Statistical design of THRio: a phased implementation clinic-randomized study of a tuberculosis preventive therapy intervention

Abstract

Background: Tuberculosis (TB) is a major public health problem in Rio de Janeiro, where a high proportion of HIV-infected adults are co-infected with latent TB. Health officials in Brazil have recommended that HIV patients be tested for TB infection and given TB prophylaxis (isoniazid) if positive. In practice, although Brazil is a model for provision of antiretroviral therapy to patients with advanced HIV disease, relatively few such patients receive TB testing and prevention services.

Purpose: We initiated a randomized study of a health services intervention to train health personnel in implementation of the recommended routine of TB testing and isoniazid prophylaxis. The primary goal is to reduce incident TB disease in the HIV clinic population.

Methods: The clinic-level intervention will be phased in gradually over the study period until all clinics have received the intervention. The clinics' order of initiation of intervention was randomized and subjected to constraints based on clinic-level covariates. This phased intervention cluster-randomized trial required special attention to power/sample size calculation and randomization procedures, of which we provide the relevant details.

Results: Special design considerations accounted for within-clinic correlation, variation in the clinic size, time-varying ratio of intervention to control clinics and guaranteed post-randomization covariate balance. These were successfully implemented for the estimation of power and execution of the randomization strategy.

Limitations: Although the design features of randomization by clinic and phased implementation of the intervention meet logistic and local needs, they substantially lower the statistical power of the study.

Conclusions: Studies with cluster-randomized order of intervention introduction can provide useful information on intervention effects. Their design and analysis are more complicated than for individually randomized parallel design trials. The methods we describe represent practical approaches to the challenges raised in the course of designing this study.