The impact of pesticide suicide on the geographic distribution of suicide in Taiwan: a spatial analysis

Suicide and population data

Suicide data for people aged 15+ years were extracted from the Taiwanese national mortality data file for years 2002-2009. Suicide by pesticide poisoning was identified using the International Classification of Diseases, Tenth Revision (ICD-10) codes X68 (suicide by pesticide poisoning), Y18 (undetermined death by pesticide poisoning) and X48 (accident by pesticide poisoning), as the latter two cause-of-death categories contain probable 'missed' suicides in Taiwan [14]. Total suicides (all methods) were identified using the following ICD-10 codes: X60-X84 (suicide), Y10-Y34 (undetermined death), W75-W76 (accident by suffocation) and X48 (accident by pesticide poisoning) [14]. 'Non-pesticide' suicides were suicides by methods other than pesticide poisoning; these methods included i) poisoning using solid or liquid substances other than pesticides (X60-X65, X69, Y10-Y15, Y19); ii) poisoning using non-domestic gases (X67 and Y17); iii) hanging (X70, Y20, W75-W76); iv) drowning (X71 and Y21); v) jumping (X80 and Y30); and vi) all other methods. Each suicide was assigned to one of Taiwan's 358 administrative districts (median population aged 15+ years = 27 000) according to the individual's registered residence as recorded on the death certificate. Sensitivity analyses were conducted by restricting the analyses to certified suicides only (X60-X84).

Population data were obtained from the Demographic Fact Books (2002-2009) published by Taiwan's Ministry of the Interior (http://www.moi.gov.tw/stat/year.aspx). These provide annual population statistics for each district, based on the census and the national household registration system.

Rural-urban definition

Area characteristics

To investigate the associations between the proportion of people working in farming in a particular area and the incidence of pesticide suicide in that area, data for the proportion of workers involved in agriculture in each district were obtained from the most recent (2000) census [17]. In the main multivariable analyses we assessed the possible confounding effects of two area factors: the proportion of lone-parent households (an indicator of social fragmentation) and median household income (an indicator of poverty) because a recent study showed that these factors are strongly associated with small-area variations in overall suicide rates in Taiwan [18]. In a sensitivity analysis we assessed the effect of controlling for additional variables associated with area suicide rates [18] - population mobility (i.e. the proportion of people who moved in or moved out from the district during 2000) and the proportion of divorced/separated population (also an indicator of social fragmentation) - as they both showed weak associations with area suicide rates in our previous analysis. In a separate sensitivity analysis we additionally controlled for area population density, which is also strongly associated with area suicide rates in Taiwan [18]; however, it should be noted that this variable was highly correlated with the proportion of people working in agriculture (r = -0.73) rendering interpretation of independent associations of population density/the size of the agricultural workforce problematic due to collinearity [19]. Data on lone-parent households and divorced/separated population for 2000 were extracted from the census [17] and those of median household income were from the Income Tax Statistics [20]. Data for population mobility and population density were extracted from the Taiwanese Demographic Fact Books (2000).

Although suicide data were from 2002-2009, rural-urban definitions and area characteristics were based on data from 2000 because that was the most recent census information available at the time of the study. All data used in the study are publicly available but access to the suicide and census data is subject to review by the Taiwanese government. This study used officially published data and therefore ethical approval is not required.

Statistical analysis

'Raw' (unsmoothed) standardized mortality ratios (SMRs), i.e. the ratio of the observed to expected number of pesticide suicides, were calculated for each of the 358 administrative districts for those aged 15+ years. The 'expected' number of suicides in each area was estimated by multiplying the national sex-age-specific suicide rates (in 5-year age-bands) and the corresponding sex-age-specific population in each district. Data over the 8-year period (2002-2009) were aggregated to ensure sufficient events in each area. However, SMR estimates can still be unstable in small areas/populations as suicide is a relatively rare outcome (annual incidence in Taiwan = 24.5 per 100 000 per year) - a small change in the number of suicides may have a marked impact on the SMR estimates [21]. To address this issue Bayesian hierarchical models were used to estimate the 'smoothed' SMR for each district. The 'smoothed' estimates are the weighted averages of the SMRs for the area of interest and the mean risk in the neighboring areas or in the whole region studied, with the weights depending on the variability observed among areas of the study region. In this way, smoothed estimates of SMRs for each area can 'borrow strength' from data in the neighboring areas which share similar characteristics or the whole region being studied [22]. This improves the precision of the SMR estimates and takes account of small area variations in rates, thus enabling an investigation into the real differences in suicide risk between geographic areas.

