.0 PANEL INVESTIGATIONS:

2.1 The significant literature on the HWE has been assembled in the Evidentiary Base (listed in Appendix B). some of which was reviewed in a staff paper (Gallina, 1986) prepared for the Panel. The Panel then decided to obtain advice from a noted epidemiologist, Dr. Geoffrey Howe, Director of the Epidemiology Unit for the National Cancer Institute of Canada at the University of Toronto, in the form of a presentation, discussion and paper (Howe, 1987). A summary of the methodological details in Howe's paper formed the basis for a request for short papers on the HWE from a number of epidemiologists with international stature. The following material was requested:

a) Examples of the presence of the HWE from the respondent's own occupational epidemiological studies (both mortality and morbidity) and the means employed to deal with it;

b) The knowledge the respondent derived about the HWE therefrom; and

c) The conclusions the respondent now believes to apply about this phenomenon, particularly in the case of cancer.

2.2 As a result, the Panel received ten papers in all from contributors in five countries (four from Canada, three from the United States, and one each from Great Britain, Sweden and Australia). Nine of these papers form Appendix A of this Report. The tenth, by Professor R. S. Roberts of McMaster University, was subsequently withdrawn by the Panel from its Evidentiary Base because the paper's release to Panel had not been fully cleared with its co-authors and sources of information by Professor Roberts. The respondents and their affiliations are listed below:

1. Axelson, O. Dept. of Occupational Medicine, University Hospital, Linkoping, Sweden.

2. Doll, Sir Richard. Imperial Cancer Research Fund Cancer Epidemiology and Clinical Trials Unit, Radcliffe Infirmary, Oxford, United Kingdom.

3. Enterline, P.E. Department of Biostatistics, School of Public Health, University of Pittsburgh.

4. Howe, G.R. Epidemiology Unit, National Cancer Institute of Canada, University of Toronto.

5. McMichael, A.J. Department of Community Medicine, University of Adelaide, Adelaide, South Australia.

6. Miettinen, O.S. Department of Epidemiology and Biostatistics, Faculty of Medicine, McGill University, Quebec.

7. Monson, R.R. Harvard School of Public Health, Boston, Mass.

8. Nicholson, W.J. Division of Environmental and Occupational Medicine, Mount Sinai School of Medicine, New York.

9. Sterling, T.D. School of Computing Science, Simon Fraser University, British Columbia.

2.3 Lastly, the Panel received a compendium of these short papers on the HWE (Heller, 1988).

3.0 PANEL FINDINGS AND RECOMMENDATIONS:

3.1 McMichael defined the HWE as "the consistent tendency for actively employed people to have a more favourable mortality (or morbidity) experience than the population at large". "The HWE is not an intentional measurement of the relative good health of a working populations; nor does it quantify the beneficial effects of the occupational environment upon those working within it. Rather, it is an unintended bias, of uncertain magnitude, in an unavoidably imperfect comparative measure of the health status of the working population." (McMichael, 1987)

3.2 The HWE is to be expected in an epidemiological study with a historical cohort design in which the health experience of an employed group is compared with that for the general population. If a comparison with another working group (with similar health status) were possible, then the HWE would not appear. All contributors agreed that the HWE is the result of:

selection bias (both by employers and by employees at the time of hire and afterward);
classification bias; and of
confounding effects (between the observed worker population and the referent population).

A few contributors noted a specific kind of selection bias in the form of a survivor effect which produces a reduced mortality after a longer period of time beyond hire among workers in a known setting of occupational risk.

3.3 Selection bias occurs for various reasons. Employment naturally screens against illnesses which occur at younger ages and which prevent or modify continuing work. However, employment for younger people does not discriminate against diseases of older ages (such as cancer). Nor does it discriminate against diseases for which no clinical manifestations have revealed themselves. Pre-employment medical screening by the employer in favour of robustness (for jobs requiring physical exertion) and against known risk factors among applicants would necessarily lead one to expect deficits in mortality (or morbidity) for some time following hire in cardiovascular or cerebrovascular disease, and for various other non-malignant diseases (e.g. nonmalignant diseases of the respiratory, digestive, endocrine and urinary systems). Unless screening against smoking and dietary fat intake and workers with previous diagnoses of cancer took place, there is no known underlying biological mechanism to explain reduced cancer risks in the years after the first period of employment.

3.4 Selection bias can continue to operate during employment. It is recognized that work confers a number of health-enhancing benefits including income, self-esteem, access to employer-sponsored quality medical care or occupationally conferred health benefits.

3.5 Less well known perhaps is the selection bias caused by the survivor effect in some industries with known occupational risks. In such industries, a lesser mortality for occupationally-related causes of death can occur with prolonged follow up. This appears to occur for two reasons. There is a decline in the number (and proportion) of workers with associated risk factors (e.g. smoking among asbestos workers) in the course of longer employment. As a result, the remaining workers show a less severe occupational risk estimate when compared to a standard population. Another reason for a continuing selection bias would be individual worker susceptibility where the departure of those susceptible would lead to a reduced worker population with a less unfavourable health experience.

3.6 Classification bias can occur when differences arise between the methods used to establish the data for both the worker and reference populations. These include differences in diagnostic criteria, in the quality of death ascertainment and in the quality of data recording. Poor industrial hygiene data can lead to the assignment of workers into incorrect exposure categories.

3.7 Confounding is the result of external factors whose effect can, perhaps, be controlled for during the data analysis. The use of inappropriate comparison rates can lead to confounding. Geographically corresponding rates are more appropriate for use in comparisons with the worker population and usually result in a lessening of the apparent HWE. However, local rates can be inadequate because the small numbers contribute an inherent lack of stability. The date of hire can be used to control for an apparent HWE. Thus, workers hired during World war II or during labour shortages tend to be less fit and less healthy then would otherwise be the case. Other potentially confounding factors include: age at hire, sex, specific cause of death. Specific occupation, socioeconomic or social class.

3.8 One respondent suggested using a single parameter to adjust for the presence of the HWE. He based his contention principally on his reanalysis of the Dorn study data (Kahn, 1966) which traced the mortality experience of a cohort of U.S. veterans (principally, World War I veterans) who held active government life insurance policies in December, 1953. It was conducted to determine the relationship between tobacco use and mortality experience. The policyholders were white males drawn from the middle and upper socioeconomic classes. Using the U.S. national population as a basis for comparison, the respondent found that the relative risk of mortality from all causes and from all cancers was reduced. His conclusions concerning the use of a single parameter to adjust for the HWE arose principally for two reasons: the Dorn study group was incorrectly regarded as a typical occupational cohort; there was no control for the confounding factors of socioeconomic class and race.

3.9 Ideally, the reference population used to generate risk estimates should be as closely matched as possible to the worker population. This can be achieved externally using another employed population, or internally using comparisons between high and low exposure groups within the same cohort. Although randomized clinical trials (RCTs) have been singled out as the paradigm for epidemiologists, when considering questions of causation, indirect means of duplicating RCT study designs must be used in occupational epidemiological studies since they are observational by definition.

3.10 Most contributors suggested incorporating a latency period (of from two or three years to twenty years) into the determination of estimates for all cancer standardized mortality ratios (SMRs) or site-specific cancer SMRs.

3.11 With these observations in mind, the Panel makes the following recommendations:

RECOMMENDATION 1: THE WORKERS' COMPENSATION BOARD SHOULD TAKE THE HEALTHY WORKER EFFECT (HWE) INTO ACCOUNT IN EVALUATING THE EPIDEMIOLOGICAL DATA FOUND IN MORTALITY AND MORBIDITY STUDIES.

RECOMMENDATION 2: EACH EPIDEMIOLOGICAL STUDY, ESPECIALLY OF THE HISTORICAL COHORT TYPE, SHOULD BE ASSESSED TO DETERMINE IF THERE IS ANY EVIDENCE OF THE HWE IN THE FORM OF A REDUCED (I.E. < 100%) STANDARDIZED MORTALITY RATIO (SMR) FOR ALL CAUSES MORTALITY OR FOR ALL CARDIOVASCULAR DISEASES MORTALITY; OR CORRESPONDINGLY REDUCED STANDARDIZED MORBIDITY RATIOS (SMbR).

RECOMMENDATION 3: A CORRECTION FACTOR SHOULD NOT BE EMPLOYED TO ADDRESS THIS POTENTIAL SOURCE OF BIAS SINCE EACH STUDY REQUIRES INDIVIDUAL INTERPRETATION CONCERNING THE EXTENT TO WHICH THE HWE MAY HAVE BIASED THE POINT ESTIMATE OF THE STANDARDIZED MORTALITY OR MORBIDITY RATIOS FOR EACH CONDITION OF INTEREST.

RECOMMENDATION 4: WHERE THERE IS EVIDENCE OF THE PRESENCE OF THE HWE, AND THERE IS THE POSSIBILITY OF EXCESS MORTALITY (OR MORBIDITY) FROM NON-CARDIOVASCULAR DISEASE CAUSES, THE EPIDEMIOLOGICAL ESTIMATES OF MORTALITY OR MORBIDITY SHOULD IN GENERAL BE DERIVED AFTER REMOVING FROM THE ANALYSIS THE INITIAL GROUP OF YEARS FROM THE TIME OF FIRST EMPLOYMENT. THE NUMBER OF YEARS OF FOLLOW UP TO BE SO REMOVED SHOULD BE APPROXIMATELY EQUAL TO THE AVERAGE ESTIMATED DURATION IN TIME FROM THE EARLIEST CLINICAL MANIFESTATIONS OF THE DISEASE TO FINAL OUTCOME (BASED, FOR EXAMPLE, ON THE USE OF SURVIVAL CURVES). FOR LUNG CANCER, FOR EXAMPLE, IT IS SUGGESTED THAT THE INITIAL 5 YEARS FOLLOWING FIRST EMPLOYMENT SHOULD BE REMOVED FROM THE ANALYSIS.

3.12 Panel considers that all of the above recommendations concerning the manner in which the Workers' Compensation Board should take the Healthy Worker Effect into account in evaluating epidemiological data should apply to Panel in its own work.

APPENDIX A

CONTRIBUTED PAPERS

ON

THE HEALTHY WORKER EFFECT

APPENDIX A CONTRIBUTED PAPERS ON THE HEALTHY WORKER EFFECT TABLE OF CONTENTS
Olav Axelson	Department of Occupational Medicine, University Hospital, Sweden
	VIEWS ON THE HEALTHY WORKER EFFECT AND RELATED PHENOMENA
Sir Richard Doll	Imperial Cancer Research Fund Cancer Epidemiology & Clinical Trials Unit, Radcliffe Infirmary, Oxford
	HEALTHY WORKER EFFECT
Philip E. Enterline	Department of Biostatistics, University of Pittsburgh
	COMMENTS ON THE "HEALTHY WORKER EFFECT" IN OCCUPATIONAL EPIDEMIOLOGY
Geoffrey R. Howe	National Cancer Institute of Canada, Epidemiology Unit, University of Toronto
	COMPONENTS AND MODIFIERS OF A HEALTHY WORKER EFFECT: EVIDENCE FROM THREE OCCUPATIONAL COHORTS, AND IMPLICATIONS FOR INDUSTRIAL COMPENSATION
A. J. McMichael	Department of Community Medicine, University of Alberta
	ASSIGNING HANDICAPS IN THE MORTALITY STAKES: AN EVALUATION OF THE "HEALTHY WORKER EFFECT"
Olli S. Miettinen	Department of Epidemiology and Biostatistics, Faculty of Medicine, McGill University
	THE HEALTHY WORKER EFFECT
Richard R. Monson	School of Public Health, Harvard University
	HEALTHY WORKER EFFECT OBSERVATION ON THE HEALTHY WORKER EFFECT
William J. Nicholson	Division of Environmental and Occupational Medicine, Mount Sinai School of Medicine, New York
	COMMENTS ON THE HEALTHY WORKER EFFECT
T. D. Sterling Weinkham, J. J.	School of Computing Science, Simon Fraser University, Burnaby, B.C.
	OBSERVATIONS ON POSSIBLE SOURCES, EXTENT, PERSISTENCE, CONSTANCY, AND CORRECTIONS FOR THE HEALTHY WORKER EFFECT

VIEWS ON THE HEALTHY WORKER EFFECT
AND RELATED PHENOMENA

Olav Axelson, Department of Occupational Medicine,
University Hospital, Sweden

The mortality or morbidity experience of a working population is often found to be less than that of a general population, a phenomenon usually referred to as the healthy worker effect. This concept was earlier discussed by McMichael (1976) and has become a popular and loosely used term in epidemiology without much distinction as to the underlying mechanisms. There are several different explanations to consider in this respect, however, some of which occur in some situations but not in other contexts.

