Statistical analysis

The prevalence of most individual mental disorders was too low to permit comparison of rates in the case and control groups across different demographic cohorts, so we combined all specific diagnoses into six diagnostic groups.

1. (a) Organic mental disorders, including dementia and intellectual disability.

2. (b) Substance use disorders, including alcohol and other substance use and dependence.

3. (c) Psychotic disorders – schizophrenia and associated disorders.

4. (d) Mood disorders, including bipolar disorder, depressive disorder and dysthymic disorder.

5. (e) Anxiety disorders, including generalised anxiety disorder, panic, phobias, obsessive–compulsive disorder, post-traumatic stress disorder and acute stress disorder.

6. (f) Other mental disorders, including borderline personality disorder, conversion disorder, hypochondriasis, adjustment disorder, pathological gambling disorder and miscellaneous disorders.

Individuals who had current co-occurring diagnoses from different groups of disorders were coded as having both types of disorders.

Chi-squared tests were used to compare the rates of specific disorders and groups of disorders between the case and control groups. Logistic regression models were used to estimate crude odds ratios, adjusted odds ratios and associated 95% confidence intervals for suicide for the different diagnostic groups before and after stratification by gender, age group (over or under the median age of 40), and urban (i.e. city and town) versus rural (i.e. village) residence. Only one injury death decedent had an anxiety disorder making it impossible to compute both odds ratios in the stratified cohort analyses, so it was necessary to combine anxiety disorders with other mental disorders in these analyses. The median split was used for age to maximise sample size in the subgroups; a separate analysis was performed comparing rates of any mental disorder in the case group with the controls in six age categories (10–19, 20–29, 30–44, 45–59, 60–74, 75+), but it was not possible to do this for separate diagnostic groups because some age groups had no individuals with the corresponding diagnoses. Comparison of the risk of suicide between cohorts (e.g. comparing the risk of suicide among men with mood disorders to that of women with mood disorders) was assessed by computing the ratio of two adjusted odds ratios (i.e. the odds ratio of suicide among men with mood disorders divided by that for women) and the significance of the ratio was assessed by adding the corresponding interaction term (mood disorder×gender) to the logistic model. ^{Reference Altman and Bland18}

The study was approved by the institutional review boards of the Beijing Hui Long Guan Hospital and the Chinese Academy of Preventive Medicine.

Main findings

In accordance with our a priori hypothesis, this study found that both the prevalence of different categories of mental disorders among suicide decedents and the risk of suicide among people with different types of mental disorders vary across demographic groups and by urban versus rural residence. Previous studies have suggested that risk factors for suicide may not be uniform in all population cohorts. ^{Reference Hunt, Kapur, Webb, Robinson, Burns and Turnbull19–Reference Zivin, Kim, McCarthy, Austin, Hoggatt and Walters21} We have advanced this discussion by comparing cohort-specific odds ratios in different categories of mental disorders for three factors that are potential moderators of the risk of suicide – gender, age and urban versus rural residence.

The overall prevalence of mental disorders in the suicide decedents (63%) is substantially lower than the prevalence reported in Western psychological autopsy studies ^{Reference Arsenault-Lapierre, Kim and Turecki2,Reference Mortensen, Agerbo, Erikson, Qin and Westergaard-Nielsen3,Reference Wærn, Runeson, Allebeck, Beskow, Rubenowitz and Skoog6} but somewhat higher than that reported by another psychological autopsy study in China, ^{Reference Zhang, Wieczorek, Conwell, Tu, Wu and Xiao22} which used the SCID to assess young rural suicide decedents and found that only 55% of male decedents and 39% of female decedents had a mental illness at the time of the suicide. The reasons for the lower prevalence of mental disorders in Chinese suicide decedents are unknown but we believe it is related to higher levels of impulsive behaviour combined with easy access to highly lethal means (i.e. pesticides), particularly in rural areas.

