Overall participation of private pesticide applicators in the telephone interview five years after enrollment was 66%. Patterns of response associated with age, education, and marital status were consistent with what has frequently been observed in other studies.[Benfante, et al. 1989
, Shahar, et al. 1996
, Osler, et al. 2008
, Russell, et al. 2001
] Participants in the follow-up interview tended to have healthier behaviors with regard to smoking, alcohol intake, vegetable consumption, and vitamin and mineral supplementation than non-participants. In apparent contrast to these healthy behaviors, we found that body mass index was higher among participants and that individuals who reported a health condition at enrollment were more likely to participate in the follow-up.
These findings together support the “worried ill” hypothesis proposed by Veenstra.[Veenstra, et al. 2006
] Having been diagnosed with a health condition, these participants were likely instructed to improve their health habits and would therefore be more likely to report having healthier habits than participants without a diagnosis. Furthermore, these participants may have a greater vested interest in the completion of the study than those without any diagnosis. We also know the health conditions reported at enrollment were not severe enough to interfere with participation, and provided that these conditions did not seriously progress in the intervening time, we would expect that participants who reported a condition at enrollment should be capable of participating in the follow-up interview.
One limitation of this analysis is that for health conditions other than overall cancer we were unable to measure the probability of participation among incident cases. However, it was reassuring to find that incident cancer cases were not significantly different from non-cancer cases in their probability of follow-up at interview. Furthermore, we did not observe a significant effect of selection bias when we examined the association of incident cancer with smoking status in the subset of applicators who completed the follow-up interview. It should be noted that while incident cancer was not significantly associated with follow-up, it is still possible for selection bias to occur. [Greenland 1977
In addition to having limited information on incident disease status, we also could not adequately determine whether the impact of losses due to fatal incident conditions was similar to that from non-fatal health conditions. The loss of deceased participants would be expected to result in an underestimation of disease incidence, and if these participants were more (or less) likely to report pesticide exposure than those with non-fatal health conditions, this could lead to selection bias. Although we were unable to consider specific causes of death because of the small number of deaths between enrollment and follow-up, a comparison of deceased participants with participants who completed the follow-up did not indicate any remarkable differences. Since the percent of participants who died before follow-up was only 3%, we would not expect this exclusion to result in substantial selection bias.
Applicators were more likely not to participate if they had never mixed or applied pesticides or if they personally applied pesticides less than one-half of the time, consistent with the idea that those with more of a connection to the subject of the study would be more likely to participate. Conversely, larger farm size was associated with increased probability of non-participation, suggesting that how busy a participant was had an effect on participation. The relationship of other measures of pesticide use to likelihood of participation differed between the two states. Overall, participation in the telephone interview was greater in Iowa than in North Carolina. Farming activities differ between the two states. North Carolina has a longer growing season and increased crop variety compared to Iowa. Frequency of pesticide use was higher in North Carolina and was associated with increased probability of non-participation, whereas in Iowa there was no significant association. In Iowa applicators reported a significantly greater number of years of pesticide use at enrollment despite being younger on average, and this measure was proportional to the probability of participation. Overall, there did not appear to be a general trend with respect to the level of pesticide exposure and probability of follow-up. Furthermore, the potential for differential non-response or selection bias is likely to vary for specific pesticides.
Farming status at time of interview might also affect participation rates. We did not have information on pesticide license status at time of follow-up and do not know if non-respondents had disproportionately left farming. However, since nearly 20% of those who completed a follow-up interview were no longer farming at follow-up (11.5% in Iowa and 33.5% in North Carolina), leaving farming cannot entirely explain non-response at follow-up.
We investigated the potential for selection bias to affect estimates of exposure-disease associations in the subset of applicators who completed the follow-up by comparing odds ratio estimates between the original cohort and the CATI interview subset. We observed no significant changes in the estimates for any of the three associations. However, the differential loss of participants with no report of depression who were exposed to chlorpyrifos in North Carolina did increase the association from 1.07 to 1.20.
With the exception of cancer, our exploration of selection bias was based on prevalent conditions reported at enrollment. Incident conditions might have a greater impact. The severity of such conditions and the timing with regard to follow-up interview would no doubt influence participation and in turn the possibility of selection bias. For large cohort studies collecting prospective information on self-reported diseases and exposures, it is important to consider the potential for selection bias to occur and to estimate the extent to which it may bias associations.