Maternal residential pesticide use and risk of childhood leukemia in Costa Rica

Study population

Subject recruitment and procedures for the CRCLS have been described elsewhere.22 Briefly, all cases of childhood leukemia (ages 0–14 years at diagnosis, n = 334) diagnosed between 1995 and 2000 in Costa Rica were identified using information from the Costa Rican Cancer Registry and the National Children's Hospital. Consent to participate in CRCLS was obtained from 90% of eligible cases. Population controls were randomly selected from the National Birth Registry and frequency-matched to cases by birth year. We used information reported by the mother at the time of birth to identify addresses of controls. Because addresses in the National Birth Registry were sometimes restricted to neighborhoods, we also used national electoral databases and local social security clinics to ascertain the exact address of 62% of potential controls. If the control could not be located (either because the exact address was not identified or no one was available at the time of the home visit) or the control refused to participate, we randomly selected a new control of the same age living in the same neighborhood as the originally selected control. Among eligible controls, 91% consented to participate. A total of 300 children with leukemia (252 ALL cases) and 579 healthy controls were enrolled in CRCLS. Once consent was obtained from at least one parent of the child enrolled in the study, we conducted an extensive face-to-face interview with the mother and, when available, the father. The interview included the administration of three questionnaires: one about the mother, one about the father, and one about the child (frequently completed by the mother, and occasionally by the father or a caregiver). Due to the number of participants missing paternal questionnaires (13%) and the etiological differences between ALL and other types of leukemia, we restricted the present analyses to data from maternal questionnaires and included only cases diagnosed with ALL. After excluding two controls with missing maternal exposure data and one case who was missing the date of diagnosis, a total of 251 ALL cases and 577 controls were included in our analyses.

The Ethical Committees of the National Children's Hospital, Ministry of Health of Costa Rica, and Karolinska Institutet in Sweden approved all study materials and procedures. Informed consent was obtained from at least one parent or legal representative for each participant before their data collection began.

Exposure assessment

We collected information on socio-demographic and lifestyle characteristics (e.g., parental age, education, smoking habits, dietary intake, and household income), known or suspected risk factors for childhood leukemia (e.g., X-ray exposure, family history of leukemia, and birth defects), parental medical history, and parental occupational and residential exposure to pesticides and other environmental toxicants. More specifically on residential pesticide exposure, we asked mothers about their home use of pesticides from the following groups: (i) insecticides, (ii) herbicides, (iii) fungicides, and (iv) commercial/professional pest control treatments, as well as exposure to (v) pesticides sprayed by the Ministry of Health for vector control, and (iv) pesticides sprayed on farms or companies near the home. We also asked mothers about the frequency (i.e., daily, several times per week, weekly, several times per month, monthly, every three months, several times per year, once per year, less than once per year) and periods of their pesticide use inside the home [i.e., in the year prior to pregnancy, during pregnancy, during breastfeeding, <1 year, 1–2 years, 2–4 years, ≥5 years but up until leukemia diagnosis for cases and either the interview date or age of 15 years for controls (whichever occurred first), and always].

Within each period of potential exposure, we estimated in a standardized way the average frequency of exposure in days per year in order to account for the differences in length of the exposure period (e.g., the year before pregnancy had a length of 12 months and pregnancy had a length of ∼9 months) and variation among study participants (e.g., the length of the period after birth varied based on the censoring date). We estimated the frequency of pesticide use across periods of exposure by (i) converting each of the individual pesticides that mothers reported using to days per year and then (ii) calculating the average frequency of pesticide use (in days per year) for the five exposure periods (i.e., the year before pregnancy, pregnancy, while breastfeeding, after birth, and at any time) for each of the six pesticide groups. For instance, values of 365 were assigned to pesticides used “daily”, 208.5 to “several times per week”, 52 to “weekly”, 32 to “several times per month”, 12 to “monthly”, 2.5 to “several times per year”, 1 to “once per year”, and 0 to “less than once per year” or “no use”. Values for categories “several times per week”, “several times per month”, “several times per year”, and “less than once per year” were estimated by averaging the values of the categories above and below [e.g., the average of 208.5 days per for “several times per week” was estimated by averaging 365 (value for pesticides used daily) and 52 (value for pesticides used weekly)].