The Bayesian hierarchical models were based on a Poisson assumption for the observed number of suicides, with random effects allowing for non-structural variability (heterogeneity across all areas in the study region) and structural variability (autocorrelation between neighboring areas) [21, 23]. Sets of districts that share a border were defined as neighboring areas. Bayesian hierarchical models were estimated using the Markov chain Monte Carlo methods implemented in WinBUGS version 1.4. Vague prior distributions were used, and convergence of the simulations was assessed using the Gelman-Rubin statistic [24], based on three parallel chains. The built-in conditional autoregressive (CAR) distribution in WinBUGS was used to specify the spatially structural variability [25]. Hierarchical models were also constructed to investigate the associations of area characteristics with suicide both before and after controlling for all other characteristics. All analyses used standardized levels (z scores) of area characteristics or their log-transformed values when the distributions of raw values were positively skewed.

Method-specific age-standardized suicide rates were calculated for rural and urban areas, as well as areas in the seven urbanization levels, based on the World Health Organization (WHO) world standard population [26].

Mapping

Choropleth maps, with different shades of color representing different incidence rates, were used to represent the geographic patterning of suicide. A divergent red-blue color scheme was used [27]; with varying shades used for areas with high (red) and low (blue) suicide rates; the central (average) category was shaded white. Maps were produced using ArcGIS Version 9.3.

Unsmoothed and smoothed SMRs for each district were ranked into eight categories; each of these was assigned a different color/shade as described above. Since the distribution of SMRs was skewed, symmetric categories on the logarithmic scale centered around 1.00 (the global mean of the study region) were preferred to dividing districts equally into quartiles/quintiles etc., as such categorization is sensitive to the distribution of data and may group values that differ greatly into broad and heterogeneous bands. The central category was chosen to include a range spanning 0.5 standard deviation below one to 0.5 standard deviation above one (i.e. SMRs of 0.9-1.1). Second, the most extreme categories at two ends of the distribution represented rate ratios that were a half or a double of the national average (i.e. < 0.5 and > 2.0 respectively). Last, the intermediate two breaks between the middle and the most extreme categories were separated at values that were two standard deviations away from one (i.e. 0.5-0.67 and 0.67-0.9 below one, and 1.1-1.5 and 1.5-2.0 above one). These resulted in seven categories that are symmetrical on the logarithmic scale (< 0.5, 0.5-0.67, 0.67-0.9, 0.9-1.1, 1.1-1.5, 1.5-2.0 and 2.0-5.0); one additional category was used to represent the highest values (> 5.0).

Spatial patterning of suicide

Taiwan's urban region includes most of the western part of the island, with four major cities - Taipei in the north, Taichung in middle Taiwan, and Tainan and Kaohsiung in the southwest; population aged 15+ in these four cities was 2 176 000, 815 000, 617 000 and 1 244 000 respectively. Hualien (population aged 15+ = 87 000) is the only city in East Taiwan, which otherwise comprises mostly rural and mountainous areas.

Between 2002 and 2009 there were 36 110 suicides in Taiwan; 4913 (13.6%) of these were pesticide poisonings. 3950 (80.4%) of the pesticide deaths were certified as suicides, 639 (13.0%) as undetermined deaths and 324 (6.6%) as accidents. The proportion of males among the pesticide suicides (71.7%) was similar to that for non-pesticide suicides (68.2%), though pesticide suicides tended to be older (mean age 55.1 years) than non-pesticide suicides (mean age 47.7).

As shown in Table 2, the mean and median of raw and smoothed SMRs for pesticide suicides were higher than one (i.e. the national average); this was because a large number of districts (mostly rural) with small populations showed higher than the national average rates while a smaller number of districts that were urban and had much larger populations showed lower than the national average rates.

Figure 1 shows the geographic distributions of smoothed SMRs for (A) pesticide suicide, (B) non-pesticide suicide and (C) overall suicide. The patterns for unsmoothed SMRs were similar but somewhat less clear due to unstable estimates in sparsely populated areas (data not shown). Higher rates of pesticide suicide were found in East and Central Taiwan, with a concentration of the highest rates in the most rural areas (Figure 1A). One hundred and thirty (36.6%) districts had a SMR above two; they accounted for nearly half (46.6%) of all pesticide suicides, but only 13.4% of Taiwan's population aged 15+ live in these areas. In contrast, five major cities showed the lowest rates; they covered 26.8% of overall population but accounted for only 6% of pesticide suicides.

In contrast to the distinct geographic pattern of pesticide suicide, there was no strong spatial patterning for non-pesticide suicide, although above average rates were found in Kaohsiung city and some areas surrounding Taipei city (Figure 1B). The geographic distribution of overall suicides (Figure 1C) was more similar to that of pesticide suicides than non-pesticide suicides, indicating the impact of pesticide suicide on the overall spatial patterning of suicide in Taiwan.

