• suicide
• unemployment
• socioeconomic status

Many capitalist economies are characterised by business cycles with concomitant increases in joblessness during recessions and depressions, and reductions in unemployment in periods of economic expansion. In view of the potentially debilitating consequences of joblessness on health and related outcomes, research on unemployment and suicide continues to be relevant in both epidemiology and sociology. One controversy that continues is the issue of selection bias. The essential question that remains unresolved is whether the observed association between unemployment and suicide reported in some studies reflects direct causation or whether there is some variable that is causally prior to both unemployment and suicide.

The report by Blakely et al¹ presents an analysis of the New Zealand Census Mortality Study (NZCMS) that attempts to shed some light on the above question. Using logistic regression models on census mortality linked data on 1.65 million men and women aged 18 to 64 years, they have observed that unemployment is strongly associated with suicide among women and men in the age group 25–64. At the same time, no significant associations were observed in other age groups. In an effort to support a causal argument, the authors have controlled for the usual socioeconomic variables (education and income), and to convince readers that there is no confounding (selection bias) Blakely et al¹ have also reported results of various sensitivity analyses using information from other studies.

The analysis was competently done, but the study is not without serious limitations. Firstly, the key independent variable in the report, employment status is a time varying covariate, but it is not treated as such in the analysis. Failure to account for multiple occurrences even in a given calendar year can distort results by underestimating or overestimating the consequences of joblessness. In short, imprecision and inaccuracies are introduced into the analysis, and despite confidence intervals the validity of conclusions become suspect. Previous studies^2,3 suggest that the effect of unemployment on suicide may be more pronounced immediately after job loss. As time progresses the newly unemployed adjust to their novel status and they may be less inclined to commit suicide. Furthermore, with passage of time previously unemployed people may find work and thus vacate the “unemployed” status. Secondly, it is unknown in the analysis when job loss occurred. All that is known is that at some point before census night, cohort members became unemployed, but the timing of unemployment is unknown. There is also no information on whether they had experienced more than one episode of unemployment.

Blakely et al seem to have linked the mortality information to census data in the three years after census night.¹ While this practice may have perhaps reduced the problem regarding the transitional nature of employment status, it did not eliminate it because of the long inter-census period. In the time lag between the current census (the one linked to mortality data) and the prior census, people may have still moved across the three categories of employment status.

In view of the above issues, it is imperative that researchers find ways to accommodate peculiarities associated with time varying covariates in cohort and other longitudinal studies. A significant part of the problem in this and most studies of this type is their dependence on official (government collected) datasets. Censuses are not taken primarily for epidemiological research. In many countries enumeration occurs only once in ten or five years (depending on national mandates). As only one enumeration is done there is no provision for follow up data collection on the same people on a weekly or monthly basis. Even if that were possible, the logistics and accompanying financial costs would be prohibitive. One result of this dependence is that very often information is needed by the researcher but it is unavailable in administrative (government) statistics. At other times (like in the present situation), data are available but not in the format appropriate for the selected research problem.

Ideally, one would have liked to see controls for mental illness and general health status both of which could directly affect unemployment and suicide. For instance, mentally ill persons may be at higher risk of becoming unemployed; they may also be at higher risk of committing suicide.⁴ Blakely et al admit that they lack such data.

Related to the above, another important flaw in epidemiological and sociological research on unemployment and suicide using census data is in the conceptualisation and measurement of employment status itself. In the New Zealand Census Mortality Study, the employed are those already at work. The unemployed are persons that are actively seeking work and available for work.¹ Everyone else is placed into a residual category called the “non-active”. The primary limitation in this conceptualisation of employment status is that it fails to take into account people who are jobless, but have become discouraged in the labour market and have given up looking for work. Some US sociologists euphemistically refer to this group as “discouraged workers”.⁵ Their number is never known, but in periods of severe and sustained economic downturns, it is never negligible especially among racial/ethnic minorities and other marginalised groups.

In view of the above, it is no surprise that Blakely et al¹ found a highly significant association between non-active status and suicide in two of their multivariate models (OR=2.63, CI=1.63 to 4.25 for women; OR=2.59. CI=1.89 to 3.55 for men). Although the non-active group includes students, homemakers, the permanently sick, and retired, as it is a residual category of persons not elsewhere classified, it most probably has a large number of persons that had given up looking for work before the census. The odds ratio for the non-active is greater in magnitude than that obtained for the unemployed in both multivariate models. This is noteworthy in view of the fact that the analyses were limited to persons in the age group 25–64 years.