Is Pesticide Use Related to Parkinson Disease? Some Clues to Heterogeneity in Study Results

1Parkinson disease (PD) is an idiopathic degenerative disorder of the central nervous system that impairs motor skills, cognitive processes, and other functions. The etiology of PD is largely unknown, although some genetic factors have been identified. Based on published epidemiological and toxicological studies, pesticides may be involved in the etiology of PD. However, epidemiological evidence is far from conclusive, as considerable heterogeneity has been observed in study results. Possible methodological causes of heterogeneity in study results have been suggested and include differences in study design, control selection, diagnosis of patients, and statistical analysis. Differences in exposure assessment methods could contribute to heterogeneity as well. Most previous studies relied almost exclusively on self-reported exposures, a process that is prone to recall bias, especially in case–control studies, and could lead to false-positive associations. Alternatively, one could speculate that PD patients might underreport pesticide exposure because of cognitive deficits, leading to false-negative associations. Furthermore, differences in the definition of exposure to pesticides (occupational vs. nonoccupational use, ever/never vs. regular use) could also result in heterogeneous study results. Lastly, the regions where the studies have been conducted could be of importance as regulation, types, and use of pesticides may differ from region to region.

2Several recent studies have been published on pesticide exposure and PD risk, including some prospective (cohort) studies. In the present analysis, we aimed at providing an update of the literature published since the last systematic review on PD and exposure to pesticides and pesticide subcategories by performing a systematic review and meta-analysis. We specifically set out to address the question of whether the previously described heterogeneity in study findings could be explained by differences in study design and exposure assessment methods.

3Data source. We searched the databases Embase (http://www.Embase.com/), starting with 1974, and Medline (http://www.ncbi.nlm.nih.gov/pubmed/), starting with 1950, through November 2010 using the search term “Parkinson” in combination with “pesticide*,” “insecticide*,” “fungicide*,” “herbicide*,” “rodenticide*,” “organochlorine*,” “organophosphate*,” “carbamate*,” “glyphosate*,” “paraquat,” “maneb,” “lindane,” “dieldrin,” “rotenone,” “DDT,” or “environmental factors.” The search was limited to publications in English, French, German, or Dutch; to human studies; and to original publications. We also searched the reference lists of the retrieved publications.

4Study selection. We included studies that specifically investigated PD or parkinsonism. We included cohort studies, case–control studies, and cross-sectional studies. No reviews, case reports, or conference abstracts were included. We excluded studies that summarized results of pesticide exposure only within a broad category of “chemical exposure.” Exposure to pesticides was defined as use of pesticides by the subject, thus excluding environmental studies.

5Data extraction. Two reviewers (M.M., M.B.) independently extracted reported risk estimates [i.e., odds ratios (ORs), risk ratios (RRs), or prevalence ratios], study designs, exposure assessment methods, and types of source population for the controls. We also evaluated subcategories of pesticides and extracted data about exposure–response relations and individual pesticides. Two other researchers (A.H., R.V.) acted as referees in cases of any differences. If authors reported adjustment for potential confounders, we preferred adjusted risk estimates over crude risk estimates. In cases where no risk estimate or 95% confidence interval (CI) was reported, we calculated crude risk estimates and 95% CIs with the reported numbers. Where risk estimates were reported separately for men and women, we pooled the risk estimates with a within-study meta-analysis. Of studies with more than one control group, the results of population controls were preferred above the results of hospital controls because population controls are generally considered to be a more representative comparison group than hospital controls.

6Statistical analysis. Because of the observed heterogeneity in study results, we conducted a random effects meta-analysis to pool the results of the separate studies for risk for pesticide exposure and the subgroups of herbicides, insecticides, and fungicides. We also stratified by whether or not nonoccupational exposure (e.g., gardening) was included in the exposed group. This was because of potential differences between occupational and nonoccupational exposures in intensity and frequency of exposures. In one publication, results both for occupational and for occupational and/or nonoccupational exposure were reported. We chose to include risk estimates of the more inclusive exposure definition, although final results did not differ when we included the risk estimates based on only occupational exposure (data not shown).

