Covariate adjustment in cluster randomised trials with binar y outcomes focussing on relative risks and risk difference
Lead Research Organisation:
University of Birmingham
Department Name: Institute of Applied Health Research
Abstract
Randomised controlled trials frequently have binary primary outcomes yet methods for the analysis of binary outcomes is under researched. The CONSORT guideline for reporting of results from randomised trials recommends that for binary outcomes researchers report both relative and absolute measures of effect. Furthermore, when reporting relative measures of effect, because odds ratios are often misinterpreted as risk ratios and because risk ratios and odds ratios diverge as the prevalence of the outcome becomes increasingly less common, relative risks can be the preferred measure of quantifying effects on the relative scale.
Whilst relative risks and risk differences are relatively straight forward to calculate without adjustment for any covariates, in practice covariate adjustment is often warranted either because of chance imbalance or to improve precision of the treatment effect. Methods to compute relative risks and risk difference whilst adjusting for covariates are more complex and have increased possibility of model non-convergence. In practice, researchers often fail to report absolute measures of effect and when reporting relative measures of effect do so using odds ratios. This is likely due to a multitude of reasons, but contributing factors include complexity of methods.
Covariate adjustment in cluster randomised trials: There are a number of reasons why covariate adjustment is likely to be important in cluster randomised trials (CRTs). Firstly, in CRTs randomisation is at the level of the cluster, and individuals might be recruited after randomisation, meaning there can be important risks of "selection bias" (known as recruitment or identification biases) [Eldridge 2008]. These manifest as differences in the characteristics of individuals under treatment and control arms - and in extreme cases can render the trial more like an observational study [Bolzern 2018; Easter 2021]. Thus, in CRTs, covariate adjustment can have an important role of protecting against confounding [Leyrat 2014]. Furthermore, as under individual randomisation, even when the risk of selection bias is negligible (e.g., in studies with pre-randomisation recruitment), covariate adjustment might still improve statistical precision [Li 2016; Li 2017].
However, there are many nuances around covariate adjustment in cluster trials. For example, in CRTs, covariates might be measured at the level of the cluster, or individual, or both; and there can thus be important decisions around how the covariates are to be included [Begg 2003]. For example, cluster size is a typical stratification factor, often implemented using an historical measure of cluster size. At the analysis stage, questions can arise around what covariate should be adjusted in the analysis: the categorized historical version; or a more contemporary continuous version. Moreover, there are numerous ways to approach covariate adjustment in CRTs, including direct covariate adjustment using generalised linear mixed models or generalised estimating equations, propensity score approaches and marginal standardisation (also known simply as standardisation or G-computation) [Benkeset 2021; Morris 2022], or cluster-level analysis.
What the studentship will encompass
Objective: To establish a methodological framework for covariate adjustment in cluster randomised trials when estimating relative risks and risk differences
Plans for project:
1. Review of the literature and case study
a. Review of reporting and analytical approaches for covariate adjustment in CRTs for binary outcomes. This project will systematically review and document contemporary approaches to covariate adjustment in CRTs when reporting binary outcomes- considering how covariates are selected for inclusion; whether they are adjusted for at the level of the cluster or individual; as well as analytical approaches. This will inform WPs 1 and 2.
b. Identification of several case studies to illustrate the methodology to imp
Whilst relative risks and risk differences are relatively straight forward to calculate without adjustment for any covariates, in practice covariate adjustment is often warranted either because of chance imbalance or to improve precision of the treatment effect. Methods to compute relative risks and risk difference whilst adjusting for covariates are more complex and have increased possibility of model non-convergence. In practice, researchers often fail to report absolute measures of effect and when reporting relative measures of effect do so using odds ratios. This is likely due to a multitude of reasons, but contributing factors include complexity of methods.
Covariate adjustment in cluster randomised trials: There are a number of reasons why covariate adjustment is likely to be important in cluster randomised trials (CRTs). Firstly, in CRTs randomisation is at the level of the cluster, and individuals might be recruited after randomisation, meaning there can be important risks of "selection bias" (known as recruitment or identification biases) [Eldridge 2008]. These manifest as differences in the characteristics of individuals under treatment and control arms - and in extreme cases can render the trial more like an observational study [Bolzern 2018; Easter 2021]. Thus, in CRTs, covariate adjustment can have an important role of protecting against confounding [Leyrat 2014]. Furthermore, as under individual randomisation, even when the risk of selection bias is negligible (e.g., in studies with pre-randomisation recruitment), covariate adjustment might still improve statistical precision [Li 2016; Li 2017].
However, there are many nuances around covariate adjustment in cluster trials. For example, in CRTs, covariates might be measured at the level of the cluster, or individual, or both; and there can thus be important decisions around how the covariates are to be included [Begg 2003]. For example, cluster size is a typical stratification factor, often implemented using an historical measure of cluster size. At the analysis stage, questions can arise around what covariate should be adjusted in the analysis: the categorized historical version; or a more contemporary continuous version. Moreover, there are numerous ways to approach covariate adjustment in CRTs, including direct covariate adjustment using generalised linear mixed models or generalised estimating equations, propensity score approaches and marginal standardisation (also known simply as standardisation or G-computation) [Benkeset 2021; Morris 2022], or cluster-level analysis.
What the studentship will encompass
Objective: To establish a methodological framework for covariate adjustment in cluster randomised trials when estimating relative risks and risk differences
Plans for project:
1. Review of the literature and case study
a. Review of reporting and analytical approaches for covariate adjustment in CRTs for binary outcomes. This project will systematically review and document contemporary approaches to covariate adjustment in CRTs when reporting binary outcomes- considering how covariates are selected for inclusion; whether they are adjusted for at the level of the cluster or individual; as well as analytical approaches. This will inform WPs 1 and 2.
b. Identification of several case studies to illustrate the methodology to imp
Organisations
Studentship Projects
| Project Reference | Relationship | Related To | Start | End | Student Name |
|---|---|---|---|---|---|
| MR/W006049/1 | 01/10/2022 | 30/09/2028 | |||
| 2893574 | Studentship | MR/W006049/1 | 01/10/2023 | 30/09/2029 | Jack Hall |