Observational studies, matching and propensity scores: Applied to colorectal cancer data

Abstract

Many studies in biostatistics are concerned with comparing two groups of patients, a treatment and a control group, to determine the effect of the treatment. Observational studies typically make use of available data, such as from a cancer registry, and consequently lack explicit design and randomisation which makes analysis and the drawing of definitive conclusions difficult. One potential problem is that there are often large imbalances in the distributions of covariates across the treatment groups. In a randomised control trial, this imbalance in the distribution of observed covariates would usually be accounted for in the design of a study and randomisation would balance over the observed and unobserved covariates. However, when the data comes from an observational study, this imbalance affects subsequent analysis and usual statistical methods are often not enough to correct for this imbalance, leading to biased estimates of treatment effects.

First, we discuss the use of observational studies and randomised control trials, contrasting and comparing them. Following this, methods to allow observational studies to be analysed in a meaningful and correct manner are presented, including matching and inverse probability weighting. However, when there is an extreme imbalance in the sizes of the treatment and control groups these methods seem to perform unreliably. We consider the use of methods, both existing ones from the literature and newly developed proposals, to overcome this deterioration in performance. We attempt to quantify the point at which the degree of imbalance in the numbers over the two groups causes serious deterioration.

The various methods, existing and newly proposed, are compared using a simulation study to evaluate their performance. Throughout, the ideas are illustrated in a survival analysis setting for a study examining the effect of inflammatory bowel disease on survival in colorectal cancer patients, using data from a cancer registry.