SOME THEORETICAL REMARKS

Most important for the appearance of a healthy worker effect is probably the selection of healthy individuals for employment, either actively by the employer or through self-selection. It is likely also that the more qualified the job, the more probable is the occurrence of a strong healthy worker effect through selection. This viewpoint also implies that the healthy worker effect could be low or nonexistent in some unqualified jobs. Furthermore, good health is also promoting an employment to continue, and the complimentary aspect may be true as well.

There might also be various external causes or study-related reasons for the phenomenon to appear in a particular study. For example, regional differences in the occurrence of certain disorders may contribute in this respect, e.g. if the worker population belongs to a region with better health than the country at large. Also the reversed situation may be at hand but would not attract much attention as the healthy worker effect would then be decreased and obscured. There might be even more spurious reasons behind the appearance of a healthy worker effect, however. For example, the regional differences in the occurrence of a particular disease may be perhaps influenced by the quality of the health care as well as of local peculiarities in diagnostic criteria or terminology. This might contribute to the overall healthy worker effect seen in a study, i.e. there is rather a "pseudo-healthy" worker effect as having nothing to do with good health of workers but rather with bias.

Selectional errors may occur and distort what is known as the study base, i.e. the specific population-time segment that is involved in any particular study, and from which the information is harvested (Miettinen 1982). Such selection phenomena may take place through various mechanisms, also long before any investigation is even planned, but may also be the direct result of inadequate design or study procedures. For example, the material of subjects for a study would be inadequate if dead individuals were sorted out from company records or from trade union registers, which might be used for the study. Then the result would be a rather strong undermortality without any clear effect from the exposure even if existent. Such a selection error could be the result of a reconstruction of a company, i.e. the active work-force of a department or some other sector of a factory could have entered a new registry at a particular time, whereas deceased or sick individuals may not have been transferred to the new register in the same manner. Afterwards such circumstances may be difficult to reveal and the subsequent bias in the study might go undetected and be mistaken for a healthy worker effect.

There is also another issue of validity, which deserves attention and which has to do with the character of the reference population, i.e. when there is a specific such population for comparison (rather than numbers of expected cases as derived from national rates). Hence, also another industrial population, than that under study, i.e. a tentative reference population, could have some totally different exposure, that might cause the same disorder(s) as the exposure under study. For example, the choice of a group of copper smelter workers (even matched on smoking, etc.) as a reference population for miners would fail to reveal any actual excess risk of lung cancer, since radon daughter exposure in the mines, and arsenic exposure in the copper smelter would both cause lung cancer in excess (cf. Pershagen 1985). Again, a comparison in this respect would result in some degree of a pseudo-healthy worker effect, which may or may not be seen in the data of the study.

It is also quite obvious that the reference population should not come from an urbanized area if the index population is rural, or the reverse, especially if lung cancer or some other type of tumour or disease, as more frequently occurring in urban areas, is under study. Similar considerations are always necessary, but the circumstances may be less obvious than in these examples, especially when it comes to studies of moderate increases in disease rates or mortality with regard to widespread but rather diffuse exposures. The full effect of an agent causing cancer may therefore not be revealed in rural populations because any (broader) comparison group tend to involve sectors of an urbanized population as often having an increased risk of various cancer forms.

The principles for good comparability of populations are well appreciated for cohorts, and the healthy worker effect is usually thought of as reflecting a deficiency in this respect, but similar aspects should also apply to case-referent studies. Hence, it is questionable to just consider one particular exposure in case-referent studies, disregarding the fact that other, more or less ill-defined determinants of the disorder may operate in the unexposed sector of the study base (i.e. a situation that may also be seen as one of negative confounding). Instead it is important to try and identify an unexposed sector of the base, which is believed to be at least free from a priori known determinants of the disease and to use this subpopulation as the reference. For example, should miners and copper smelter workers live in the same area, it is necessary to identify and separate these categories and to use the remaining population sector as the reference category for estimating risk in any of these occupations (cf. Pershagen 1985 for an example).

EXAMPLES AND EXPERIENCES

By allowing for some latency time from first exposure to start of observing deaths or cancer cases, the healthy worker effect is usually decreased, since the good health at first employment gets lesser and lesser influence with time. For example, in a cohort of workers exposed to trichloroethylene, a requirement of 10 years of latency changed the observed to expected numbers from 49/62.0 to 37/39.8 for total mortality and from 11/14.5 to 9/9.5 for all tumors (Axelson et al 1976). This experience may be a simple illustration of what usually happens when allowing for latency time. This may even be done indirectly by considering the observed to expected number of deaths or cancers in various time periods after first exposure.

A further illustration of the healthy worker effect phenomenon can be obtained from a study by Ott et al (1983), who studied the workers of two factories with the same type of production but one of them not using the agent of interest, namely methylene chloride. Both populations showed nothing but strong healthy worker effects in comparisons to expected numbers based on the national cause-specific mortality, but interestingly, the workers with the exposure had a higher cardiovascular mortality compared to the others. Although the numbers were small and the two factories located in different states in the U.S., this comparison of two seemingly well comparable populations indicated the possibility that methylene chloride could be a risk factor for cardiovascular disease (say, by its metabolism through carbon monoxide).

A further illustration of the healthy worker effect might be obtained from two cohort studies on herbicide exposure (Axelson et al 1980; Riihimaki et al 1982). One of them had some excess of cancer and a slight healthy worker effect for other causes of death, the other showed a very strong healthy worker effect as even increasing for cancer with a strengthening of the latency time requirement (one observed versus 5.7 expected with 15 years of latency; P < 0.05 for prevention). The interpretation would have to be that there is either a real prevention or, perhaps more likely, a selection bias explaining the figures obtained (an overall interpretation of these and other studies on herbicides is another matter as not attempted here at all).

In case-referent studies, a healthy worker effect may also be seen, for example in the younger age strata in a study of exposure to nitroglycerine and nitroglycol exposure in the explosives industry (Hogstedt and Axelson 1977), in this case presumably an effect of rather rigorous pre-employment health check-ups. No particular efforts seemed possible to specifically deal with this phenomenon, since no other design involving a comparison with similarly selected group was possible. On the other hand, in older ages there was a clear increase in cardio-cerebrovascular deaths.

Sometimes a restriction of the study base underlying the cases of a case-control study may be helpful. A study design illustrating this point may be seen in a study of neuropsychiatric disorders and solvent exposure, where a restriction of the study base was done so as to involve only individuals with employment in construction activities, e.g. excluding white-collar workers as well as farmers, forestry workers and other groups where the frequency of especially mental disorders as a reason for pre-time pensioning was likely to be different for one reason or another (Axelson et al 1976).

SOME CONCLUDING REMARKS

With regard to positive studies, the healthy worker effect is less problematic to deal with although resulting in a lowered risk. The problem comes instead in the non-positive or negative studies, where the question arises whether there is no effect of the exposure or if a moderate effect has been obscured through the operation of a healthy worker effect or some other more or less related phenomenon. The application of some adjustment factor to the expected numbers could be thought of, but any suggested magnitude of a correction would hardly be generally accepted. The application of an adequately long latency period, or an analysis comparing observed and expected cases in various periods of time since first exposure, may be helpful in decreasing the influence of the healthy worker effect, at least in cancer studies, where the early exposure may play the greater role. Usually, however, the healthy worker effect has to be dealt with on a judgmental basis and to the extent that there is no clear excess of disease, that is strong enough to break through this effect, no positive conclusions can be drawn (but obviously no conclusions about a lack of effect either).

The radical solution of the problem of the healthy worker effect would in principle be to use a proper comparison group whenever possible. Still, it will probably be unavoidable also in the future, for economic and other reasons, to calculate expected numbers from national or regional rates for comparison with the observed number of cases, even if such a study design is known to be less adequate, not to say inadequate from a scientific standpoint (cf. Wang & Miettinen 1982).

However, even an opposite view may be taken because of the difficult problem to know what reference group that would be proper, and therefore there is perhaps some justification for using national rates in the context of cohort studies in spite of the problems caused by the healthy worker effect. This might be especially true for cancer studies since the healthy worker effect is usually moderate for cancer. For case-referent studies, the corresponding justification would be to have just all nonexposed as the reference rather than a particular sector of these. The comparison population might therefore in both instances be thought of as a reasonable average for comparison and evaluation of overmortality or overmorbidity in general, more or less leading to a sort of ethical point of view regarding the risk taken by a particular worker population. This could be relatively adequate from the point of view of the society, but not necessarily so for the exposed individual or a worker group. This view is not in agreement with the principles argued above, however, and does not lead to a clear scientific evaluation of agent-specific effects. A reasonable and practical compromise could be therefore, to apply at least some refinement as discussed, e.g. to exclude white collar professions as easily done in a case-referent design regarding some exposure in blue-collar work. This is in contrast to the possibilities in cohort studies, since national rates inevitably include all kinds of professions as well as unemployed with their high mortality, and the resulting problems of comparability. Rates for social subgroups of a population could therefore be desirable but are rarely available.

REFERENCES

Axelson O., Andersson K., Hogstedt C., Holmberg B., Molina G., de Verdier A. (1976). A cohort study on trichloroethylene exposure and cancer mortality. J. Occup. Med. 20:194.

Axelson O., Hane M., Hogstedt C. (1976). A case-referent study on neuropsychiatric disorders among workers exposed to solvents. Scand. J. Work Environ. Health 2:14.

Axelson O., Sundell L., Andersson K., Edling C., Hogstedt C., Kling H. (1980). Herbicide exposure and tumor mortality. An update epidemiologic investigation on Swedish railroad workers. Scand. J. Work Environ. Health 6:73.

Hogstedt C., Axelson O. (1977). Nitroglycerine-nitroglycol exposure and the mortality in cardio-cerebrovascular diseases among dynamite workers. J. Occup. Med. 19:675.

McMichael A. J. (1976). Standardized mortality ratios and the "healthy worker effect": Scratching beneath the surface. J. Occup. Med. 18:165.

Miettinen O. S. (1982). Design options in epidemiologic research - An update. Scand. J. Work Environ. Health 8, 1:7.

Ott M. G., Skory L. K., Holder B. B., Bronson J. M., Williams P. R. (1983). Health evaluation of employees occupationally exposed methylene chloride. Mortality. Scand. J. Work Environ. Health 9, 1:8.

Pershagen G. (1985). Lung cancer mortality among men living near an arsenic-emitting smelter. Am. J. Epidemiol. 122:684.

Riihimaki V., Asp S., Hernberg S. (1982). Mortality of 2,4-dichloro-phenoxy acetic acid and 2,4,5-trichlorophenoxy acetic acid herbicide applicators in Finland. First report of an ongoing prospective study. Scand. J. Work Environ. Health 8:37.

Wang J. D., Miettinen O. S. (1982). Occupational mortality studies. Principles of validity. Scand. J. Work Environ. Health 89:153.