Individuals missing data on the time period in which they were potentially exposed a pesticide were excluded from multivariate analyses for that pesticide group (e.g., 1 case and 1 control for insecticides; 13 cases and 8 controls for pesticides sprayed on nearby farms or companies). Missing values for frequency of maternal use of a pesticide during any exposure period were imputed using the participant's average frequency of exposure for that pesticide group from other exposure periods (when available) or the frequency most commonly reported by other participants for that pesticide group at any exposure period (i.e., one time per year for herbicides, professional fumigation, and fumigation for vector control; 2.5 times per year for insecticides, fungicides, and pesticides sprayed on nearby farms or companies).

Statistical analysis

We estimated bivariate associations between the exposures, outcome, and covariates using t-tests for continuous variables and χ² tests for categorical variables. We used Spearman correlation coefficients (r_s) to examine whether the average frequency of use of one pesticide group was correlated with (i) average frequency of use of other pesticide groups and (ii) average frequency of use of that pesticide group at other time periods. We examined exposure-outcome associations using unconditional logistic regression models to estimate ORs and their 95% CIs. We ran separate regression models for each exposure period and three pesticide groups: herbicides, insecticides, and spraying on farms or companies near the home. We did not fit multivariate regression models for the remaining pesticide groups (i.e., fungicides, pesticides used for professional fumigation, and pesticides sprayed for vector control by the Ministry of Health) due to the small number of mothers of cases and controls who reported these exposures. We adjusted our models for birth year (matching variable at enrollment), and child sex and socioeconomic status [two variables identified a priori that were associated with at least one of the pesticide groups and the outcome of interest in bivariate analyses (p < 0.20)]. Covariates with missing information (i.e., socioeconomic status) were randomly imputed based on observed probability distributions.

Because the average frequency of use for most pesticide groups was low, exposure to each of the groups was dichotomized as none vs. any use. Insecticide use was reported frequently enough to create three categories: low [no use and average use up to 2.5 times per year (which was the median for most exposure periods)], medium [average use between 2.5 and 36 times per year (the latter represented the 75th percentile for most exposure periods)], and high (average use >36 times per year). We examined whether child sex modified the exposure-outcome associations using a product interaction term between child sex and each of the dichotomized exposures and also stratifying by sex. We did not examine sex differences in associations of average frequency of insecticide use (three categories) and childhood leukemia due to our relatively small sample size.

We conducted sensitivity analyses to assess the robustness of our results: (i) we fitted our regression models using socioeconomic status variables imputed in two different ways: by substituting the missing observations with medium socioeconomic status (the most frequent category for both cases and controls) and by creating a new category for the missing values; (ii) we adjusted our models for parental education (highest combined level) instead of socioeconomic status [because we had imputed a relatively large percentage of socioeconomic status missing values (∼20%)]; (iii) we adjusted our regression models for additional potential confounders or strong predictors of ALL [i.e., maternal smoking and alcohol consumption during pregnancy, birth order, birth weight, and breastfeeding] to examine their impact on effect estimates (referred to henceforth as fully-adjusted models); (iv) we included all 299 leukemia cases enrolled in the CRCLS (not just ALL cases) in the unconditional logistic regression models; (v) we assessed the independent effect of multiple exposures by adjusting for exposure to insecticides, herbicides, and pesticides sprayed on farms or companies near the home in the same model; and (vi) we examined the potential impact of the mother being home during the day when agricultural pesticide spraying would be more common by adjusting for parity (1 vs. >1 pregnancies) in models including child sex, year of birth, and socioeconomic status.

Study population characteristics

ALL was the most common type of leukemia in the CRCLS, accounting for nearly 85% of the cases (AML and other types = 10% and 5% of cases, respectively). ALL cases and controls were similar on most attributes, including maternal age at child's birth, parental education, maternal smoking and alcohol consumption during pregnancy, child sex, and birth order (Table 1). Cases were more likely to have been born between 1991 and 1995 (50%) than controls (42%) and to have a lower socioeconomic status (18%) than controls (14%). The average time period between age of diagnosis (censoring date) and age of interview for the cases was 3.9 years. The average age at the time of interview for controls was 11 years.