Despite the lower prevalence, the relative importance of different classes of mental disorders as risk factors for suicide found in this study is comparable to that seen in high-income countries. Similar to other studies, ^{Reference Wærn, Runeson, Allebeck, Beskow, Rubenowitz and Skoog6,Reference Ran, Wu, Conwell, Chen and Chan8} mood disorders are, by far, the most important category of mental disorders. The relatively less important (although statistically significant) role of substance use disorders seen in this study compared with that reported in high-income countries ^{Reference Wærn, Runeson, Allebeck, Beskow, Rubenowitz and Skoog6,Reference Cherpitel, Borges and Wilcox23} is partly due to the lower prevalence of these conditions in China (although rates of alcohol misuse have been increasing recently). ^{Reference Phillips, Zhang, Shi, Song, Ding and Pang14,Reference Hao, Young, Xiao, Li and Zhang24,Reference Hao, Su, Liu, Zhang, Yang and Chen25} Another possibility is that higher rates of substance misuse in accidental death decedents than in the general population could result in an underestimate of the odds ratio for substance misuse in suicide. Given the similar prevalence of substance use disorders reported in males in the general population ^{Reference Phillips, Zhang, Shi, Song, Ding and Pang14} to that found in our male accidental decedents (11.3% v. 10.0%), this does not appear to be the case for men. However, the prevalence of substance misuse in female accidental decedents (1.8%) is substantially higher than reported in females in the general population ^{Reference Phillips, Zhang, Shi, Song, Ding and Pang14} (0.3%), so the reported odds ratio for substance misuse in females (1.77) may be an underestimate.

Our results highlight the important distinction between population-based attributable risk and individual-based clinical risk. For example, we found that the prevalence of psychotic disorders in female suicide decedents is double that in male suicide decedents (12.9% v. 6.4%), so psychotic disorders are a relatively more important cause of suicide in women than in men. ^{Reference Phillips, Yang, Li and Li26} But, similar to other studies in China, ^{Reference Ran, Chen, Conwell, Chan, Yip and Xiang27} we found that the risk of suicide among males with psychotic disorders is 3.4-fold higher than that of females with psychotic disorders so clinicians need to have a greater ‘index of suspicion’ for males with psychotic disorders. Thus the target of universal preventive steps aimed at the general population may need to be different from the focus for selective and indicated interventions and treatments aimed at high-risk groups and individuals.

The most important specific finding was that mental illnesses are much less important as risk factors for suicide among rural residents than among urban residents. The adjusted odds ratio for suicide for rural residents with any mental disorder was 7.6 compared to an odds ratio of 15.5 for urban residents with mental disorders (P = 0.008). Previous reports from China have shown major differences in the characteristics of urban versus rural suicide decedents ^{Reference Yang, Phillips, Zhou, Wang, Zhang and Xu12} and the important role of impulsivity in the suicides of both men and women from rural communities. ^{Reference Conner, Phillips, Meldrum, Knox, Zhang and Yang28} The current analysis found that, with the exception of substance use disorders (which conferred a greater risk of suicide among rural residents), the risk of suicide among individuals with mood disorders, psychotic disorders, organic mental disorders or other mental disorders (including anxiety disorders) living in urban areas was more than double the risk among individuals with these disorders living in rural areas. The prevalence of mental disorders in rural versus urban injury death decedents was similar (17.8% v. 16.4%), so it is unlikely that the much lower risk of suicide among rural residents with mental disorders is the result of underdiagnosis in rural decedents. Nor is it likely that the lower suicide risk associated with mental illness is the result of better services and support for the mentally ill in rural areas, many of which lack any mental health services whatsoever. We believe this is a robust finding with important theoretical and practical implications.

Urban versus rural residence is a composite indicator that captures a range of social and contextual variables relevant to suicide (e.g. economic and educational status, density of family relationships, employment, access to means of suicide, access to health resources), but it is not included in most multidimensional models of suicide, ^{Reference Maris9,Reference Wasserman and Wasserman10} possibly because these models are primarily based on work in high-income countries where a much smaller proportion of the population live in rural areas and, perhaps, where the urban–rural divide is not as pronounced as in low- and middle-income countries. Based on our findings, we hypothesise that urban versus rural residence explains a substantial proportion of the within-country variance in suicide rates in low- and middle-income countries with large agriculture-based populations and that it is a more powerful predictor of suicide rates than its component variables (e.g. economic level, education, employment). The relationship of urban versus rural residence to suicide risk may become attenuated with urbanisation and with high rates of rural-to-urban work migration, so eventually it may be more fruitful to dissect it into its component variables. But even as the urbanisation process progresses, major urban versus rural differences in social structure and healthcare systems will persist, so it will remain an important dimension that is relevant to the provision of preventive and treatment services for suicide.