In the sensitivity analyses including certified suicides only, the main features of the spatial patterning of pesticide, non-pesticide and overall suicides were similar to those based on data for certified and possible suicides combined (data not shown).

Rural-urban differences

Ecological analyses

Ecological analyses showed that a district's pesticide suicide rate was strongly associated with the proportion of the workforce involved in agriculture (Table 4). Figure 2 shows the maps for the proportions of workforce involved in agriculture, lone-parent households and median household income across Taiwan's 358 districts in 2000. In the unadjusted model, one standard deviation increase in the proportion of workforce in agriculture was associated with a 78% increase in district rates of pesticide suicide (rate ratio [RR] = 1.78, 95% Credible Interval [CrI] 1.63-1.94), although the strength of association decreased after controlling for an area's levels of lone-parent households and median household income (RR = 1.58, 95% CrI 1.44-1.74). When additionally controlling for area levels of divorced/separated population and population mobility in a sensitivity analysis, the strength of the association further decreased (RR = 1.46, 95% CrI 1.31-1.61). In another sensitivity analysis that additionally controlled for an area's level of population density, the association of pesticide suicide rates with the proportions of workforce involved in agriculture attenuated markedly (RR = 1.24, 95% CrI 1.10-1.38), but this may be an underestimate as population density was strongly correlated with an area's level of agricultural workforce (Pearson correlation coefficient = -0.73) and likely to be a partial indicator of agricultural workforce and exposure to pesticides in itself.

Area characteristics	Rate ratio (95% Credible Intervals), unadjusted	Rate ratio (95% Credible Intervals), adjusted for all other characteristics
Pesticide suicideb
Agricultural workforce (%)	1.78 (1.63, 1.94)	1.58 (1.44, 1.74)
Lone-parent households (%)e	1.06 (0.97, 1.17)	1.09 (1.10, 1.17)
Median household income (1000 Taiwan Dollar)e	0.63 (0.58, 0.68)	0.75 (0.69, 0.82)
Non-pesticide suicidec
Agricultural workforce (%)	0.93 (0.90, 0.96)	0.91 (0.87, 0.95)
Lone-parent households (%)e	1.12 (1.08, 1.15)	1.10 (1.06, 1.13)
Median household income (1000 Taiwan Dollar)e	0.98 (0.96, 1.10)	0.95 (0.92, 0.98)
Overall suicided
Agricultural workforce (%)	1.09 (1.06, 1.13)	1.06 (1.03, 1.10)
Lone-parent households (%)e	1.10 (1.07, 1.13)	1.10 (1.07, 1.13)
Median household income (1000 Taiwan Dollar)e	0.90 (0.88, 0.92)	0.94 (0.91, 0.96)

^aStandard deviation (SDs for agricultural workforce, lone-parent households and median household income were 16.4%, 1.3% and 78,000 Taiwan dollar [= 2500 US dollar] respectively)

^bIncluding certified suicides, undetermined deaths and accidents from pesticide poisoning

^cIncluding certified suicides, undetermined deaths and accidents from suffocation

^dIncluding certified suicides, undetermined deaths and accidents from pesticide poisoning/suffocation

^eValues were log-transformed before standardization due to their skewed distributions

In contrast, rates of non-pesticide suicide were lower in districts with a higher proportion of workforce involved in agriculture (unadjusted RR = 0.93, 95% CrI 0.90-0.96; adjusted RR = 0.91, 95% CrI 0.87-0.95). Overall suicide rates were positively associated with an area's level of agricultural workforce (unadjusted RR = 1.09, 95% CrI 1.06-1.13; adjusted RR = 1.06, 95% CrI 1.03-1.10), although the strengths of associations were much weaker than those for pesticide suicide.

In the sensitivity analyses including certified suicides only, results were similar to those based on certified and possible suicides combined (detailed data not shown). In the models controlling for lone-parent households and median household income, districts where a higher proportion of population worked in agriculture had higher rates of pesticide suicide (adjusted RR = 1.76, 95% CrI 1.48-2.09), lower rates of non-pesticide suicide (adjusted RR = 0.92, 95% CrI 0.87-0.99) and higher overall suicide rates (adjusted RR = 1.09, 95% CrI 1.03-1.16).