7Subsequently, we explored whether heterogeneity in observed risk estimates could be explained by study and exposure assessment characteristics. We did so by stratification and used meta-regression to explore statistical significance of these characteristics. Given the limited number of studies, we only explored one characteristic at a time. Characteristics explored were the type of exposure assessment (self-reported ever/never pesticide exposure, self-reported regular pesticide exposure, or exposure assessment based on reported job titles by expert judgment and/or applying a job-exposure matrix), source of control population [hospital, general population, or other (studies using family members or case acquaintances as controls, or studies that used a combination of different sources)], geographical area (North America, Europe, or other), and whether adjustments were made for potential confounders. The I2 measure was used to quantify the heterogeneity between studies; I2 can be interpreted as a measure of the percentage of the total variation that cannot be explained by chance. p-Values for heterogeneity are based on the Q-statistic. Small study effects were tested with funnel plots and Egger’s test. All analyses were performed with Stata (version 10; StataCorp, College Station, TX, USA) with the metan, metareg, metafunnel, and metabias commands. All statistical tests were two sided, and a p-value of < 0.05 was considered statistically significant.

8The search in Embase and Medline yielded 883 publications, of which 52 publications met the inclusion criteria. We excluded 3 publications where the study population had been included in subsequent publications. Lastly, one study was excluded because the reported data showed risk per year of pesticide exposure, which was not comparable with reported risk ratios of other studies. Among the remaining 48 publications, there were two studies for which the relevant results were reported in two separate publications each. Thus, results of a total of 46 studies were used in the meta-analysis.

9An overview of the study characteristics of the included studies can be found in Table 1. There were 39 case–control studies, 4 cohort studies, and 3 cross-sectional studies; 40 publications reported on pesticides, 15 on herbicides, 15 on insecticides, and 9 on fungicides. Three studies included all parkinsonism; the rest studied idiopathic PD. Four studies showed only results in men. One study included only cases with a disease diagnosis before 51 years of age, which is much lower than the average age of disease onset in all other studies (generally ~ 60 years of age). Information about participation rates was provided for only 13 of the 39 case–control studies. Studies that reported participation rates had rates between 69% and 100% for cases and between 41% and 100% for controls.

10Figure 1 shows PD relative risk estimates for any pesticide exposure based on studies of occupational and/or nonoccupational exposures, and studies of occupational exposures only. The summary risk ratios (sRRs) between these two groups were very similar, with sRRs of 1.69 (95% CI: 1.38, 2.06) and 1.52 (95% CI: 1.23, 1.89), respectively, and an overall sRR for all studies combined of 1.62 (95% CI: 1.40, 1.88). The I2 for all studies combined was 63.7%. Only three studies estimated effects of nonoccupational exposure only, with an sRR of 1.18 (95% CI: 0.86, 1.63).

11Meta-analyses by herbicide, insecticide, and fungicide exposure are shown in Figure 2. In line with the results for any pesticide exposure, we did not observe noticeable differences between studies of occupational exposures only and studies of nonocccupational and occupational exposures combined. The sRR for exposure to fungicides did not indicate an association with PD (overall sRR = 0.99; 95% CI: 0.71, 1.40; Figure 2C), in contrast with positive sRRs for exposure to herbicides (overall sRR = 1.40; 95% CI: 1.08, 1.81; Figure 2A) and insecticides (overall sRR = 1.50; 95% CI: 1.07, 2.11; Figure 2B).

12Funnel plots of effect estimates for exposure to pesticides and pesticide subcategories were suggestive of small study effects, with a tendency for smaller studies to report higher relative risks compared with larger studies (Figure 3), with Egger’s test p-values of 0.057, 0.338, 0.208, and 0.680 for pesticide, herbicide, insecticide, and fungicide effect estimates, respectively.

13Figure 4 presents subgroup sRR estimates for those factors a priori hypothesized to be related to the observed heterogeneity in study results. The only study characteristic that was suggestive of contributing to heterogeneity was the exposure assessment method, with the lowest summary estimates observed for self-reported exposures (n = 36) and highest sRR for studies with exposures estimated based on reported job titles (n = 3). However, these differences were not statistically significant (p = 0.30). There was no evidence for a difference in summary estimates by adjustment of results for potential confounders, type of control population source, geographical area, or by study design. We also investigated whether adjustment for smoking had an effect on the summary risk estimate. Almost identical results were found for studies that did or did not correct for smoking (data not shown). Similar analyses for the subcategories herbicides and insecticides rendered similar results as for all pesticides (data not shown).

14Our systematic review indicated that PD is related to pesticide exposure with an sRR of 1.62 (95% CI: 1.40, 1.88). However, there was substantial heterogeneity among individual study estimates (I2 = 63.7%). Summary estimates also indicated positive associations of PD with herbicides and insecticides, but not with fungicides. We systematically investigated several factors that could explain heterogeneity in study results, but none appeared to be related to the observed heterogeneity, with the possible exception of the method of exposure assessment. Studies that based their exposure assessment on job titles reported somewhat higher risk estimates than studies that used self-reported exposures, but the difference did not reach statistical significance, in part because of low numbers of studies relying on job title and expert judgment.

