The impact of measurement differences on cross-country depression prevalence estimates: A latent transition analysis

Background: Depression is currently the second largest contributor to non-fatal disease burden globally. For that reason, economic evaluations are increasingly being conducted using data from depression prevalence estimates to analyze return on investments for services that target mental health. Psychiatric epidemiology studies have reported large cross-national differences in the prevalence of depression. These differences may impact the cost-effectiveness assessments of mental health interventions, thereby affecting decisions regarding government and multi-lateral investment in mental health services. Some portion of the differences in prevalence estimates across countries may be due to true discrepancies in depression prevalence, resulting from differential levels of risk in environmental and demographic factors. However, some portion of those differences may reflect non-invariance in the way standard tools measure depression across countries. This paper attempts to discern the extent to which measurement differences are responsible for reported differences in the prevalence of depression across countries. Methods and findings: This analysis uses data from the World Mental Health Surveys, a coordinated series of psychiatric epidemiology studies in 27 countries using multistage household probability samples to assess prevalence and correlates of mental disorders. Data in the current study include responses to the depression module of the World Mental Health Composite International Diagnostic Interview (CIDI) in four countries: Two high-income, western countries—the United States (n = 20, 015) and New Zealand (n = 12,992)—an upper-middle income sub-Saharan African country, South Africa (n = 4,351), and a lower-middle income sub-Saharan African country, Nigeria (n = 6,752). Latent class analysis, a type of finite mixture modeling, was used to categorize respondents into underlying categories based on the variation in their responses to questions in each of three sequential parts of the CIDI depression module: 1) The initial screening items, 2) Additional duration and severity exclusion criteria, and 3) The core symptom questions. After each of these parts, exclusion criteria expel respondents from the remainder of the diagnostic interview, rendering a diagnosis of “not depressed”. Latent class models were fit to each of the three parts in each of the four countries, and model fit was assessed using overall chi-square values and Pearson standardized residuals. Latent transition analysis was then applied in order to model participants’ progression through the CIDI depression module. Proportion of individuals falling into each latent class and probabilities of transitioning into subsequent classes were used to estimate the percentage in each country that ultimately fell into the more symptomatic class, i.e. classified as “depressed”. This latent variable design allows for a non-zero probability that individuals were incorrectly excluded from or retained in the diagnostic interview at any of the three exclusion points and therefore incorrectly diagnosed. Prevalence estimates based on the latent transition model reversed the order of depression prevalence across countries. Based on the latent transition model in this analysis, Nigeria has the highest prevalence (21.6%), followed by New Zealand (17.4%), then South Africa (15.0%), and finally the US (12.5%). That is compared to the estimates in the World Mental Health Surveys that do not allow for measurement differences, in which Nigeria had by far the lowest prevalence (3.1%), followed by South Africa (9.8%), then the United States (13.5%) and finally New Zealand (17.8%). Individuals endorsing the screening questions in Nigeria and South Africa were more likely to endorse more severe depression symptomology later in the module (i.e. they had higher transition probabilities), suggesting that individuals in the two Western countries may be more likely to endorse screening questions even when they don’t have as severe symptoms. These differences narrow the range of depression prevalence between countries 14 percentage points in the original estimates to 6 percentage points in the estimate taking account of measurement differences. Conclusions: These data suggest fewer differences in cross-national prevalence of depression than previous estimates. Given that prevalence data are used to support key decisions regarding resource-allocation for mental health services, more critical attention should be paid to differences in the functioning of measurement across contexts and the impact these differences have on prevalence estimates. Future research should include qualitative methods as well as external measures of disease severity, such as impairment, to assess how the latent classes predict these external variables, to better understand the way that standard tools estimate depression prevalence across contexts. Adjustments could then be made to prevalence estimates used in cost-effectiveness analyses.