Squeezing observational data for better causal inference: Methods and examples for prevention research

Randomised controlled trials (RCTs) are typically viewed as the gold standard for causal inference. This is because effects of interest can be identified with the fewest assumptions, especially imbalance in background characteristics. Yet because conducting RCTs are expensive, time consuming and sometimes unethical, observational studies are frequently used to study causal associations. In these studies, imbalance, or confounding, is usually controlled with multiple regression, which entails strong assumptions. The purpose of this manuscript is to describe strengths and weaknesses of several methods to control for confounding in observational studies, and to demonstrate their use in cross-sectional dataset that use patient registration data from the Juan Pablo II Primary Care Clinic in La Pintana-Chile. The dataset contains responses from 5855 families who provided complete information on family socio-demographics, family functioning and health problems among their family members. We employ regression adjustment, stratification, restriction, matching, propensity score matching, standardisation and inverse probability weighting to illustrate the approaches to better causal inference in non-experimental data and compare results. By applying study design and data analysis techniques that control for confounding in different ways than regression adjustment, researchers may strengthen the scientific relevance of observational studies.