HEALTHY WORKER EFFECT

Sir Richard Doll

Imperial Cancer Research Fund Cancer Epidemiology
& Clinical Trials Unit
Radcliffe Infirmary, Oxford

INTRODUCTION

I first became aware of the complication that the so-called "healthy worker effect" introduces into studies of occupational mortality in 1965, when I analyzed data that had been collected in a study of coal gasworkers employed by four of the British Area Gas Boards (Doll et al., 1965). All men were included in the study who, on 1 September, 1953 were aged between 40 and 65 years of age, had selected occupations, had been employed by the industry for more than five years, and were currently in employment or in receipt of a company pension. The last, rather unusual, criterion (being in receipt of a company pension) was included to ensure that the results were not biased by the exclusion of men who had retired early on grounds of ill-health. Altogether 11,499 men were included and all but 50 (0.4%) were followed successfully for 8 years or until death, whichever was the earlier. The 50 untraced men were assumed to have been alive so that the recorded mortality rates may have been (and in fact were) slightly underestimated. The men were divided into three broad occupational categories according to whether they had heavy, intermittent, or no exposure to the products of coal carbonization and their mortality from all causes and ten specific causes or groups of causes was compared with that expected if the men had the same mortality rates as all men in England and Wales of the same ages over the same period. The results are shown in Table 1. They showed the anticipated occupational hazards (of cancers of the lung, bladder, and scrotum for the heavily exposed workers) and suggested hazards of bronchitis for the heavily exposed men and of pneumoconiosis for the maintenance men. They also showed, however, a reduced mortality from arteriosclerotic and degenerative heart disease and from a residual group of other causes in all three occupational groups (Standardized Mortality Ratios^* of 75, 79, and 82 and of 83, 86, 73) and a substantially reduced mortality from all causes for the men with intermittent or no exposure (SMRs of 90 and 84).

The first explanation for the reduced mortality rates that we considered was that the comparison with national rates was inappropriate and we, therefore, re-examined the results, comparing the observed mortality in each Area Board with that observed in the corresponding regional conurbations, choosing the regional conurbation for comparison rather than the region as a whole on the grounds that the great majority of the gasworkers lived in the large conurbations. The results showed that the low mortality rates in the men with intermittent and no exposure could not be wholly attributed to the choice of national rather than local rates for comparison; but they did show that the low mortality from occupational causes of death was limited to two of the four geographical groups (Table 2).

Our discussion of the reasons for these low mortality rates is perhaps of some interest, as it was one of the first to have been published: that is, apart from the many previous discussions of the low mortality commonly observed in the first year or two after recruitment of an active population that was known to occur as a result of selective factors that excluded some of the seriously ill. I have, therefore, quoted it in full:

"The reason for the low mortality from all other non-occupational' diseases is also unexplained. It is not due to a beneficial effect of heavy physical work in retort houses, since it is observed among all types of employees...; nor is it an attribute of the industry as a whole, since it is observed among the employees of only two of the four Boards.... One explanation might be that the mortality in two of the Boards was underestimated as a result of recording some men as alive who were in fact dead. A check on a randomly selected 10% sample at Board III failed, however, to indicate that any substantial number of deaths could have been missed. The status of all but one of the men was confirmed; this man was said to have been still employed by the Board, whereas he had actually left though he was still alive. If the explanation is that some deaths have been missed, the mortality from occupational diseases is presumably also underestimated.

An alternative explanation is that some selective bias resulted in the inclusion of a relatively healthy group of employees in two of the Boards. Work in horizontal retort houses is heavy, and it would be understandable if chronic invalids failed to qualify for inclusion in the study by continuing to work for five years. It is, however, difficult to see why such a bias should affect men employed as meter collectors, meter readers or gas fitters (class C).^* The high rate of employment in the south east may perhaps have resulted in a more rapid labour turnover among unhealthy men, but if this were the explanation we should have anticipated that the bias would have worn off after the first few years of follow-up. In fact the mortality from non-occupational causes was practically the same in the first three and in the last five years of the study (Board III, 9. 6 and 9.4 per 1,000; Board IV, 8.2 and 9.1 per 1,000 respectively).

Whatever the explanation, it is difficult to see how the deficiency could have produced a spuriously high mortality from occupational diseases, which, it may be noted, was raised in all four Boards. It is not, we think, reasonable to suggest that the deficiency of deaths from non-occupational causes is due to bias in favour of diagnosing cancer of the lung or bronchitis. First, the deficiency is apparent in all three occupational classes, whereas the excess mortality from lung cancer and from bronchitis is present only among class A workers. Secondly, the causes of death have been classified according to the information given on death certificates; these were completed by many different doctors who were unaware of the existence of the present study and not likely to be biased by knowledge of the subjects occupation."

PERSONAL EXPERIENCE

Since reporting this work my colleagues and I have obtained mortality data for many other groups of workers, some of which are the results of continued observations on workers whose mortality we had reported previously. I have preferred, however, to cite the latest data, even if they are still unpublished, as the larger numbers reduce the element of random variation. To these I have added data on two very large groups of workers that I have analyzed recently at the request of the European Office of the World Health. Organization and of the Chemical Manufacturers Association (on, respectively, men and women manufacturing man-made mineral fibers and men exposed to vinyl chloride). These two sets of data were not collected by me or my colleagues, but many of the data are still only in press, the reviews are original, and I thought that the results would be of interest as they relate to so many workers followed for long periods.

Table 3 lists the studies and the types of disease that were found to be (or suspected of being) specifically associated with the occupations under investigation. Deaths attributed to the diseases listed are excluded from the data presented in Tables 4 and 5. A very brief account of each study is given in an Appendix.

Table 4 shows, for each study, the SMRs for all cancers combined, other than those listed in Table 3, derived from the numbers of deaths that would have been expected had the workers experienced the same mortality rates as all men and women of the same ages in the country as a whole over the corresponding periods. In some instances it has also been possible to include SMRs corrected for locality, by the use of regional or county rates or by the use of correcting factors derived from knowledge of the difference between national and local rates for the group of diseases concerned. Inspection of Table 4 shows that the SMRs for cancers not known to be related to the various specific occupations, are nearly all close to 100. Eight are greater than 100 and five are less than 100. Only one difference is statistically significant (the deficiency of cancers other than cancers of the lung and mesothelioma in male textile asbestos workers) and that is only marginally so (P=.02 when compared with national rates and .04 when compared with local rates).

Table 5 shows SMRs for all non-occupationally induced diseases other than cancer in the same 10 occupational groups plus radiologists, whose experience could not be included in Table 4 because all types of cancer (except possibly chronic lymphatic leukaemia) are believed to be induced by exposure to ionizing radiations. The data are limited to deaths from non-malignant diseases, excluding, whenever possible, deaths due to injury and poisoning, as this latter group will inevitably include an unknown proportion attributable to industrial accidents. Diseases due to non-malignant diseases certainly or possibly due to hazards associated with the specific occupations are also excluded (see Table 3). The results are substantially different from those in Table 4. Six SMRs are greater than 100, one equals 100, and 14 are less. In many instances moreover, the differences are statistically significant. Four excesses are significant (the excess of deaths due to diseases of the circulatory system in nickel salt manufacturers, of respiratory disease (other than asbestosis) and of circulatory disease in asbestos textile manufacturers, and of diseases other than circulatory disease in mustard gas workers). In three instances, however, the excess disappears or becomes statistically non-significant when a correction is made for the locality of the factory. Whether it would disappear in the fourth (nonmalignant diseases other than circulatory disease in mustard gas workers) depends on whether it is more appropriate to compare the workers with the population of Cheshire Urban Districts (in one of which the factory was situated) or with the population of neighboring Merseyside, whence many of the workers are likely to have been drawn. Of the 14 SMRs less than 100, eight are significantly low. Unfortunately, correction for locality or, in the case of radiologists for social class, could be made for only three. In one instance the deficiency ceased to be statistically significant (circulatory disease in nickel foundry workers) in one it became marginally significant (all non-malignant diseases in radiologists) and in one it remained highly significant (diseases other than circulatory disease in nickel foundry workers).

DISCUSSION

In interpreting these and other similar results, we need to take the following into account:

1. The low mortality rates reported in many occupational cohort studies are partly an artifact due to incomplete follow-up, inaccurate information, and the failure to obtain information about the cause of death for all workers known to have died. This is not a major factor, but it is certainly one of the reasons why low SMRs are often reported. In our own study of gasworkers (Doll et al., 1965) we initially failed to trace 50 out of 11,499 men (0.4%) as noted above. A continued follow-up (Doll et al., 1972) found that 5 of the untraced men were dead, 7 others who had been reported as alive and in receipt of a company pension had also actually died before the end of the stated period, and 3 others had died who had been reported as alive (although not in receipt of a pension). In total the mortality rate should have been increased by 1.2%. Failure to obtain information about the cause of death is uncommon in British studies, but it is common in US studies, which require searching for death certificates in up to 50 States. When all three types of error are allowed for (incomplete tracing, erroneous reports, and failure to trace death certificates) it would not be surprising if many studies (with follow-up rates of the order of 95% or less) underestimated SMRs by 5 or occasionally even 10%.

2. Low SMRs derived by comparing the experience of a cohort with that of the whole country may be misleading, because the national standard may have been inappropriate. When a factory is in a highly industrialized zone the SMRs may be overestimated (particularly in the case of lung cancer) but low SMRs for respiratory and circulatory disease may well be due to the location of the factory in a relatively low risk area, as was the case with our nickel foundry workers.

3. After these factors are allowed for, it still remains true that SMRs for non-malignant disease are commonly found to be well below 100 in occupational studies, a deficiency that has been attributed to the "healthy worker effect". At least four factors contribute to it:

(i) selection by the employer to exclude those obviously at high risk, e.g. individuals suffering from chronic bronchitis and emphysema, congenital heart disease, gross obesity, alcoholism, and psychosis, or convalescent from the treatment of a potentially fatal disease, etc. Some of this selection wears off within a couple of years, but much of it persists for decades.

(ii) selection to exclude those whose ill health makes the work unsatisfactory or uncongenial. In so far as this leads to people giving up work within a few months it will affect all studies which include only people who have worked for a minimum specified period. The longer the qualifying period for entry to the cohort, the more likely it is for workers with poor health to be excluded. This is particularly important if the great majority of workers included in a cohort are already in employment at the start of the study and becomes progressively less important as the proportion of the cohort recruited after the start of the study increases.

(iii) self selection by workers who, by virtue of their personal characteristics (for example, a tendency to alcoholism) change jobs frequently. Nearly all occupational studies exclude very short term workers; most exclude individuals who have worked for less than 6 months, many exclude those who have worked for less than a year, and a few concentrate on individuals who have worked for 5 years or more, with the intention of ensuring that those studied have had prolonged exposure to the suspect hazard. Evidence that short-term workers tend to have high mortality rates from many non-malignant diseases is accumulating and the exclusion of such workers inevitably results in some low SMRs.

(iv) a beneficial effect of work. This includes not only the benefit of employment as compared to unemployment (which, there is increasing evidence to suggest, is in itself harmful to health) but also such effects as still occur in some industries, like the beneficial effect of physical exertion in reducing blood pressure and the risk of myocardial infarction.

It is not, of course, to be expected that these factors will always have the same effect even after artifacts have been excluded. Their effects may vary to some extent depending on (a) the length of time individuals are followed (the longer the time, the smaller the effect); (b) the criteria for inclusion in the study (the longer the period of employment required, the greater the effect) and (c) the social conditions at the time of employment (for example, cohorts that include a substantial proportion of men recruited during periods of high employment and, particularly during the second world war, are less likely to have excluded individuals at high risk of sickness). Whether the factors referred to in paragraph 3 have much effect on the SMR for cancer is another matter. A few certainly will, such as the exclusion of alcoholics. This effect, however, is likely to be small and in general it is doubtful whether the selective factors that reduce the SMR for non-malignant diseases will have any corresponding effect on the SMR for cancer. A recent history of (say) gastric or lung cancer will certainly reduce the opportunities for starting a new appointment, but the effect of this sort of selection will mostly wear off within two years. It is extremely difficult to predict who will get cancer (apart from knowledge of the individual's smoking habits) and, unless there is selection against smokers, it is not evident that any of the factors referred to in paragraph 3 will have any material effect on the risk of cancer after (at the most) 5 years.

