Confounding

Published on March 13, 2014 by

Dealing with multiple confounders in observational studies: beyond logistic regression

While traditional regression methods have been proven as a powerful research tool, no technique is suitable for all circumstances. Two common situations in which regression techniques alone are likely to produce biased results is when the outcome is rare and the number of measured confounders is large; and/or when important confounders are neglected.
Analysing briefly an example in which both circumstances are present simultaneously, this article shows how propensity score matching associated with Monte Carlo sensitivity analysis can be considered an interesting complement to traditional multivariate modelling.

Published on September 14, 2011 by

Use of fractional polynomials in medical research

Most multivariable models in clinical and epidemiology research consider predictor variables as linear terms or as dummy variables after categorization of continuous variables. Clinically it may be desirable to classify patients into different prognosis groups, but categorization of continuous variables assumes homogeneity of the trait under consideration within each specified category. This may however be unrealistic especially when few categories are used. Categorization may result in overparameterized models and there is usually loss of efficiency. Important relevant predictor variables are sometimes missed in prognostic or diagnostic models because the true functional form of a predictor variable may be non-linear. Categorizing confounding variables may result in residual confounding. Fractional polynomials have been proposed in epidemiological studies to investigate functional forms of continuous predictor variables and confounders.