Strengths and limitations

To our knowledge this is the first nationwide analysis of small-area variations in suicide from pesticide poisoning. The availability of small-area data allowed a detailed investigation of the spatial distribution of pesticide suicide. The study has a number of limitations. First, data for non-fatal self-poisoning using pesticides were unavailable. This study was unable to determine whether the spatial patterning of pesticide suicide was due to variations in the incidence of self-poisoning or in case fatality following pesticide poisoning; area differences in case-fatality could be caused by area differences in access to medical services or variations in the toxicity of the pesticides commonly taken in episodes of self-poisoning (due to variations in agricultural practices). Second, an area's proportion of workforce in agriculture was used as the proxy indicator of access to pesticides; other variables may be better markers of pesticide accessibility, e.g. pesticide sales, for which small-area data were unavailable. Third, we aggregated data for 8 years (2002-2009) to ensure a sufficient number of events in small areas. Such aggregation may mask area-specific changes in suicide rates over this period; analysis of relevant data revealed that although pesticide suicide rates did decline over this period, the reduction was similar in rural and urban areas (32% and 33% respectively). Last, there are geographic variations in the quality of suicide statistics in Taiwan [28]; however, results were similar when using data for certified suicides only or certified and possible suicides combined.

Comparison with previous studies

The study showed substantial geographic variation in the incidence of pesticide suicide and its impact on the overall spatial patterning of suicide in Taiwan. In contrast, non-pesticide suicides showed no strong spatial patterning. A recent study from Taiwan showed that self-poisoning using solid or liquid substances demonstrated the greatest geographic variations compared with other common suicide methods [18]; the present study showed that 69% of such solids/liquids poisoning suicides were due to pesticides. This study also showed that pesticide suicide contributed to most of the rural-urban disparity of suicide in Taiwan. A recent study demonstrated that such disparity had attenuated over the last decade after the emergence of charcoal-burning suicides around 1998-2000, which concentrated in mostly urban areas [29].

Areas with a high proportion of their labour force working in agriculture experienced higher rates of pesticide suicide, but not non-pesticide suicide, both before and after controlling for area socioeconomic disadvantage. The association suggested the role of means availability in the spatial distribution of pesticide suicides - suicide rates were higher in areas where there was easier access to pesticides, indicated by a higher proportion of people working in agriculture. Likewise, an ecological analysis in South Korea found strong association of pesticide self-poisoning mortality rates with an index of farming activities and population [12]. In contrast, a recent study in southern rural Sri Lanka found that areas with a high proportion of population working in agriculture had low rates of self-poisoning, although a greater proportion of episodes in these areas involved pesticides [10]. However, this study was limited to investigating only a relatively small, mainly agricultural area in southern Sri Lanka. The socioeconomic characteristics of workers in this area are likely to be quite different from those in Taiwan and South Korea. Furthermore, farmers in different countries may be involved in different types of cultivation with varying requirements of pesticide application, and thus the 'population working in agriculture' may be a good indicator of pesticide accessibility in some but not other countries/regions. Recent person-based studies in rural China provided further evidence for pesticide access as an independent risk factor for suicide in residents controlling for individuals' socioeconomic conditions and mental disorders [30].

Public health implications

The WHO acknowledges that pesticide ingestion is among the mostly frequently used methods of suicide worldwide [31]. Pesticides are readily available in rural areas of developing countries and commonly used in impulsive acts of self-poisoning following acute life crisis [32, 33]. This leads to many tragic deaths due to the high case fatality associated with pesticide ingestion [34]. One strategy to reduce these deaths is to restrict the availability of toxic pesticides. Studies in countries such as Finland [3] and Sri Lanka [4] showed marked reductions in pesticide suicides after restricting their use. In keeping with these, the present study showed that the spatial distribution of pesticide suicide was strongly associated with the distribution of agricultural workforce, an indicator of pesticide availability/accessibility. Other strategies for preventing pesticide suicides included safe storage in lockable boxes [35, 36], centralized communal storage, education to users, retailers and community leaders [37], stopped using all pesticides by adopting non-pesticide management policy [38] and improved medical treatment of pesticide poisoning [39]. A study of the effectiveness of locked boxes is currently underway in Sri Lanka [40]. It is not yet clear which intervention is the most promising [41] because no comparative cost-effectiveness analysis is possible at the moment. A hazard reduction approach suggests that banning the more toxic pesticides from agricultural practice would be most effective [42]. However, the agricultural costs and the healthcare benefits of switching to less toxic pesticides are not currently known.

This study showed that, in Taiwan, some areas had disproportionately high numbers of pesticide suicides. In the eastern mountainous part and central rural areas of Taiwan, where rates of pesticide suicide were high, strategies such as restricting access to the highly toxic pesticides and improving the medical care of self-poisoning may reduce suicide mortality. Such prevention strategies may help to tackle inequality in suicide burden in Taiwan as pesticide suicide was the major contributor to geographic variations in overall suicide rates. Identification of areas with high rates of pesticide suicide will inform the allocation of prevention resources targeted at reducing these deaths.