Is it possible to improve the analytical approach to the evaluation of cluster-randomised trials where the complexity of the intervention demands a small number of clusters? the case of the triage plus Integrated TB-HIV community intervention project in Lilongwe Rural, Malawi

摘要

Introduction In this thesis, analytical approaches for the design and evaluation of cluster randomised trials are presented and reviewed. In particular, statistical power/ sample size issues relating to the design of cluster randomised trials for which only a limited number of clusters are available are assessed using a series of simulation studies. The use of computer simulation methods made it possible to investigate how well cluster randomised trials with limited numbers of clusters available can be optimised both in terms of statistical power and also the accuracy of parameter estimates. The study design conditions performing best in the simulation studies were then applied to a community intervention study involving informal healthcare providers: the 'Triage Plus integrated tuberculosis (TB) and human immuno-deficiency virus (HIV) community intervention project in Lilongwe rural, Malawi'. Aims and objectives The general aims of this dissertation were to: 1. investigate if it is possible to improve the analytical approach to the evaluation of cluster-randomised trials where the complexity of the intervention demands a small number of clusters and in which the primary outcome measure is a count of events occurring in a specified time interval; 2. investigate the effectiveness of engaging informal healthcare providers in integrated TB and HIV community intervention in treatment initiation rates and testing access rates, a cluster randomised trial was conducted in Malawi for which only a limited number of clusters were available to the researchers. The specific objectives were: 1. to review cluster randomised trials and the statistical methods used in the assessment of the effectiveness of the intervention in this type of trial when the primary outcome measure is a count of events occurring in a specified time interval; 2. to assess the statistical efficiencies of different design conditions in terms of statistical power and the accuracy of parameter estimates when determining the effectiveness of complex interventions with a limited number of clusters in this situation; iii 3. to identify the circumstances under which each of the statistical methods would be most robust in detecting significant intervention effects or providing accurate estimates of intervention effects; 4. to apply these statistical approaches to the data collected in the cluster randomised clinical trial of community based interventions for TB and HIV (the 'Triage Plus' study); 5. to assess the effect of involving non-paid informal healthcare providers in integrated TB and HIV community interventions aimed at improving testing and treatment initiation rates for these two diseases. Methods Two research approaches were used in this dissertation: 1. Simulation studies were used to investigate statistical efficiencies in terms of statistical power and accuracy in parameter estimation under different study design conditions for cluster randomised trials in which the primary outcome measure is a count of the number of events occurring during a specified period of time. 2. These statistical approaches were then applied to obtain robust estimates of the effect of the test intervention using the data collected during the “Triage Plus” study. The Triage Plus intervention, implemented in rural areas of Lilongwe, involved informal healthcare providers in an integrated TB and HIV community intervention. This intervention specifically involved empowering the informal healthcare providers in disease recognition, sputum specimen collection, referral of presumptive TB cases, and conducting community TB and HIV awareness meetings. Results The simulation studies showed that statistical efficiency and power both varied considerably under the different design conditions investigated. Non-coverage rates within the nominal value of 5% and negligible biases in the estimated fixed effects parameters (regression coefficients) were observed for all scenarios investigated including the (minimal) 3 cluster per arm design. However, it was discovered that, in order to achieve adequate power in low incidence disease conditions such as TB treatment initiation rates, more repeated measurement times were required to achieve adequate power of 80% with a true effect size of 20% or lower (for example, 12 measurement times were needed to achieve adequate power in this situation in a 3 cluster per arm design when the ICC was 0.00154). With an ICC of 0.081 iv at least 9 clusters were needed to achieve adequate statistical power of ≥80% with an effect size of 20% with 6 and 12 measurement time points respectively for high and low incidence disease conditions. For an effect size of 40%, at least 3 clusters per arm were needed to achieve adequate power with 4 repeated measurement times in low incidence diseases and 3 measurement times for high incidence diseases. For ICCs of 0.321 and above, no adequate statistical power was achieved with an effect size of ≤40% in both high and low disease incidence conditions. In the analysis of the TB services access data from the Triage plus study, the intervention significantly increased the number of presumptive TB cases accessing testing sites by 15.2% (p=0.003) in the first 12 months of the intervention; however, this was followed by a statistically non-significant reduction of 18.3% (p=0.224) when the intervention was rolled-out into the control clusters. Overall, the intervention was associated with a non-significant increase in TB treatment initiation rates of 18% (p=0.112). In the analysis of HIV services access rates, antiretroviral therapy (ART) initiation rates increased significantly by 34.7% (p=0.048) in the intervention clusters in the first 12 months of intervention, and the ART initiation rates were similar after rolling-out the intervention to the control clusters. Overall, the intervention was associated with a 61% increase in HIV testing uptake rates (p