CONTROLLING THE HEALTHY WORKER SURVIVOR EFFECT - AN EXAMPLE OF ARSENIC EXPOSURE AND RESPIRATORY CANCER

Objective-This investigation sought to examine whether methods propose d to control the healthy worker survivor effect would influence the sh ape or magnitude of the dose-response curve for respiratory cancer ind uced by arsenic. Methods-Results from an unadjusted analysis are compa red with results obtained by applying four different methods for contr ol of the healthy worker survivor effect to data on arsenic exposure a nd respiratory cancer. The four methods are: exposure lag, adjustment for work status, cohort restriction, and the G null test. Results-Coho rt restriction gave erratic results depending upon the minimum years o f follow up used. Exposure lag substantially increased the rate ratios and a non-linear shape (decreasing slope) compared with an unlagged a nalysis. Adjusting for work status (currently employed v retired or ot herwise not employed) yielded slightly higher rate ratios than an unad justed analysis, with an overall shape similar to the baseline analysi s. Results from the G null test procedure of Robins (1986), although n ot directly comparable with the baseline analysis, did show an adverse effect of exposure that seemed to reach a maximum when exposure was l agged between 10 and 20 years. Conclusions-All results confirm an adve rse effect of arsenic exposure on respiratory cancer. In these data, i t seems that the healthy worker survivor effect was not strong enough to mask the strong effect of arsenic exposure on respiratory cancer, N evertheless, several methods show a stronger association between arsen ic exposure and respiratory cancer after adjustment for the healthy wo rker survivor effect, suggesting that for weaker causal associations, studies not controlling for this source of bias will have low power to detect results. Although the G methods are theoretically the most unb iased, further work elucidating the validity of the assumptions underl ying lagging, adjustment for work status, and the G methods are needed before clear recommendations can be made.