Experience of industrial cohorts shows that the SMR from cancers other than those known to be due to occupational hazards are most commonly close to 100 and that this also pertains when no specific occupational cancer hazard exists. In some cohorts, of course, a relatively high SMR for cancers not known to be due to occupational hazards compared to the SMRs for non-malignant diseases can be due partly to the misdiagnosis of cancers of other types that are known to be occupationally induced (as certainly used to happen with mesotheliomas in asbestos workers and probably still also occurs whenever there is a peculiarly high incidence of lung cancer) and partly to a more widespread effect of a carcinogen than has been recognized (as might be the case with vinyl chloride). In these circumstances, however, the SMR for other cancers is likely to be over 100.

CONCLUSION

I conclude that the healthy worker effect is a real phenomenon, but that it is irrelevant to the interpretation of SMRs for cancer in occupational studies, so long as the first five years' observations after recruitment to the study are excluded.

REFERENCES

Al-Dabbagh, S., Forman, D., Bryson, D., Stratton, I., Doll, R. (1986) Mortality of nitrate fertiliser workers. Brit. J. industr. Med., 43, 507-515.

Doll, R. (1987a) Symposium on MMMF, Copenhagen, October: Overview and Conclusions. Ann. occup. Hyg. (in press).

Doll, R. (1987b) Effects of exposure to vinyl chloride: an assessment of the evidence. Report to Chemical Manufacturers Association.

Doll, R., Fisher, R. E. W., Gammon, E. J., Gunn, W., Hughes, G. O., Tyrer, F. H., Wilson, W. (1965) Mortality of gasworkers with special reference to cancers of the lung and bladder, chronic bronchitis, and pneumoconiosis. Brit. J. industr. Med., 22, 1-12.

Doll, R., Mathews, J. D., Morgan, L. G., (1977) Cancers of the lung and nasal sinuses in nickel workers: a reassessment of the period at risk. Brit. J. industr. Med., 34, 102-105.

Doll, R., Vessey, M. P., Beasley, R. W. R., Buckley, A. R., Fear, E. C., Fisher, R. E. W., Gammon, E. J., Gunn, W., Hughes, G. O., Lee, K., Norman-Smith, B. (1972) The mortality of gas-workers - final report of a prospective study. Brit. J. industr. Med., 29, 394-406.

Easton, D. F., Doll, R., Morgan, L., Peto, J. (1988a) Mortality of men working with soluble nickel compounds. (to be published)

Easton, D. F., Doll, R., McKean, C. W. F., Peto, J. (1988b) Respiratory cancer due to metallic nickel exposure. (to be published)

Easton, D. F., Peto, J., Doll, R. (1988c) Cancers of the respiratory tracts in mustard gas workers. (to be published)

Peto, J., Doll, R., Hermon, C., Clayton. R., Goffe, T., Binns, W. (1985) Relationship of mortality to measures of environmental asbestos pollution in an asbestos textile factory. Ann. occup. Hyg., 29, 305-335.

Smith, P. G., and Doll, R., (1981) Mortality from cancer and all causes among British radiologists. Brit. J. Radiol., 54, 187-194.

Wald, N., Boreham, J., Doll, R., Bonsall, J. (1984) Occupational exposure to hydrazine and subsequent risk of cancer. Brit. J. industr. Med., 41, 31-34.

Table 1

MORTALITY OF GASWORKERS COMPARED WITH NATIONAL EXPERIENCE

(Doll et al, 1965)

Cause of death SMR for men with

heavy
exposure intermittent
exposure no
exposure

Occupational causes

      Cancer of lung 169 113 100

          "      "   bladder 221 143   57

          "      "   skin and scrotum 350     0     0

Possible occupational causes

      Bronchitis 213   99   94

      Pneumoconiosis   71 186     0

Other causes

      Other respiratory disease 104   77   92

      Arteriosclerotic and degenerative
          heart disease   75   79   82

      Other disease   83   86   73

      Injury and poisoning 106   88   79

All causes 105   90   84

Table 2

  MORTALITY OF GASWORKERS COMPARED WITH REGIONAL EXPERIENCE

(Doll et al., 1965)

Gas Board SMR for men with

heavy
exposure intermittent
exposure no
exposure

     1      Occupational^*
Non-occupational^** 179
97   62
  97   65
91

2 Occupational
Non-occupational 136
107 103
110   68
90

3 Occupational
Non-occupational 145
76 113
  77   77
79

4 Occupational
Non-occupational 162
58   79
  70   91
78

^* Including possible occupational causes, see Table 1.
^**Including injury and poisoning, see Table 1.

Table 3

STUDIES REVIEWED

Authors Occupational groups studied      Occupational hazards

Doll et al., 1965
     "    "     1972 Gasworkers exposed to
combustion products of coal Cancers of lung, bladder, skin
and scrotum.
Bronchitis (?)^* Pneumoconiosis

Doll et al., 1977 Nickel refiners Cancers of nasal sinuses & lung

Easton et al., 1988a      Nickel salt manufacturers Cancer of lung (?)

Easton et al., 1988b Nickel foundry workers Cancer of lung (?)

Wald et al., 1984 Hydrazine manufacturers Cancer of lung (?)

Peto & Doll, 1985 Asbestos textile workers Cancer of lung & mesothelioma.
Asbestosis

Al-Dabbagh et al.,
    1986 Nitrate fertiliser
    manufacturers Cancer of stomach (?)

Easton et al., 1988c Mustard gas workers Cancers of buccal cavity,
pharynx, nose, larynx & lung.

Smith & Doll, 1981 Radiologists Cancers of all sites

Doll, 1987a Man-made mineral fibre
    workers Cancer of lung

Doll, 1987b Vinyl chloride workers Cancer of liver (angiosarcoma)

^*Hazards marked with a query (?) were suspected but unproven.

⁺In maintenance men with intermittent exposure.

Table 4

MORTALITY FROM CANCERS OTHER THAN THOSE SUSPECTED OF ASSOCIATION WITH THE

OCCUPATION UNDER INVESTIGATION

(numbers of deaths in parentheses)

Occupational group SMR derived from comparison with:

  National
  Mortality   Regional or
  Local Mortality

(a) Gasworkers - heavy exposure
                            intermittent exposure
                            no exposure 111      (94)
101      (94)
  88    (111)

Nickel refiners 102      (75)

Nickel salt manufacturers   96      (16)

Nickel foundry workers   90      (45)          98

Hydrazine manufacturers   76        (7)

Asbestos textile workers - men
                                           women   81    (117)
105      (15)          83

Nitrate fertiliser manufacturers        106      (79)
(b)  103

Mustard gas workers 106    (284)   (b)  99,  (c)  89

Man-made mineral fibre workers 103  (1060)

Vinyl chloride workers 102    (609)

(a) Different from the results shown in Table 1 because of the inclusion of
      additional data from Doll et al., 1972.

(b) Based on local county rates.

(c) Based on neighbouring conurbation rates.

Table 5

MORTALITY FROM ALL CAUSES OTHER THAN CANCER, DISEASES DUE TO THE OCCUPATION

UNDER INVESTIGATION, AND INJURY AND POISONING

(numbers of deaths in parentheses)

Occupational group SMR derived from comparison with:

        National
        Rates     Regional or
    Local Rates

(a) Gasworkers, heavy exposure
                           intermittent exposure
                           no exposure         95    (345)
        86    (340)
        79    (414)

Nickel refiners (d)  101    (413)

Nickel salt manufacturers (e)  144      (75)
(f)    94      (26)       116

Nickel foundry workers   (e)  80    (102)
  (g)  65      (26)         98
        60

Hydrazine manufacturers   (h)  81      (37)

Asbestos textile manufacturers, men
                                                      "
                                                      "
                                                   women (j)   115    (536)
(e)  130    (169)
(g)    95      (85)
        87      (27)         97
        93

Nitrate fertiliser manufacturers   (e)  80   (145)
  (g)  67     (56)

Mustard gas workers (e)  108    (877)
(g)  132    (525) (b)  103,  (c)    91
(b)  143,  (c)  100

Radiologists registered before 1921
          "              "         after    1920         95    (257)
        76    (339) (k)  104,  (1)    97
(k)    89,  (1)    87

Man-made mineral fibres       100  (4703)

Vinyl chloride workers (d)   84  (1547)

(a), (b), (c) See corresponding footnotes to Table 4.
(d) Including injury and poisoning.     (e) Circulatory disease.
(f) Other diseases and injury and poisoning.     (g) Other diseases.
(h) All other diseases and injury and poisoning.     (j) Respiratory disease
(k) Compared with men in social class I.
(l) Compared with all medical practitioners.

APPENDIX

1. Coalgas workers (Doll et al., 1965 and 1972)

11,499 men employed by one or other of four area Gas Boards for at least 5 years were followed for 8 years from 1.9.53. All but 50 (0.4%) were traced. The men were all those employed in carbonizing plants (regarded as having heavy exposure to the products of coal combustion), all employed on maintenance work in gas-producing plants and as process men in gas-producing plants other than retort houses (regarded as having intermittent exposure) and all employed as process or maintenance workers in by-products plants, prepayment meter collectors, credit meter readers, and gas fitters (regarded as having no exposure).

A further study included the 2,444 men with heavy exposure and the 579 men who had worked in by products plants (with no exposure) in the first study and an additional 4,687 men, meeting similar criteria to those in the first study, who were employed by four other area Gas Boards. These consisted of 1,176 men with heavy exposure, 1,430 men with intermittent exposure, and 2,081 prepayment meter collectors, credit meter readers, and gas fitters with no exposure. All were followed for four years from 1.9.61. All but one of the total of 7,710 (0.01%) were successfully traced.

In both studies, the observed mortality was compared with that expected from national mortality rates for England and Wales in men of the same ages over the same period and separately, for four groups of diseases, from the mortality rates recorded in the regional constructions which corresponded most closely with the areas covered by the individual area Gas Boards.

2. Nickel refiners (Doll et al., 1977)

967 men who had been employed for at least 5 years in a nickel refinery using the nickel carbonyl process in S. Wales and whose first employment was in or before April 1944 were followed from April 1934 or such other later date as qualified them for 5 years employment to 1 January 1972. All but 37 (3.8%) were traced. The observed mortality was compared with that expected from national mortality rates for England and Wales in men of the same ages over the same period. Earlier reports on subgroups of these men had been published by Doll (1958) and Doll et al. (1970).

3. Nickel salt manufacturers (Easton et al., 1988a)

289 men employed for at least one year in either the wet treatment or chemical products plant at a nickel refinery in S. Wales who had not been employed at the refinery before 1933 (when the hazard associated with the refining process was considered to have been almost or completely eliminated) were followed from the time they qualified for inclusion (some time after 1937 when the plants first opened) to 31.12.85. The observed mortality was compared with that expected from national mortality rates for England and Wales in men of the same ages over the same period. Expected mortality rates from lung cancer and from circulatory disease were also obtained by the use of rates for rural Glamorganshire, where the refinery was located.

An earlier report on these men was published by Cuckle, Doll, and Morgan (1980).

4. Nickel foundry workers (Easton et al., 1988b)

1,907 men employed for at least 5 years in a nickel foundry in Hereford, between its opening in 1953 to April 1978 were followed to 1.4.85. All but 21 (1.1%) were traced. The men were divided into five categories according to the extent to which they were likely to have been exposed to nickel dust and their mortality compared with that expected from the national mortality rates for England and Wales in men of the same ages over the same period. Expected numbers of deaths were also obtained by the use of rates appropriate for the urban areas of Herefordshire, in one of which the foundry was located.

