Content area
Background: Epidemiological studies of mental health among sexual minority men (SMM) show considerable heterogeneity in prevalence estimates. While population-based surveys with probability sampling are considered the gold standard for prevalence estimates, misclassification biases can still occur due to reluctance to accurately share sexual orientation in government surveys. Community-based surveys can reduce misclassification for hard-to-reach populations such as sexual minorities but are prone to selection/participation bias due to nonprobability sampling. The purpose of this thesis is to develop and apply different methods for correcting for selection bias in nonprobability samples and misclassification bias in probability samples, applied to SMM mental health.
Study Objectives: 1) estimate the prevalence of mental health and social connectedness among SMM using the adjusted logistic propensity (ALP) method that corrects for selection bias due to nonprobability sampling, 2) evaluate the performance and statistical properties of a newly proposed two-step method that can simultaneously correct for selection bias in a nonprobability sample and misclassification bias in a probability sample with a simulation study, 3) apply the two-step method to estimate prevalence of social connectedness among SMM and improve the prevalence estimation of some outcomes obtained in Objective 1.
Data Sources: Canadian Community Health Survey (CCHS) 2015-2018 and the community-based Sex Now 2019 survey
Results: The ALP resulted in prevalence estimates that fell between the Sex Now and CCHS estimates, reducing heterogeneity in between-survey estimates. The simulation showed that the two-step method produced estimates with the smallest relative bias and the best coverage probability compared to other methods of prevalence estimation, including ALP alone. Applying the two-step method resulted in minimal changes in prevalence estimates compared to the ALP-weighted estimates.
Conclusion: The two-step method is the most effective tool in reducing biases in the analytic stage for hard-to-reach populations. Our findings suggest that the biases in our SMM data were minimal, providing more confidence in the robustness of previous analyses with these data. Better recruitment and data collection strategies are still the optimal approach to reducing biases. However, in the presence of selection or misclassification bias, the analytical methods proposed in this thesis provide improved approaches for prevalence estimation.