Insecticides were the pesticides most frequently reported to have been used by mothers inside their homes (between 58% and 84% of participants reported using them across all five exposure periods), whereas fungicide use, professional fumigation, and fumigation for vector control were relatively rare (<10% exposed in at least one period of exposure; Table 2). Self-reported maternal insecticide and herbicide use was relatively similar among cases and controls for all exposure periods. More mothers of cases than controls reported living near farms or companies that sprayed pesticides in all periods of exposure (39% vs. 31%, p = 0.01 at any time period), while more mothers of controls reported living in areas sprayed with pesticides for vector control than mothers of cases after birth (32% vs. 22%, p < 0.01) and at any time period (33% vs. 24%, p < 0.01; Table 2). Average frequencies of maternal use within a specific pesticide group were highly correlated throughout the different exposure periods (r_s = 0.61–0.99); however, average frequencies of use were either weakly or not at all correlated between pesticide groups (see Supporting Information Table 1). For instance, the strongest correlations between pesticide groups were observed for herbicides and spraying on farms or companies near the home (r_s = 0.24–0.30).

Associations between maternal pesticide use inside the home (none vs. any) and childhood ALL

Self-reported maternal insecticide use in the year before pregnancy, during pregnancy, and while breastfeeding was associated with increased odds of ALL among boys [aOR = 1.63 (95% CI: 1.05–2.53), 1.75 (95% CI: 1.13–2.73), and 1.75 (95% CI: 1.12–2.73), respectively, all p-int < 0.20; Table 3]. In contrast, we observed that maternal insecticide use after birth and at any time period were associated with small decreases in odds of ALL among girls [aOR = 0.69 (0.39–1.23) and 0.66 (0.37–1.17), respectively, all p-int > 0.20]. Maternal herbicide use inside the home was not associated with childhood ALL in combined and sex-stratified analyses (Table 3).

Maternal report of pesticides sprayed on farms or companies near the home was also associated with childhood ALL. For example, maternal report of these pesticide applications during pregnancy, while breastfeeding, and during any time period was associated with increased odds of leukemia in combined analyses of boys and girls [aOR = 1.43 (95% CI: 1.00–2.05), 1.41 (95% CI: 0.98–2.03) and 1.52 (95% CI: 1.11–2.09), respectively; Table 3]. In sex-stratified analyses, we observed that these exposure-outcome associations were stronger among boys than among girls, but sex differences were not statistically significant (all p-int > 0.20).

Associations between average frequency of maternal insecticide use inside the home (low, medium, high) and childhood ALL

Children whose mothers reported a high average frequency of insecticide use inside their homes (>36 times/year) in the year before pregnancy, during pregnancy, and while breastfeeding had higher odds of ALL compared to children whose mothers reported a low frequency of insecticide use (<2.5 times/year) during these exposure periods [aOR = 1.56 (95% CI: 1.07–2.27), 1.58 (95% CI: 1.08–2.31), and 1.56 (95% CI: 1.07–2.29), respectively; all p-trend < 0.05; Table 4].

Sensitivity analyses

Effect estimates did not change appreciably after (i) using different methods of imputing/handling missing observations for socioeconomic status, (ii) adjusting for parental education instead of socioeconomic status, or (iii) adjusting regression models for additional potential confounders or strong predictors of the outcome (fully-adjusted models; data not shown). Notably, associations between average frequency of insecticide use (low, medium, high) and childhood leukemia were slightly weaker when we included all leukemia cases (n = 299) in the adjusted models, as compared to analyses with ALL cases only (see Supporting Information Table S2). Results did not change appreciably when we stratified by socioeconomic status or year of birth (See Supplementary Information Tables S3 and S4, respectively). Associations between maternal residential pesticide use (none vs. any) and childhood leukemia did not change when we included all leukemia cases (data not shown). In general, effect estimates were slightly lower when we assessed co-pollutant confounding by adjusting for exposure to insecticides, herbicides, and pesticide spraying on farms or companies near home in the same models. Importantly, associations highlighted previously remained in the same direction (data not shown). Because parity was not associated with the outcome (p = 0.88) and effect estimates did not change appreciably when the variable was included in the models (See Supplementary Information Table S5), we decided to maintain our original covariate selection process and did not include parity in the final models.