An earlier report on these men was published by Cox et al. (1981).

5. Hydrazine manufacturers (Wald et al., 1984)

427 men employed for at least 6 months in a factory in which hydrazine was produced between 1945 and 1971 were followed from the time they qualified for inclusion to 31.7.82; 21 (5.2%) were untraced. The men were divided into three groups according to the amount of exposure they were likely to have had and their mortality was compared with that expected from the national mortality rates for England and Wales in men of the same ages over the same period.

6. Asbestos textile workers (Peto et al., 1985)

Three groups of men and women employed at a Rochdale asbestos textile factory were followed to 30.6.83, i.e. 145 men first employed before 1933 who had served 20 years or more in scheduled areas, 238 women first employed between 1.1.33 and 31.12.62 who had served 10 years or more in scheduled areas, and 3,211 men first employed between 1933 and 1974, including all who had been employed for at least 5 years with some time in scheduled areas or on maintenance and a 10% sample of all other employees. Men with Asian surnames and men with previous occupational exposure to asbestos were excluded from this third group. Each group was followed to 30.6.83. 135 (4.2%) of the men in the third group were untraced; none was untraced in either of the other groups. Mortality was compared with that expected from the national mortality rates for England and Wales for men and women of the same ages over the same periods and corrected for the mortality from broad groups of diseases by use of the SMR for Rochdale (the town in which the factory was situated) for 1969-73.

Data for the first group of men employed before 1933 are not shown in this report because exposure to asbestos was so gross that it had a major but unassessable effect on the mortality from respiratory and circulatory diseases other than that recorded as due to asbestosis.

Previous reports on some of these workers have been published by Doll (1955), Knox et al. (1968), and Peto et al. (1977).

7. Nitrate fertiliser workers (Al-Dabbagh et al., 1986)

1,327 men employed for at least one year in the production of nitrate based fertilisers between 1946 and 1981 were categorized as having had high, intermediate, or low exposure to nitrates or nitrogen oxides. All but 2 (0.2%) were followed to 1.3.81. Their mortality was compared with that expected from Northern regional rates for men of the same ages over the same period and the expected deaths were corrected crudely for locality by multiplying by the ratio between the SMRs for the locality and the Northern region in and around 1973.

8. Mustard gas workers (Easton et al., 1988c)

2,498 men employed at a factory in Cheshire in the manufacture of mustard gas at any time between 1.7.38 and 31.12.44 and 1,052 women employed for at least one year at the same factory over the same period were followed from 1.1.45 to 31.12.84: 176 (5.0%) of the total were untraced. Their mortality was compared with that expected from national rates for men and women of the same ages over the same period. The expected deaths were corrected crudely for locality in two ways: by adjusting by the ratio between Cheshire urban and national rates at ages 15-64 years and by adjusting similarly using the neighbouring Merseyside rates instead of the Cheshire urban rates.

A preliminary report about a subcohort of these workers has been published earlier (Manning et al., 1981).

9. Radiologists (Smith and Doll, 1981)

All medically or dentally qualified men who joined a British radiological society between 1897 and 1954 have been followed up to 1.1.77, excluding only those who were abroad or in the Colonial or Armed Services at the time of registering with the society. 339 men joined before 1921, when rigorous measures for protection against ionizing radiations were introduced and 999 joined subsequently. All the former group were traced, but 5 of the latter (0.5%) were untraced. The observed deaths were compared with the numbers expected from national mortality rates for (i) all men of the same ages over the same period, (ii) all men in social class I, and (iii) all registered medical practitioners.

An earlier report of the mortality of these same men was published by Court Brown and Doll (1958).

10. Man-made Mineral Fibre workers (Doll, 1987a)

Three cohorts of men and women employed in the production of man-made mineral fibres have been studied in Canada (1 plant), Europe (13 plants) and the USA (17 plants). Altogether information has been obtained about 7,862 deaths in 41,185 workers observed for varying periods between the 1930s and the end of 1982. The results were reviewed by RD at a symposium organized by the European Office of the World Health Organization in October 1986 and the individual papers and the review are in press. Notable difference in mortality were observed between rock and slag wool workers, glass wool workers, and glass filament workers and between the results for lung cancer when mortality was compared with that expected on the basis of national rates and on the basis of local county rates in Europe and the USA. The mortality of Canadian workers was compared only with that expected on the basis of provincial rates.

The full results obtained in the three studies are to be published shortly (Enterline et al., 1987; Shannon et al., 1987; Simonato et al., 1987).

11. Vinyl-chloride workers (Doll, 1987b)

The published reports of the mortality of men occupationally exposed to vinyl chloride are reviewed. Material of substantial value has been provided only by four studies: (i) A study of 16,173 men employed in 37 plants in the USA, including all men who had been exposed to vinyl chloride for at least one year between 1942 and 1972 inclusive. Most were followed to 31.12.82, but a few were followed only to 31.12.77: 7.3% were untraced and no cause was identified for 6.3% of the 1,439 who had died. (ii) A study of 5,498 men employed for at least one year on jobs involving exposure to vinyl chloride between 1940 and 1974 in 9 plants in the UK and followed to 31.12.84: 1.1% were untraced. (iii) A study of 451 men exposed to vinyl chloride by virtue of their employment for at least 5 years in one Canadian plant since 1943 and followed to 31.12.77: all were traced (by initial definition). (iv) A study of 618 men employed for at least 6 months in one or other of two Italian plants between 1953 (when operations began) and 31.12.81 and followed to 31.12.84: 3 (4.9%) were untraced. Certified causes of death were obtained for all the 905 men in the UK, Canadian, and Italian cohorts who had died. In each study the numbers of deaths were compared with the numbers expected if the man's mortality had been the same as that recorded in men of the same ages over the same periods in the whole country or the corresponding province.

The results of the Canadian and Italian studies have been published (Theriault and Allard, 1981; Belli et al., 1986). The results of the US and UK studies are to be published (Environmental Health Associates, 1986; Jones, 1986).

REFERENCES

Al-Dabbagh, S., Forman, D., Bryson, D., Stratton, I., Doll, R. (1986) Mortality of nitrate fertiliser workers. Brit. J. industr. Med., 43, 507-515.

Court Brown, W. M. and Doll, R. (1958) Expectations of life and mortality from cancer among British radiologists. Brit. med. J., 2, 181-187.

Cox, J. E., Doll, R., Scott, W. A., Smith, S. (1981) Mortality of nickel workers: experience of men working with metallic nickel. Brit. J. Industr. Med., 38, 235-239.

Cuckle, H., Doll, R., Morgan, L. C. (1980) Mortality study of men working with soluble nickel compounds. In: Nickel Toxicology, ed. S. S. Brown and F. W. Sunderman. pp. 11-14. Academic Press, London.

Doll, R. (1955) Etiology of lung cancer. In: Advances in Cancer Research, vol. 3. Academic Press, Inc., New York.

Doll, R. (1958) Cancer of the lung and nose in nickel workers. Brit. J. industr. Med., 15, 217-223.

Doll, R. (1987a) Symposium on MMMF, Copenhagen, October: Overview and Conclusions. An. occup. Hyg. (in press)

Doll, R. (1987b) Effects of exposure to vinyl chloride: an assessment of the evidence. Report to Chemical Manufacturers Association.

Doll, R., Morgan, L., Speizer, F. (1970) Cancers of the lung and nasal sinuses in nickel workers. Br. J. Cancer, 24, 623-32.

Doll, R., Mathews, J. D., Morgan, L. G. (1977) Cancers of the lung and nasal sinuses in nickel workers: a reassessment of the period at risk. Brit. J. industr. Med., 34, 102-105.

Easton, D. F., Doll, R., Morgan, L., Peto, J. (1988a) Mortality of men working with soluble nickel compounds. (to be published)

Easton, D. F., Doll, R., McKean, C. W. F., Peto, J. (1988b) Respiratory cancer due to metallic nickel exposure. (to be published)

Easton, D. F., Peto, J., Doll, R. (1988c) Cancers of the respiratory tract in mustard gas workers. (to be published)

Enterline, P. E., Marsh, G. M., Henderson, V., Callahan, C. (1987) Mortality update of a cohort of UK man-made mineral fiber workers. Ann. occup. Hyg. (in press)

Environmental Health Associates (1986) An update of an epidemiological study of vinyl chloride workers, 1942-82. Final report to the Chemical Manufacturers Association. Environmental Health Associates, Oakland, California.

Jones, R. D. (1986) A mortality study of vinyl chloride monomer workers employed in the United Kingdom 1940-1984. (in press)

Knox, J. F., Holmes, S., Doll, R., Hill, I. D. (1968) Mortality from lung cancer and other causes among workers in an asbestos textile factory. Brit. J. industr. Med., 25, 293-303.

Manning, K. P., Skegg, D. C. G., Steet, P. M., Doll, R. (1981) Cancer of the larynx and other occupational hazards of mustard gas workers. Clin. Otolaryngol. 6, 165-170.

Peto, J., Doll, R., Howard, S. V., Kinlen, L. J., Lewinsohn, H. C. (1977) A mortality study among workers in an English asbestos factory. Brit. J. industr. Med., 34, 169-173.

Peto, J., Doll, R., Hermon, C., Clayton, R., Goffe, T., Binns, W. (1985) Relationship of mortality to measures of environmental asbestos pollution in an asbestos textile factory. Ann. occup. Hyg. 29, 305-335.

Shannon, H. S., Jamieson, E., Julian, J. A., Muir, D. C. F., Walsh, C. (1987) Mortality experience of glass fibre workers: extended follow up. Ann. Occup. Hyg. (in press)

Simonato, L., Fletcher, A. C., Cherrie, J., Andersen, A., Bertazzi, P. A., Charnay, N., Claude. J., Dodgson, J., Esteve, J., Frentzel-Beyme, R., Gardner, M. J., Jensen, O. M., Olsen, J. H., Teppo, L., Winkelmann, R., Westerholm, P., Winter, P. D., Zochetti, C., Saracci, R. (1987) The International Agency for Research on Cancer historical cohort study of MMMF production workers in seven European countries. Ann. Occup. Hyg. (in press)

Smith, P. G. and Doll, R. (1981) Mortality from cancer and all causes among British radiologists. Brit. J. Radiol. 54, 187-194.

Theriault, G. and Allard, P. (1981) Cancer mortality of a group of Canadian workers exposed to vinylchloride monomer. J. occup. Med., 23, 671-676.

Wald, N., Boreham, J., Doll, R., Bonsall, J. (1984) Occupational exposure to hydrazine and subsequent risk of cancer. Brit. J. industr. Med., 41, 31-34.

COMMENTS ON THE "HEALTHY WORKER EFFECT" IN
OCCUPATIONAL EPIDEMIOLOGY

Philip E. Enterline
Department of Biostatistics
University of Pittsburgh
October 12, 1987

In every worker study I have conducted where there has not been a probable occupational hazard all cause death rates have been less than rates expected based on the mortality of populations of the entire United States, of the state in which the worker lived or of the local area where the worker's place of employment was located. Some small part of this deficit could be the failure to track every individual to determine his vital status or due to inaccurate death ascertainment. In general, however, I believe that what is reflected here is simply a selection against ill health for participation in the work force. The extent of selection varies by disease classification with the greatest selection against those diseases that appear early in life, and with little selection against diseases unlikely to be manifest at time of employment. Thus, there is little selection against cancer since for the most part symptoms of this disease appear only a few years before death occurs and deaths from this condition tend to occur late in life. On the other hand there is selection against nonmalignant diseases, particularly cardiovascular disease. This is partly due to the fact that some of these diseases are manifest fairly early in life and prevent labor force participation and partly because these diseases are usually of long duration and interfere with employment during periods of life when individuals are likely to be members of the workforce.

