How to adjust for baseline heterogeneity in count data occurring from experimental studies

Abstract

When analysing the results from experimental and observational studies, the main aim is often to estimate what effect the treatment is having on the participant. For this reason, it is important that the estimate for the treatment effect is not influenced (biased) by having imbalanced (heterogeneous) treatment groups. Randomisation is often used in experimental studies to obtain homogenous treatment groups i.e. the distribution of all covariates (except treatment group) is the same between treatment groups. However, when the sample size is small, randomisation may not successfully obtain entirely homogeneous groups. Heterogeneous groups are often an issue in observational studies as randomisation cannot be used. This thesis aims to demonstrate the need to adjust for baseline heterogeneity by showing the potential consequences of not. The thesis also aims to find a method to successfully adjust for baseline heterogeneity. A hypothetical example is drawn up to demonstrate the potential bias in the results if no adjustment for the baseline heterogeneity is made. An explanation is given on how failing to adjust for heterogeneity, could lead to false conclusions to the extent that a harmful drug could be wrongly authorised. This thesis examines the properties of six potential methods for adjusting for baseline heterogeneity which include 5 parametric methods (using an Offset, Continuous Covariate, Categorical Covariate, Random Effect and a Conditional model) and a non-parametric method (Mantel-Haenszel). The ability of these methods to adjust for heterogeneity is assessed by using them to analyse three datasets containing baseline heterogeneity. Furthermore, a detailed simulation study is undertaken to analyse the bias and RMSE of each of the methods. The treatment effects obtained from the different methods (for adjusting for baseline heterogeneity) differ in the analysis of the example datasets. The AIC and BIC demonstrate that adjusting for baseline heterogeneity is required. However, the AIC and BIC cannot convincingly separate the parametric methods (AIC and BIC is not available for the Mantel Haenszel method). In order to distinguish between the 6 different trial methods, simulation studies are used. The RMSE is then calculated for a range of different scenarios (different sample sizes, risk ratios, number of treatment groups and time points). The Continuous method consistently performs well. For this reason, the 13 Continuous method appears to be the preferred method to use.

More information

Identifiers

ORCID for Dankmar Bohning: orcid.org/0000-0003-0638-7106

ORCID for Helen Ogden: orcid.org/0000-0001-7204-9776

Catalogue record

Export record