Implications

Our results suggest that the network of causal factors leading to suicide in rural areas is substantially different than that in urban areas. The ready availability of potent pesticides in many rural households may increase the likelihood that impulsive, low-intent suicidal acts following intense interpersonal conflicts among people without a diagnosable mental illness result in death. ^{Reference Zhang, Wieczorek, Conwell, Tu, Wu and Xiao22} If this is the case, then the focus of preventive efforts in rural areas should be on providing educational and social interventions aimed at improving coping skills and conflict resolution skills in community members; and on decreasing access to pesticides, reducing the lethality of pesticides, promoting public health campaigns about the safe use and storage of pesticides, and improving the treatment of pesticide poisoning. ^{Reference Gunnell, Eddleston, Phillips and Konradsen29} The lack of adequate mental health services in rural areas is a common problem in low- and middle-income countries that urgently needs to be addressed, ³⁰ but improving rural mental health services should only be one component – and not necessarily the most important component – of suicide prevention efforts in these countries.

Strengths and limitations

Several factors need to be taken into consideration when interpreting these results. The demographic profiles of mental disorders ^{Reference Phillips, Zhang, Shi, Song, Ding and Pang14} and of suicides in China ^{Reference Phillips, Li and Zhang4,Reference Phillips, Yang, Zhang, Wang, Ji and Zhou5} are quite different from that reported in most high-income countries, so the relationship between suicide risk, diagnostic groups and the three factors considered (gender, age and urban versus rural residence) may be different in China than elsewhere.

Despite using data from the largest case–control psychological autopsy yet reported, the low prevalence of some mental disorders in the injury death control group made it impossible to conduct some of the stratified analyses and resulted in large confidence intervals around some of the estimated odds ratios, increasing the risk of type II errors (i.e. some important differences did not meet the criteria of statistical significance). Overcoming this problem will require conducting meta-analyses that combine results from several large psychological autopsy studies that have used comparable methods.

As a result of the sample size considerations we arbitrarily dichotomised age by the median age and found few differences in the risk of suicide associated with different mental disorders between persons under or over 40 years of age. Our analysis by six age groups found age-related differences in risk of suicide among people with any mental disorder when comparing more narrowly defined age ranges, but the numbers of individuals in the age-specific cohorts were too few to reach statistical significance. There may also be important diagnostic group-specific differences in suicide risk in different age groups that we were unable to assess.

There are several other demographic and clinical factors which may moderate the relationship of suicide risk and mental illnesses that we did not consider: for example, education, income, marital status, prior suicide attempt, and family history of suicide. Given the practical problems associated with travelling to 23 sites around the country, the time from death to interview (median 11 months) was longer than in most psychological autopsy studies; ^{Reference Mortensen, Agerbo, Erikson, Qin and Westergaard-Nielsen3,Reference Wærn, Runeson, Allebeck, Beskow, Rubenowitz and Skoog6} but there was no significant difference in the time to interview between those in the case group and the controls, and the prevalence of mental disorders was similar in cases interviewed early and those interviewed later, so it is unlikely that the relatively long time to interview affected the reported results.

As discussed in the Method, the use of deceased controls has the disadvantage that potential risk factors for suicide may also be risk factors for other causes of death that could lead to an underestimation of the importance of such factors. But the overall prevalence of mental disorders among men and women in the injury death group (17.5% and 17.2% respectively) are similar to those found in a large community-based epidemiological study in four Chinese provinces that used similar diagnostic methods ^{Reference Phillips, Zhang, Shi, Song, Ding and Pang14} (20.1% and 14.8% respectively), so it does not appear that accidental death decedents have unusually high rates of mental illness that would obscure the importance of mental illnesses in suicide.

The specific findings of this analysis may or may not hold true in other countries where the characteristics of suicide and the demographic profile of mental illnesses are different. Similar analyses of large case–control psychological autopsy studies are needed to assess the generalisability of our specific findings. The core finding that we believe to be widely relevant is that the importance of mental illnesses as risk factors for suicide varies widely both by the type of mental illness and by demographic cohort. We also believe it essential to consider urban versus rural residence as a key variable in the understanding and management of suicide, particularly in low- and middle-income countries.