Both the employee and the employer seem to participate in the selection process. Workers who do not have strong motivation to work because of health problems do not present themselves for employment while, historically at least, employers have reserved the right to reject certain persons because of physical disabilities. The extent to which these processes occur varies with time and place. During periods of labor shortages the less fit workers are more likely to be taken into the labor force whereas during periods of labor surpluses employers can be much more selective in deciding who will be employed.

My own studies have been largely confined to manufacturing in what might be considered to be heavy industries. In general I find that workplaces that appear (to me) to be fairly clean and desirable tend to produce workers where the healthy worker effect is greater than workplaces that appear to be undesirable. In a study we recently completed on copper smelter workers we found the standardized mortality ratio for all causes of death to be 86.4 based on 1491 deaths while in a study of workers at a large chemical complex the SMR for all causes of death was 74.7 based on 1180 deaths.

One problem is trying to separate any effects that might have resulted from the employment itself from overall mortality. These effects can be both negative and positive. While health hazards clearly exist in industry it seems reasonable to suppose work itself brings some satisfaction if only in the form of monetary rewards and the ability to purchase goods, as for example medical care. The latter has never been very well documented but it seems intuitively true whereas health hazards are easily documented. In a study of retired asbestos workers, for example, I found an all cause SMR of 123.8 while in a very large study we recently completed on workers producing man made mineral fibers we found the all cause SMR to be 102.0 based on 4986 deaths.

While historically employment does not directly select for or against cancer, in recent years I have noticed a tendency for employers to select against cigarette smoking, either directly or indirectly. Some time ago I visited a large refinery chemical complex at a time when they had been making a recruitment effort to fill 30 vacancies. The day I arrived they had over 300 applicants. It interested me as to how the personnel manager would select 30 from over 300. My discussions with him led me to believe that his final choice probably represented a selection of population likely to live a very long time. Some of his criteria included no history of drinking or legal problems, no history of divorces or family troubles, a history of stable employment prior to applying for the present position, and perhaps no history of cigarette smoking. My impression was that the workers chosen were nonsmokers, nondrinkers with very stable family lives.

Clearly the question of a healthy worker effect is not a simple one. In general, however, I feel where there is no occupational hazard I would expect the workforce to have an overall death rate about 80% of the death rate in an appropriate reference general population and I would expect the cancer SMR to be higher than the all cause SMR. If the overall SMR is close to 100 I suspect a problem. An analysis by date of hire sometimes helps. If I find WWII hires or hires during periods of labor shortages are contributing heavily to the overall SMR I may modify my feelings about the SMR.

I do not attempt any adjustment for the "healthy worker effect" for specific disease categories. This is probably because I have mainly been concerned with cancer excesses where the healthy worker effect is probably minimal. If the cancer SMR is very low I suspect I've missed some deaths, have encountered a population containing many nonsmokers, or have the wrong reference population. I've also considered beneficial effects of employment as for cotton textile workers exposed to endotoxins.

COMPONENTS AND MODIFIERS OF A HEALTHY WORKER EFFECT:
EVIDENCE FROM THREE OCCUPATIONAL COHORTS, AND
IMPLICATIONS FOR INDUSTRIAL COMPENSATION

Geoffrey R. Howe
National Cancer Institute of Canada
Epidemiology Unit
University of Toronto

SUMMARY

The components and modifiers of the healthy worker effect have been examined using mortality data for three occupational cohorts, the employees of Atomic Energy of Canada Limited followed between 1950 and 1981, a ten-percent sample of the Canadian labour force followed between 1965 and 1979, and workers at the Eldorado Resources Limited Beaverlodge uranium mine followed between 1950 and 1980. Two important components have been identified in these cohorts, namely initial selection of healthy individuals, and continuing employment of healthy individuals. There is less evidence for a contribution from the existence of differential risk factors amongst employed individuals compared with the general population. The healthy worker effect is however substantially modified by time since employment, sex, age, specific cause of death, and specific occupation. It is concluded that because of this variation it is inappropriate to account for the healthy worker effect by a single parameter, and all of the above factors have to be taken into account in any appropriate analysis. The effect of the healthy worker bias on assessing the causality of any observed association, and attributing cause in an individual case is also discussed: when the only available comparison group for an occupational cohort is the general population, the healthy worker effect is unlikely to have any substantial influence upon the interpretation of either of these two components of the compensation decision process. This would be particularly true for cancer, and even more so for lung cancer, often a disease associated with industrial compensation cases.

INTRODUCTION

The healthy worker effect is the name given to the observation that cohort studies of individuals employed in some occupation or industry usually show that such individuals have a lower mortality than the general population, which includes individuals who are not currently employed for various reasons.¹ The phenomenon seems first to have been described by Ogle,² who suggested that initial selection for employment of healthy individuals, and continuing employment of those who remained healthy was responsible for the healthy worker effect. Reports of the results of occupational studies in which comparisons are made with the general population usually take cognizance of the healthy worker effect, but few attempts have been made to identify and quantify its components and modifiers. An exception to this is the report by Fox and Collier³ of the healthy worker effect as seen in a cohort study of those employed in the U.K. vinyl chloride industry, but their results are based on a relatively small total number of deaths, which limits the stability of their estimates. The healthy worker effect is important in attempting to assess the causality of any observed association for an occupational cohort based on comparison with the general population. Although various methods have been proposed for taking it into account, for example by adjusting standardized mortality ratios (SMRs) by a constant factor, 1 or by using proportionate mortality,⁴ these solutions have considerable limitations.

This paper describes the healthy worker effect as seen in three occupational cohorts whose mortality has been ascertained using computerized record linkage to the Canadian National Mortality Database.⁵ The cohorts consist of: 1) employees of Atomic Energy of Canada Limited, a Crown Corporation responsible for the research and development of the Canadian nuclear power programme, and related technologies, whose members in general show no marked excess of deaths from any particular cause;⁶ 2) a 10 percent sample of the Canadian labour force consisting of 700,335 individuals whose occupations between 1965 and 1969 have been recorded;⁷ and 3) males employed by Eldorado Resources Limited at their Beaverlodge Uranium Mine in northern Saskatchewan between 1948 and 1980, where exposure to radon decay products has led to a large excess of lung cancer.⁸ These data have been used to examine and quantify the components and modifiers of the healthy worker effect, and the distinctive nature of each of the cohorts makes comparisons amongst them particularly informative. The findings are discussed with respect to implications in assessing the causality of any observed association in an occupational cohort, and in particular to the implications for attributing causation for an individual who develops a disease, and consequently is under consideration for compensation.

THE HEALTHY WORKER EFFECT: COMPONENTS AND MODIFIERS

One may postulate at least four potentially important components of the healthy worker effect. The first arises at the time of initial employment. Individuals who have a diagnosed disease such as chronic bronchitis may either be less willing themselves to seek employment, or alternatively may seek employment but are more likely to be rejected by a potential employer on the basis of their disease. The effect could also operate for diseases not yet diagnosed, but which could still make an individual feel less able to seek employment, though presumably the effect would be less than for specifically diagnosed conditions. The second possible component again would occur primarily at the time of initial employment and relates to risk factors for disease such as smoking. If the characteristics with respect to such risk factors of those who seek employment are different from the general population, subsequent mortality from those diseases associated with the risk factor would differ in the employed cohort compared to the general population. Both the above factors are of course confounding variables in the classic sense, the first relating to confounding by current disease status, the other to confounding by current risk factor status. The third potential component of the healthy worker effect arises from preferential continued employment of those who remain healthy, and the tendency of those who develop disease to leave employment. Thus, if comparisons are made between those who remain in employment during the time period of observation, with the general population, their mortality would be lower by reason of those who have already left employment due to disease. This factor should not in general apply to studies in which mortality is determined for an entire occupational cohort, irrespective of whether or not they continue in employment. Finally, a fourth component could arise from differential diagnosis of the cause of death for an occupational cohort as compared with the general population. An intensive review of death certification, and/or clinical records for the cohort could serve to identify cases of disease which would not be identified from routine death certification on which population statistics are based. Again, this fourth component would only operate where such special surveys have been carried out and are probably the exception rather than the rule.

The effect of all of the above four components might be expected to be modified by a number of factors. If the initial selection component is important, one would expect that the effect would diminish considerably with time since initial employment, particularly if the mortality experience of those who subsequently leave employment is included. Since there will obviously be a strong correlation between time since employment and increasing age, it is also necessary to examine the effect of age at observation to determine whether any time since employment effect is independent of an age effect. In addition to time since employment and age, other obvious important modifiers of the effect could be sex, the specific cause of death, and the type of occupation. These components and modifiers are examined below with respect to the three occupational cohorts described.

THE OCCUPATIONAL COHORTS

Full details of the three occupational cohorts have been described previously.⁶,⁷,⁸ Mortality results with respect to lung cancer for the uranium miners study,⁸ and cancer mortality amongst males for the ten percent labour force study⁷ have been published. All three studies use the same method of follow-up, namely by computerized record linkage to the Canadian National Mortality Database maintained by Statistics Canada. The latter contains in machine readable form all deaths occurring in Canada since 1950, together with those of Canadian residents occurring in the U.S. Date of death and underlying cause of death coded to the appropriate revision of the international classification of diseases is available, together with personal identifying information. Since no unique identification number is available on the database, it is searched using the corresponding identifying information from the cohort records. The searching process makes use of a probablistic weighting system for comparing the individual items of identifying information, and the statistical theory and computer system used have been described elsewhere.⁹ Brief details of each of the cohorts are provided.

Atomic Energy of Canada Limited (AECL) Study: This cohort consists of 13,570 current and ex-employees of AECL known to be alive as of January 1, 1950 and constitutes 92.8% of those originally defined as eligible for the study. The majority of those not included were current employees who did not wish to participate in the study. Identifying information for the cohort was assembled from company records and by means of questionnaires completed by current employees. Mortality follow-up has been carried out between 1950 and 1981, resulting in 882 deaths amongst males and 66 deaths amongst females. A total of 159,845 person-years of observation have been accumulated by the 10,034 males in the cohort, and 54,807 person-years by the 3,536 females. The interest in this cohort arises since a number of the subjects are classified as radiation workers with exposure to well monitored doses of low-level low-linear energy transfer radiation. Based on current radiation risk-estimates,¹⁰ it is unlikely that any significant excess of radiation related deaths will be observed in this cohort, but it is being monitored to confirm the adequacy of those estimates.

Labour Force Survey (LFS) Study: This cohort consists of 700,335 members of the Canadian labour force, constituting approximately a ten-percent sample. The survey was carried out by Statistics Canada to compile employment statistics, and records are available for individuals with recorded occupation and industry between 1965 and 1971. Records for 1970 have been lost, and a different coding scheme was employed for the 1971 records which was not compatible with that used for 1965-69. Therefore, the present analysis is restricted to occupations as recorded between 1965 and 1969. The personal identifiers for this cohort were obtained from the master index file of social insurance numbers. The mortality of the cohort between 1965 and 1979 has been monitored as described above, and a total of 41,194 deaths amongst males and 7,365 deaths among females has been observed. Person-years of observation total 5,467,282 for the 415,201 males, and 2,805,141 for the 285,134 females in the cohort. The study is intended as a routine monitoring in order to detect associations between occupation and risk of death not previously reported, and to confirm associations reported from other studies.

