Analysis of longitudinal data in perinatal trials when the length of follow-up is informative

Background: Most commonly used statistical methods assume that the data consist of independent observations. Clustered data occur in many settings, such as longitudinal studies, where outcomes are repeatedly measured over time on each subject. Observations from the same subject are dependent and hence form a cluster. Two commonly used methods of analysis for clustered data are mixed models and generalised estimating equations (GEEs). Additional complexity arises when analysing clustered data where the cluster size is informative; that is, where the cluster size is related to the outcome. Most methods of analysis for clustered data, including mixed models and GEEs, generally assume non informative cluster size and hence may not be suitable when the cluster size is informative. Aim: The aim of this thesis is to compare methods for analysing longitudinal data when the cluster size (length of follow up) is informative. Methods: Both real and simulated data were used to compare methods for analysing clustered data with informative cluster size. A range of methods were considered including: GEEs with independent, autoregressive or exchangeable working correlation structures; cluster weighted GEEs; and mixed models. The real data come from a perinatal trial (the POPPET trial), which investigated the effect of high versus standard protein content human milk fortifier on the growth of 60 preterm infants. This dataset was used to investigate different methods of analysis for estimating the effect of treatment on infant growth when informative cluster size was suspected. As real data cannot be used to show which methods of analysis are performing best in general, a simulation study was conducted to compare methods when the true parameter values were known. The data were simulated based on the POPPET trial. Different treatment effects, sample sizes, and correlations between the cluster size and the outcomes were considered. Results: For the POPPET trial, evidence of informative cluster size was found. Different methods of analysis produced quite different parameter estimates but similar conclusions about the effect of the intervention. The simulation results showed that when cluster size was non informative, all methods performed very well. When cluster size was informative, mixed models and autoregressive GEEs always performed well. However, the independence, exchangeable and cluster weighted GEEs often produced low coverage probabilities and model based standard errors that differed from the standard deviation of the parameter estimates. These methods generally performed better when the trial size was larger and when there was no correlation between individual growth trajectories and cluster size. Conclusions: It is recommended that mixed models or autoregressive GEEs be used to analyse longitudinal data with informative cluster size in general, including the POPPET trial data. Independence, exchangeable and cluster weighted GEEs should only be used when the sample size is large and there is no correlation between individual growth trajectories and cluster size.

Author dc:creator

• Longstaff, Tessa

Advisors dc:contributor.advisor

• Glonek, Garique Francis Vladimir
• Yellend, Lisa

• longitudinal data
• analysis
• perinatal trial

Handle dc:identifier.uri

http://hdl.handle.net/2440/100192

OAI identifier oai:identifier

oai:digital.library.adelaide.edu.au:2440/100192