Electronic Thesis and Dissertation Repository

Thesis Format

Monograph

Degree

Doctor of Philosophy

Program

Epidemiology and Biostatistics

Supervisor

Choi, Yun-hee

2nd Supervisor

Zou, Guangyong

Abstract

Cluster randomization trials are popular in situations where the intervention needs to be implemented at the cluster level, or logistical, financial and/or ethical reason dictates the choice for randomization at the cluster level, or minimization of contamination is needed. It is very common for cluster trials to take measurements before randomization and again at follow-up, resulting in a clustered pretest-posttest design. For continuous outcomes, the cluster-adjusted analysis of covariance approach can be used to adjust for accidental bias and improve efficiency. However, a direct application of this method is nonsensical if the measures are incompatible with an interval scale, yet such data are very common in practice.

In this thesis, we propose nonparametric methods for trials with a clustered pretest-posttest design, focusing on estimation of treatment effect. We quantify treatment effects using the win probability, defined as the probability that a randomly selected subject in the treatment group has a more favourable outcome than that of a subject in the control group. The methods for data analysis and sample size planning for estimating win probability rely on subject-specific win fractions created from outcome measurements at baseline and follow-up. Specifically, the win fraction for a subject is given by the difference between the rank of the observation among all observations in the combined sample of two treatment groups and its rank among observations in its own group divided by the sample size of the comparison group. The cluster-adjusted analysis of covariance is then applied to win fractions created from baseline and follow-up measurements. The proposed methods, which may be regarded as an extension of Zou (2021) for follow-up measurements, are applicable to studies with binary, ordinal, count, and continuous outcomes without making parametric assumptions. Simulation results demonstrated that the methods for constructing confidence intervals for the win probability performed well in terms of coverage and average interval width, even when the numbers of clusters are as small as 5 clusters per arm. The methods for sample estimation also performed well in terms of the probability of achieving a pre-specified precision.

The methods are illustrated using data from two published cluster randomization trials with SAS code provided.

Summary for Lay Audience

Cluster randomization trials are popular in situations where the intervention needs to be implemented at the cluster level, or logistical, financial and/or ethical reason dictates the choice for randomization at the cluster level, or minimization of contamination is needed. It is very common for cluster trials to take measurements before randomization and again at follow-up, resulting in a clustered pretest-posttest design. For continuous outcomes, the cluster-adjusted analysis of covariance approach can be used to adjust for accidental bias and improve efficiency. However, a direct application of this method is nonsensical if the measures are incompatible with an interval scale, yet such data are very common in practice.

In this thesis, we propose nonparametric methods for trials with a clustered pretest-posttest design, focusing on estimation of treatment effect. We quantify treatment effects using the win probability, defined as the probability that a randomly selected subject in the treatment group has a more favourable outcome than that of a subject in the control group. The methods for data analysis and sample size planning for estimating win probability rely on subject-specific win fractions created from outcome measurements at baseline and follow-up. Specifically, the win fraction for a subject is given by the difference between the rank of the observation among all observations in the combined sample of two treatment groups and its rank among observations in its own group divided by the sample size of the comparison group. The cluster-adjusted analysis of covariance is then applied to win fractions created from baseline and follow-up measurements. The proposed methods, which may be regarded as an extension of Zou (2021) for follow-up measurements, are applicable to studies with binary, ordinal, count, and continuous outcomes without making parametric assumptions. Simulation results demonstrated that the methods for constructing confidence intervals for the win probability performed well in terms of coverage and average interval width, even when the numbers of clusters are as small as 5 clusters per arm. The methods for sample estimation also performed well in terms of the probability of achieving a pre-specified precision.

Creative Commons License

Creative Commons Attribution-Noncommercial 4.0 License
This work is licensed under a Creative Commons Attribution-Noncommercial 4.0 License

Download

Share

COinS