Eldorado Resources Limited (ERL) Study: The cohort consists of 8,487 males employed at some time in the Beaverlodge uranium mine operated by Eldorado Resources Limited in northern Saskatchewan between 1948 and 1980, and known to be alive as of January 1, 1950. This constitutes 77.5% of those originally deemed eligible, the majority of those excluded being due to missing birth years. Identifying information was assembled from company records, and mortality has been determined between 1950 and 1980, resulting in 603 deaths. Too few females were employed for any meaningful analysis, and the person-years at risk for the males are 118,337. This cohort represents one of the largest series of uranium miners studied to date, and has provided valuable data on the relationship between exposure to radon decay products and risk of lung cancer, of which there is a substantial excess amongst the cohort.⁸

METHOD OF ANALYSIS

The observed number of deaths and the person-years at risk were calculated for each study categorized by age group (15-19, 20-24...80-84, 85+), sex and calendar-year at risk (1950-54, 55-59...75-81). Entry to the study was defined as occurring during first year of employment for the AECL and ERL studies, and during first year for which an occupation was recorded for the LFS study. Year of exit was defined as year of death for those who died, or else the end of the last calendar year in which mortality ascertainment had been carried out (1979 for the LFS study, 1980 for the ERL study, and 1981 for AECL study). Six causes of death were analyzed: lung cancer (a disease frequently the subject of compensation awards), other cancers, circulatory diseases, chronic respiratory diseases (bronchitis and emphysema), accidents and all causes of death combined. The expected number of deaths for any particular analysis were computed from the age, sex and calendar-year specific death rates for the disease under consideration for the Canadian population, applied to the appropriate person-years at risk in the cohort. Standardized mortality ratios (SMR) were then computed as: 100-percent multiplied by the ratio of the observed number of deaths to the expected number of deaths. The individuals enrolled in the LFS study were classified by social class I-V on the basis of their recorded occupations. The occupations were grouped into social classes using a system very similar to that used in the decennial supplement to the U.K. Registrar General's reports on occupation,¹¹ slightly modified for Canadian occupation.¹² Individuals were placed in the social class corresponding to the highest value for any of their recorded occupations. Data are presented in three groups corresponding to social classes I and II (professional and managerial), social class III (other white collar) and social classes IV and V (blue collar). A small number of occupations could not satisfactorily be classified, so the total presented (Table 6) are slightly less than those for the other analyses.

RESULTS

Table 1 shows SMRs and the number of observed deaths for all causes of deaths combined by period of follow-up for the various cohorts for both males and females. Both the AECL study and LFS study show substantial healthy worker effects, which decrease with increasing time of follow-up. The effect is stronger for the AECL study than for the LFS study, presumably reflecting the fact that follow-up in the AECL study is from start of employment, whereas in the LFS study it is by time since observation started, which acts as a surrogate for time since first employment. In contrast, the ERL data do not show the healthy worker effect for all causes of death. Another noticeable feature from Table 1 is that the effect is stronger for females than for males in both the AECL and LFS studies.

The contribution of the several specific causes of death to the overall effect are examined in Table 2 for the AECL and LFS studies. In general, the specific causes show the same increasing effect with time under observation as do all causes of death (details not shown). The patterns of the effect on the various diseases are generally consistent, though the data for the female AECL cohort are sparse and consequently are difficult to draw inferences from. The healthy worker effect is less for lung cancer than for other cancers, and in turn this is less than for all causes of death. The effect for circulatory diseases is less than for all causes of death in the LFS cohort, but the opposite is true for the AECL male cohort. The effect for deaths from respiratory diseases is less than that for all causes for both male cohorts, but is greater than for all causes for females, although the AECL female data are based on a single death.

In order to see whether the healthy worker effect for the ERL data are masked by the presence of occupationally related excesses for certain causes of death, the data for ERL are examined by cause and time under observation in Table 3. There is evidence for a healthy worker effect for cancers other than lung, and for respiratory diseases, the latter however is based on small numbers. The excess of lung cancer due to radon decay products is well established,⁸ and the excess of deaths from accidents is again well recognized in mining cohorts. The pattern for circulatory diseases appears to have no obvious explanation.

Table 4 presents data for all causes of death by period of follow-up and age at risk. For all studies, and for all age groups the general pattern of a decreasing healthy worker effect with increasing period of follow-up is seen. (The AECL female data are too few to contribute to this detailed analysis). However, although the strongest healthy worker effect is seen for the youngest age group for the AECL male data, the reverse is true for both males and females in the LFS study. Some of these observations, particularly for young ages and greatest length of follow-up, and old ages and shortest length of follow-up are inevitably based on small numbers.

Table 5 examines the evidence for a component of the healthy worker effect due to the preferential continued employment of those who are healthy, by examining all causes of mortality for the male AECL cohort by a period of follow-up, classified as to whether or not the individual continued to be employed by AECL. For the first ten years of follow-up both groups show a substantial and similar decease in mortality relative to the population, but after that time period those who leave employment have a mortality approximately 50 percent greater than those who remain in employment.

Finally, Table 6 examines the effect of social class by cause of death for the LFS study cohort. For males, the greatest reduction in mortality compared to the population is for social classes I and II, compared to social classes III, IV and V. Although the overall mortality for class IV and V is similar to that for class III, there are differences for the individual causes of death. In particular, lung cancer deaths are higher for classes IV and V, but their rate for other cancers is lower. Classes IV and V also have lower rates from circulatory diseases, but substantially increased rates from accidents. For females, the differences between the social classes in general appear somewhat less than for males as judged by overall mortality. However, again lung cancer is increased in social classes IV and V, as are accidents compared to the other social classes.

DISCUSSION

Empirical Evidence for Components and Modifiers of the Healthy Worker Effect: The results reported in the previous section come from three very different cohorts. The LFS study represents a 10-percent sample of the entire Canadian labour force, and thus represents a diverse collection of occupations and industries. The AECL cohort represents an occupational group in which a priori one would expect no large excesses from any particular cause of death, and probably represents a higher than average social class group, containing many skilled white and blue collar workers. In contrast, the ERL study consists primarily of uranium miners, many of whom are non-skilled. They have expected high rates of mortality from lung cancer due to exposure to radon decay products, and from industrial accidents. The data confirm the importance of at least two components of the healthy worker effect, namely the initial selection process, and the continuing employment component. (Tables 1 and 5 respectively). However, the healthy worker effect is subject to substantial modification by time since employment, sex, specific causes of death, age, and type of job. In general, the maximal effect is observed during the first five years of employment, and has almost disappeared 20 years after first employment for the AECL males, or in the case of the LFS study, 10 years after first observation. The only exception are female AECL employees, which although based on small numbers still shows a substantial reduction in overall mortality compared to the population 20 years after first employment. It is clearly demonstrated in both AECL and LFS studies that the effect is stronger for females than males, and this appears consistent for the individual causes of death considered. The general pattern of a decreased effect for lung cancer and a smaller decreased effect for other cancers is not unexpected: cancer has a high fatality rate, particularly lung cancer, and consequently the initial selection bias disappears more rapidly for cancer than for other long term chronic diseases such as cardiovascular disease. The effect of age as seen in the LFS data is consistent with that observed by Fox and Collier for the U.K. vinyl chloride workers,³ but seems to operate in the opposite direction for the AECL cohort. There is no obvious reason for this difference, but it is marked and presumably represents different selection criteria for different industries which operate differentially at different ages.

The evidence supporting a role for confounding by risk factors as a component of the healthy worker effect is much less striking. As stated, the cohorts in general achieve the mortality of the population after a number of years under observation. The social class data shown in Table 6 do indicate that occupations in the higher social classes (I and II) do have a persistent lower mortality than the general population, as compared to other social classes both white collar and blue collar. However, the data in Table 6 again illustrate the necessity of considering factors such as specific cause of death, since this clearly indicates that there is a differential between class III and classes IV and V with respect to diseases such as lung cancer, presumably reflecting differences in smoking habits.

In summary the study has demonstrated strong evidence for two components of a healthy worker effect, namely initial selection and continuing employment, but suggests that a component resulting from confounding by risk factors makes much less contribution to overall mortality, though it may play a role in certain specific diseases. A number of factors modify the healthy worker effect, and consequently it is not possible to make generalizations about a single "healthy worker effect", and in particular the suggestion of correcting for such an effect using a single figure as proposed for example by Goldsmith¹ would not be valid, and could indeed be very misleading.

Implications for Industrial Compensation: In assessing the relevance of the healthy worker effect to the issue of industrial compensation it is necessary to consider the effect on two parts of the compensation decision process. These are 1) the decision that an observed association between occupational exposure and risk of disease is causal and 2) the model used for attributing causation in a particular individual case. The third part of the process, namely the model used for compensation is often a subjective process, and subject to wide variation in various jurisdictions and cannot realistically be considered here.

In assessing the causal nature of any observed association it is traditional to consider a number of criteria.¹³ The two criteria primarily affected by the healthy worker effect are the observed strength of the association as measured by the relative risk and the statistical significance of the association. These two factors relate to possible contribution by systematic error and random error respectively to the observed association. If no appropriate internal control group is available (see below) and the relative risk is estimated by the SMR comparing the occupational cohort to a general population, the healthy worker effect will generally tend to bias the estimated relative risk towards unity, although the magnitude of this bias will vary as discussed above with time since employment, age, sex, cause of death and type of occupation. This therefore will weaken any conclusion as to causality based on the strength of association. However, the reduction in the strength of the evidence will be most marked when the true value of the relative risk is small, and when the bias is large. For example, if the bias (B) is defined as:

B = R(u, z)/R(p,z)
    R(e,z)/R(p,z)
  = -------------                       (1)
    R(e,z)/R(u,z)
    Obs. RR
  = -------
    True RR

Where R(e,z) is the risk of disease of an individual with a vector of exposure variables (e), and a vector of covariates (z), R(u,z) is the risk for an individual without exposure and with a set of covariates (z), and R(p,z) is the risk for an individual in the comparison population with the same set of covariates. If the value of B is 0.8, which represents a fairly typical healthy worker effect, and the true relative risk is 2.0, the observed relative risk on average will be 1.6, whereas if the true relative risk is 5, on average the observed relative risk would be 4. In the former case it would appear more likely that the observed association could be due to confounding, than in the latter case. However, when considering the significance of any observed association, it is conceivable that using comparisons with the general population can increase the power in relation to a comparison with an internal control group, despite the healthy worker effect. For a given number of observed deaths in the exposed group, the increase in power will be greater as the true relative risk increases, and as the bias from the healthy worker effect decreases.

In order to assess the effect of the healthy worker effect on the attribution of causality in any individual case it is necessary to develop the appropriate model for attribution. The probability that an observed death arises from exposure will be given by:

     R(e,Z)-R(u,z)
PC = -------------                      (2)
        R(e,z)

assuming that the difference in risks arises solely because of a causal relationship between exposure and disease. Obviously, this probability of causation as expressed in equation 2 will be a function of the underlying risk for the individual i.e. will be specific for a particular set of covariates (z) and any interaction with the exposure factor (e). It is therefore necessary as usual to assume some form of probabilistic model in order to come up with stable estimates of the probability of causation. The usual model chosen is the multiplicative one i.e.:

     RR(e)-1
PC = -------                            (3)
      RR(e)

derived from equation 2 above by assuming that the relative risk is a function only of the exposure vector and is independent of the covariates. The empirical evidence for the multiplicative model has been discussed extensively¹⁴ and although it is generally difficult to definitively determine the nature of the relationship between risk from exposure to occupational factors, and risk from other factors, the multiplicative model is frequently chosen. If a bias is introduced into the relative risk estimation by the healthy worker effect, the bias (B') introduced in the probability of causation under the multiplicative model is given by:

     RR(e)-1/B
B' = ---------                          (4)
      RR(e)-1

if B is less than unity as is generally the case, the bias in the probability of causation will also be less than unity i.e. the observed probability of causation will be less than the true probability of causation. Table 7 shows some examples of the bias in the probability of causation as a function of the true relative risk and the bias in the relative risk due to the healthy worker effect. The bias in the PC decreases as the relative risk increases, and as the bias from the healthy worker effect decreases, but is often substantially less than the bias in the relative risk itself. In fact, the bias in the PC will be less than the bias in the relative risk provided that:

       1/B-B
RR = ---------                          (5)
        1-B

thus if the bias in the relative risk is 0.5, any true relative risk greater than 3.0 will show a reduced bias in the PC, and a bias of 0.8 in the relative risk estimate will result in a smaller bias in the PC for any true relative risk greater than 2.25.

