To the best of our knowledge, this is the first analysis of the geographic variations in the incidence of self-poisoning in rural Asia. Likewise, it is the first study in rural Asia to analyse the association of area-based socioeconomic and agricultural characteristics with the incidence of self-poisoning.
The overall rate of intentional self-poisoning in the study area was 315 per 100,000 inhabitants, confirming the high rates reported from other Sri Lankan studies [14, 15]. The distribution of cases varied considerably across GNs (range 0–2168 per 100,000, median 234 per 100,000), and the disproportionately high contribution to self-poisoning incidence by a fairly small number of GNs suggests that the spatial variation may not be random. This is supported by the fact that nine of the top-incidence GNs in 2002 also ranked in the top 10% of self-poisoning incidence GNs in each of the preceding 3 years (1999–2001) . However, as the numbers of self-poisoning cases per GN predominately remain quite small, no strong conclusions can be drawn.
Our investigations of the association of self-poisoning with area level markers of socioeconomic and agricultural conditions yielded surprising results. For instance, the finding that areas characterised by high levels of socioeconomic deprivation tended to have lower rates of self-poisoning is at odds with the research from high-income countries where area levels of socioeconomic deprivation are in general positively associated with risk of self-harm [9, 17]. The role of socioeconomic deprivation in the aetiology of self-harm may well be different in rural Asia. Certainly, the causal pathway is complex, as suggested in this study by the divergence in associations of the different measures of socioeconomic status (education and housing vs. unemployment). Previous Sri Lankan studies support this view. For example, in a case control study of self-harm conducted in 1999, Thalagala and Fernando  reported an increased risk among those with higher levels of education (OR 1.8 95%CI 1.1 to 2.9), although low income was also associated with increased risk. Van der Hoek and Konradsen  in their case control study of pesticide self-poisoning conducted within the current study area reported no association between socioeconomic position or debt and self-poisoning, although they did find a protective effect of higher levels of education. The only previous ecological study of suicidal behaviour in Sri Lanka  examined associations with suicide at a high level of geographic aggregation (22 districts of Sri Lanka, total population 12,7 million). The author found no association between literacy or urbanization with suicide, but reported that high levels of unemployment were associated with lower suicide rates.
It is possible that different markers of socioeconomic deprivation are associated with self-harm through separate pathways and also that the effect of the various socioeconomic indicators may vary from context to context. For instance, in farming communities, unemployment may be of a seasonal nature affecting large parts of the community at the same time, thereby entailing fewer social and psychological consequences than in urbanised areas. Similarly, in theory, the composite indicator 'housing quality' may not expose the same aspects of socioeconomic status in urban and rural areas due to differences either in construction costs or community availability of electricity and protected drinking water.
Future studies of self-harm in developing countries should assess such issues and consider plausible explanatory pathways using individual level data.
The role of the agricultural factors is similarly complex. We found that GNs with a relatively high proportion of the population involved in agriculture had lower rates of self-poisoning, while the reverse association was seen in relation to the proportion of cultivated land.
This may be due to the variables not, in fact, being clear indicators of pesticide accessibility. 'Agricultural land-use' covers different types of crop with varying requirements of pesticide application, while 'population employed in agriculture' similarly includes farmers involved in different types of cultivation. Detailed information on specific crop-types from agricultural surveys as employed in a methodologically comparable study from Brazil  might have provided more accurate indicators of pesticide usage. Lin and Lu  have previously used 'percentage of the population employed in agriculture' as an indicator of pesticide accessibility in Taiwan. A strong positive association of this variable with suicide by intentional ingestion of pesticides, led the authors to conclude that the easy availability of pesticides influenced their usage for self-harm. In support of this, our findings show that the relative contribution of pesticide ingestion in self-poisoning episodes was larger in areas with large agricultural populations.
Strengths and Limitations
We carried out detailed medical record searches to identify all cases of self-poisoning and used small-area level census and satellite image data to characterise the study area. The last census in Sri Lanka was carried out in 2001 and satellite images were taken in 2003, so our data on the characteristics of the study area correspond closely in time to the measurement of intentional self-poisoning. The low level of aggregation (median population 1416) more accurately describes the ecological exposure of the individuals studied, reducing though not entirely avoiding the inherent susceptibility of area-based studies using aggregate data to ecological bias. While using smaller areas for analysis gave a higher statistical power to detect associations (due to a larger number of analysed units), the small number of events per GN reduced precision of the incidence rate estimates. A further concern is the issue of spatial autocorrelation, if the patterns of self-poisoning in a GN are influenced by self-poisoning in neighbouring areas simply by virtue of their spatial proximity. Such correlation may potentially affect the results of the regression analyses, and it would be useful to assess and control for this in future studies .
There are a number of further limitations to our analysis. First, hospital admission data may not satisfactorily capture the incidence of self-poisoning in the community. Varying accuracy in hospital records may affect the identification of cases both with regard to whether self-poisoning was intentional or accidental, and with regard to the type of poison ingested. Also, hospital presentation may not occur in cases of minor poisoning or those episodes resulting in death before hospital treatment can be sought, while some severe poisonings may have been taken directly to larger better-equipped hospitals outside the study area. The associations with area characteristics are, however, only likely to be biased if missing data are more common in certain areas. For instance, if the lower rates of self-poisoning in areas of socioeconomic deprivation were observed simply due to reduced hospital access in such areas the resultant associations might entirely be the effect of selection bias. However, we find no reason to suppose that access to hospitals or treatment seeking might vary across the study area; hospital care is provided free of charge and all hospitals lie within a short and easy travelling distance of each other. Furthermore, our sensitivity analysis did not suggest an effect in peripheral GNs, where cross boundary treatment seeking might otherwise have been an issue. Second, we were unable to clearly identify repeated admissions for the same individual over the 12-month study period and this could potentially inflate rate estimates in some areas. Careful scrutiny of the case data with regard to sex, age and place of residence did not, however, suggest repetition of self-poisoning to be a problem. Third, some difficulties were encountered in assigning cases to GN-divisions. Any resulting misclassification is, however, likely to be random with respect to the exposure variables examined and so lead to an under-estimation of any associations.