Checking assumptions: Advancing the analysis of sex and gender in health sciences

Background. Sex and gender are dissociable constructs, each including multiple components. Based on the analytic problems associated with dichotomising continuous variables, we aimed to synthesize a new approach to collecting and analysing sex and gender data in health research, in contrast to the conventional use of dichotomous tickboxes to code sex/gender. Methods. Using a literature review and data simulations, we examined the magnitude of the statistical and methodological problems associated with the use of a single dichotomised sex/gender variable, including construct validity, predictive validity, measurement error, residual confounding, misclassification and bias due to cut points, power, and representative sampling. Results. Using the dichotomous sex/gender predictor rather than a continuous sex/gender predictor increased residual confounding up to 80% and misclassification of individual participants up to 50%. Further, there was substantial bias in model parameters when continuous sex/gender variables were dichotomised. Finally, we found that using the dichotomous sex/gender predictor decreased statistical power, in some cases by more than 50%. Conclusions. Using a dichotomous sex/gender predictor in place of continuous sex/gender predictors, when applicable, has profound impacts on the modelling and the validity of statistical inferences. Accordingly, we proposed measurement and analytic strategies for new multi-variable data collection and analyses of existing binarized data in relation to sex and gender, to reduce these statistical problems and improve model quality.