CONCLUSION

In analyzing the results of any occupational cohort study estimates of risk, both for assessing causality, and for use in an attribution of causation model, are best based on internal comparisons, in order to minimize bias due to the healthy worker effect. However, this will only be true when an appropriate internal comparison group exists i.e. one of sufficient size to yield adequate power, and one which is either comparable to the exposed group in terms of potential confounding variables, or has measures of those variables which can be controlled in analysis. It will also generally be sensible to present results compared to general population rates, even when these are not used for definitive risk estimation. It is also apparent from the results presented in this paper, that for an internal analysis, it is necessary to take account of factors such as time since first employment, and continued employment, if these factors differ among the exposed and non-exposed members of the cohort.

When an appropriate internal control group is not available. the use of general population rates (specific for age, sex, race, location and calendar year) provides a sensible alternative for risk assessment, despite the healthy worker effect. The use of such a comparison will generally increase the power of any study despite the downward bias generally introduced by the healthy worker effect, and this is particularly true as the relative risk increases and as the bias decreases. In addition, for reasonably large relative risks, and the order of the bias generally seen, the healthy worker effect is unlikely to make a substantial difference to the assessment of causality as based on the strength of the association. If a model for the probability of causation is used to attribute cause in an individual case, the bias in the probability of causation is likely to be less than the bias in the relative risk estimate itself, and this reduction in the bias will increase as the relative risk increases, and as the healthy worker effect bias decreases. Many of the associations recognized for industrial compensation involve cancer as an outcome. The healthy worker effect in general appears smaller for cancer than other causes of death. In addition, the attribution model often specifies a minimum latent period typically of ten years. The latent period generally is highly correlated with time since first employment, again this tends to minimize the bias due to the healthy worker effect. Thus, for these particular situations it is likely that any bias introduced by the healthy worker effect when using population rates for comparison will be minimal, and the bias in the probability of causation will be even smaller and thus negligible.

REFERENCES

1. Goldsmith, J.R. What do we expect from an occupational cohort? Occup. Med. 1975; 17: 126-131.

2. Ogle, W. Letter to the Registrar-General on the mortality in the registration districts of England and Wales during the ten years 1871-80. Supplement to the 45th Annual Report of the Registrar General of Births, Deaths and Marriages, in England. (1885).

3. Fox, A. J., Collier, P. F. Low mortality rates in industrial cohort studies due to selection for work and survival in the industry. Br. J. Prev. Soc. Med. 1976; 30: 225-230.

4. McMichael, A. J. Standardized mortality ratios and the "healthy worker effect": scratching beneath the surface. J. Occup. Med 1976; 18: 165-168.

5. Howe, G. R. Use of computerized record linkage in follow-up studies of cancer epidemiology in Canada. Nat'l. Cancer Inst. Monogr. 1985; 67: 117-121.

6. Abbatt, J. D., Hamilton, T. R., Weeks, J. L. Epidemiological studies in three corporations covering the Canadian nuclear fuel cycle. IAEA-SM-266/9, in Proceedings International Symposium on Biological Effects of Low Level Radiation. Vienna: International Atomic Energy Authority 1983; 351-361.

7. Howe, G. R., Lindsay, J. A follow-up study of a ten-percent sample of the Canadian labour force. 1. Cancer mortality in males, 1965-73. JNCI 1983; 70: 37-44.

8. Howe, G. R., Nair, R. C., Newcombe, H. B., Miller, A. B., Abbatt, J. D. Lung cancer mortality (1950-1980) in relation to radon daughter exposure in a cohort of workers at the Beaverlodge uranium mine. JNCI 1986; 77: 357-362.

9. Howe, G. R., Lindsay, J. A generalized iterative record linkage computer system for use in medical follow-up studies. Computers and Biomedical Research 1981; 14: 327-40.

10. Committee on the Biological Effects of Ionizing Radiation. The effects on populations of exposures to low levels of ionizing radiation. Washington D.C.: National Academy Press 1980.

11. Office of Population Census and Surveys. Classification of Occupations. London 1970. HMSO 1970.

12. Lindsay, J., Howe, G. R. A system for monitoring cancer mortality in a ten percent sample of the Canadian labour force. Proceedings of the Workshop on Computerized Record Linkage in Health Research, ed. R. A. Spasoff and G. R. Howe. Ottawa 1986.

13. Bradford-Hill, A. The environment and disease: association or causation. Proceedings of the Royal Society of Medicine 1965; 58: 295-300.

14. Breslow, N. E., Day, N. E. Statistical methods in cancer research: Volume 1: The analysis of case-control studies. IARC Scientific Publication, Chap. 2. Lyon 1980.

TABLE 1
SMRs AND OBSERVED NUMBER OF DEATHS FOR ALL CAUSES OF DEATH
EFFECT OF PERIOD OF FOLLOW-UP

Period of follow-up
(years)* Study

AECL
males females males LFS
females ERL
males

    0-4 54.1
(61) 19.6
  (3) 74.4
(77.65) 59.3
(1149) 130
(106)

    5-9 58.7
(84) 36.0
(6) 85.4
(14573) 71.4
(2436) 103
(96)

  10-14 82.9
(140) 37.7
(7) 108
(18856) 100
(3780) 102
(116)

  15-19 85.3
(157) 58.7
(12) - - 108
(132)

  20-24 95.4
(187) 70.3
(15) - - 97.3
(114)

  25+ 96.6
(253) 72.3
(23) - - 73.3
(39)

TOTAL 82.7
(882) 53.2
(66) 90.7
(41194) 78.6
(7365) 104
(603)

* Time since first employment for AECL and ERL studies, since 1965 for
LFS study.

TABLE 1 SMRs AND OBSERVED NUMBER OF DEATHS FOR ALL CAUSES OF DEATH EFFECT OF PERIOD OF FOLLOW-UP
Period of follow-up (years)*	Study
AECL males	females	males	LFS females	ERL males
0-4	54.1 (61)	19.6 (3)	74.4 (77.65)	59.3 (1149)	130 (106)
5-9	58.7 (84)	36.0 (6)	85.4 (14573)	71.4 (2436)	103 (96)
10-14	82.9 (140)	37.7 (7)	108 (18856)	100 (3780)	102 (116)
15-19	85.3 (157)	58.7 (12)	-	-	108 (132)
20-24	95.4 (187)	70.3 (15)	-	-	97.3 (114)
25+	96.6 (253)	72.3 (23)	-	-	73.3 (39)
TOTAL	82.7 (882)	53.2 (66)	90.7 (41194)	78.6 (7365)	104 (603)
* Time since first employment for AECL and ERL studies, since 1965 for LFS study.

TABLE 2

SMRs AND OBSERVED NUMBER OF DEATHS

EFFECT OF CAUSE OF DEATH

Cause of death Study

AECL LFS

males      females       males   females

Lung Cancer 102
(68)    0*
  (0)       107
  (3254)    112
  (283)

Other Cancers    83.4
(131) 82.0
(31)        94.3
  (6485)      89.1
(2482)

Circulatory diseases    94.3
(436) 31.4
(11)        87.4
(17826)      69.1
(2261)

Respiratory diseases           73.0
  (14) 108.5
  (1)        81.9
    (872)      80.9
    (69)

Accidents    56.7
  (72)   49.8
  (6)        103
   (4686)      94.0
  (669)

All Causes    82.7
(882)   53.2
(66)

Table 1 MORTALITY OF GASWORKERS COMPARED WITH NATIONAL EXPERIENCE (Doll et al, 1965)
Cause of death	SMR for men with
Cause of death	heavy exposure	intermittent exposure	no exposure
Occupational causes
Cancer of lung	169	113	100
" " bladder	221	143	57
" " skin and scrotum	350	0	0
Possible occupational causes
Bronchitis	213	99	94
Pneumoconiosis	71	186	0
Other causes
Other respiratory disease	104	77	92
Arteriosclerotic and degenerative heart disease	75	79	82
Other disease	83	86	73
Injury and poisoning	106	88	79
All causes	105	90	84

Table 2 MORTALITY OF GASWORKERS COMPARED WITH REGIONAL EXPERIENCE (Doll et al., 1965)
Gas Board		SMR for men with
Gas Board		heavy exposure	intermittent exposure	no exposure
1	Occupational^* Non-occupational^**	179 97	62 97	65 91
2	Occupational Non-occupational	136 107	103 110	68 90
3	Occupational Non-occupational	145 76	113 77	77 79
4	Occupational Non-occupational	162 58	79 70	91 78
^* Including possible occupational causes, see Table 1. ^**Including injury and poisoning, see Table 1.

Table 3 STUDIES REVIEWED
Authors	Occupational groups studied	Occupational hazards
Doll et al., 1965 " " 1972	Gasworkers exposed to combustion products of coal	Cancers of lung, bladder, skin and scrotum. Bronchitis (?)^* Pneumoconiosis
Doll et al., 1977	Nickel refiners	Cancers of nasal sinuses & lung
Easton et al., 1988a	Nickel salt manufacturers	Cancer of lung (?)
Easton et al., 1988b	Nickel foundry workers	Cancer of lung (?)
Wald et al., 1984	Hydrazine manufacturers	Cancer of lung (?)
Peto & Doll, 1985	Asbestos textile workers	Cancer of lung & mesothelioma. Asbestosis
Al-Dabbagh et al., 1986	Nitrate fertiliser manufacturers	Cancer of stomach (?)
Easton et al., 1988c	Mustard gas workers	Cancers of buccal cavity, pharynx, nose, larynx & lung.
Smith & Doll, 1981	Radiologists	Cancers of all sites
Doll, 1987a	Man-made mineral fibre workers	Cancer of lung
Doll, 1987b	Vinyl chloride workers	Cancer of liver (angiosarcoma)
^*Hazards marked with a query (?) were suspected but unproven. ⁺In maintenance men with intermittent exposure.

Table 4 MORTALITY FROM CANCERS OTHER THAN THOSE SUSPECTED OF ASSOCIATION WITH THE OCCUPATION UNDER INVESTIGATION (numbers of deaths in parentheses)
Occupational group	SMR derived from comparison with:
Occupational group	National Mortality	Regional or Local Mortality
(a) Gasworkers - heavy exposure intermittent exposure no exposure	111 (94) 101 (94) 88 (111)
Nickel refiners	102 (75)
Nickel salt manufacturers	96 (16)
Nickel foundry workers	90 (45)	98
Hydrazine manufacturers	76 (7)
Asbestos textile workers - men women	81 (117) 105 (15)	83
Nitrate fertiliser manufacturers		106 (79) (b) 103
Mustard gas workers	106 (284)	(b) 99, (c) 89
Man-made mineral fibre workers	103 (1060)
Vinyl chloride workers	102 (609)
(a) Different from the results shown in Table 1 because of the inclusion of additional data from Doll et al., 1972. (b) Based on local county rates. (c) Based on neighbouring conurbation rates.

TABLE 2 SMRs AND OBSERVED NUMBER OF DEATHS EFFECT OF CAUSE OF DEATH
Cause of death	Study
	AECL		LFS
	males	females	males	females
Lung Cancer	102 (68)	0* (0)	107 (3254)	112 (283)
Other Cancers	83.4 (131)	82.0 (31)	94.3 (6485)	89.1 (2482)
Circulatory diseases	94.3 (436)	31.4 (11)	87.4 (17826)	69.1 (2261)
Respiratory diseases	73.0 (14)	108.5 (1)	81.9 (872)	80.9 (69)
Accidents	56.7 (72)	49.8 (6)	103 (4686)	94.0 (669)
All Causes	82.7 (882)	53.2 (66)

REPORT OF THE INDUSTRIAL DISEASE STANDARDS PANEL ON THE HEALTHY WORKER EFFECT

APPENDIX A CONTRIBUTED PAPERS ON THE HEALTHY WORKER EFFECT

APPENDIX A

CONTRIBUTED PAPERS

ON

THE HEALTHY WORKER EFFECT