15Including persons who were nonoccupationally exposed to pesticides together with those occupationally exposed resulted in a very similar sRR. Given that occupational pesticide applications are in general more frequent and on larger areas than are nonoccupational exposures, one would have anticipated higher RRs for studies focusing only on occupational exposures. On the other hand, use of protective equipment during nonoccupational applications may be less. The fact that summary results were similar for both types of studies could indicate that nonoccupational and occupational pesticide exposures carry similar risks, or that most of the exposures in the combined studies were occupational. In the three studies that exclusively reported on nonoccupational pesticide exposures, only a small increase in relative risk was observed (sRR = 1.18; 95% CI: 0.86, 1.63), suggesting that risks associated with nonoccupational pesticide exposures are lower than from occupational exposures. Nevertheless, nonoccupational pesticide exposure cannot be ruled out as a risk factor for PD based on these analyses.

16Studies used different methods for exposure assessment and assignment. Most studies (36 of 39) were based on self-reported exposure to pesticides, defined as ever versus never use or as regular versus nonregular use. No difference in sRR was seen between these two definitions of self-reported exposure, although it could have been expected that using a more stringent definition of exposure would have resulted in stronger associations. Studies that used reported job titles and expert judgment, and/or that used a job-exposure matrix to estimate exposures, resulted in a higher sRR compared with studies using self-reported pesticide exposures. This difference cannot be explained by recall bias, because in that case, higher risk ratios would have been expected for studies relying on self-reported exposures. A more likely explanation is that subjects are not able to reliably report exposures to pesticides, resulting in nondifferential exposure misclassification and bias toward the null. The fact that some heterogeneity is observed in study results by exposure assessment method indicates that this may be an important factor that should be taken into consideration when designing or interpreting studies.

17A broad range of different pesticides exist with different chemical compositions and working mechanisms. In line with the conclusions of, we found that both herbicides and insecticides, but not fungicides, were associated with PD. However, it is difficult to disentangle the effect of herbicides and insecticides given that the use of these two pesticide groups is often highly correlated. This is illustrated by the fact that we observed a correlation coefficient of 0.79 between the study-specific RRs of herbicides and insecticides. Few studies have focused on specific pesticides precluding any meaningful meta-analyses. However, it is interesting to note that the subgroup of organochlorines was significantly associated with PD in three studies. This is also in line with studies on biomarkers in serum and in the brains of deceased patients. Organochlorines are mainly insecticides, including DDT (dichlorodiphenyltrichloroethane), dieldrin, and heptachlor.

18Funnel plots gave some indication for a small-study effect, such that larger effect estimates appeared to be associated with smaller studies, which suggests that the sRR might be slightly overestimated. In addition, the studies included were generally small, resulting in imprecise effect estimates that could have contributed to the substantial heterogeneity in study results. Meta-regression analyses provided no evidence for a difference in sRRs based on study design, geographical area, adjustment for potential confounders, or type of control population. As such, factors explaining the heterogeneity observed remain largely elusive. We were not able to investigate the effect of differences in criteria used for the diagnosis of PD because there was substantial variation in the exact inclusion criteria among the studies that reported on the criteria used. However, in most of the studies, the diagnosis was made by a physician and included the presence of two or three of the cardinal symptoms of PD, often with some additional inclusion and exclusion criteria. Variations in participation rates could also contribute to study heterogeneity. The ability to investigate this factor was limited because only 13 of the case–control studies reported participation rates. The same is true for differences across sex. Only 8 studies showed separate results for men and women, but the results were not conclusive: RRs were higher for men than for women in 3 studies, higher for women than men in 3 other studies, and comparable between men and women in the remaining 2 studies. Heterogeneity in the results could also arise from both quantitative and qualitative differences in types of agriculture in the study areas. Although we compared large regions (i.e., North America, Europe, and other), this analysis would not have captured regional differences in the types of agriculture and pesticides used. Analyses by time periods might provide some clues because pesticide use has changed over the decades, but data were insufficient to perform a meaningful analysis of changes over time.

19Our overall summary risk estimates strongly suggest that exposure to pesticides, and to herbicides and/or insecticides in particular, increases the risk of developing PD. Heterogeneity among study-specific RRs could not easily be explained by methodological differences, except for a suggestive effect of exposure assessment characteristics. Future studies should therefore focus on using more objective semiquantitative methods for exposure assessment such as job- or crop-exposure matrices, rather than relying solely on self-report. Although classes of pesticides have been linked to PD, it remains important to identify the specific chemicals responsible for this association. Therefore, in new, preferably prospective studies, attention should be given to collecting detailed information on specific pesticide use.