Different Methods to Calculate Effect Estimates in Cross-sectional Studies

Abstract

According to results from the epidemiological literature, it can be expected that the prevalence odds ratio (POR) and the prevalence ratio (PR) differ with increasing disease prevalence. We illustrate different concepts to calculate these effect measures in cross-sectional studies and discuss their advantages and weaknesses, using actual data from the ISAAC Phase III cross-sectional survey in Münster, Germany. We analyzed data on the association between self-reported traffic density and wheeze and asthma by means of the POR, obtained from a logistic regression, and the PR, which was estimated from a log-linear binomial model and from different variants of a Poisson regression. The analysis based on the less frequent disease, i.e. asthma with an overall prevalence of 7.8%, yielded similar results for all estimates. When wheezing with a prevalence of 17.5% was analyzed, the POR produced the highest estimates with the widest confidence intervals. While the point estimates were similar in the log-binomial model and Poisson regression, the latter showed wider confidence intervals. When we calculated the Poisson regression with robust variances, confidence intervals narrowed. Since cross-sectional studies often deal with frequent diseases, we encourage analyzing cross-sectional data based on log-linear binomial models, which is the 'natural method' for estimating prevalence ratios. If algorithms fail to converge, a useful alternative is to define appropriate starting values or, if models still do not converge, to calculate a Poisson regression with robust estimates to control for overestimation of errors in the binomial data.