Comparing Baseline Adjustment Techniques in Individual and Community Randomized Trials

Frequently we are interested in comparing a count as primary endpoint in a study where randomisation
only involves entire communities such as schools or villages. We might be interested in
comparing one or several interventions or treatments with a control. As not all communities might
be on the same level of the clinical endpoint to be compared, adjustment for baseline heterogeneity
is usually required. This is a serious problem for community of cluster randomized studies, but
it also occurs in individually randomized studies, in particular, if the trial size is small. Baseline
adjustment can be done in different ways and a comparison of these different approaches to baseline
adjustment is the topic of the thesis.

The project idea was motivated by a community–randomized trial on pre-school children in
Belo Horizonte (Brasil) targeting the prevention of caries occurrence [1] where also the associated
data are available. Four interventions were randomized to four schools, one school received all four
interventions simultaneously, and one schol served as control school. The outcome measure was
the count of tooth surfaces with decay, missingness or filling (DMFS). This is also available as the
number of teeth with decay, missingness, or filling (DMFT). Treatment lasted 2 years and at the
end of the period the DMFS (DMFT) status was measured. However, it is clear that any evaluation
of treatments need to take into account the DMFS (DMFT) value when children entered the study
as any unadjusted analysis will lead to potentially largely biased intervention effect estimates. The
major objective of the project is to compare various methods for adjusting for baseline heterogeneity.
We have motivated the project using a count outcome, but clearly the problem applies to any
positive continuous outcome or endpoints as well.

We look at the following methods:

1. A Mantel-Haenszel approach for a stratified computation of the risk ratio.
2. A Poisson regression with baseline as continuous covariate.
3. Poisson regression with log-baseline as offset.
4. Poisson regression with baseline as categorical covariate.
5. Mixed Poisson regression with baseline as random effect.

Note that in all Poisson models is the log-risk ratio which is the parameter of interest. Note
that these models differ in the way they handle baseline heterogeneity.

The project has three major aims:

1. Do the five approaches provide different or similar estimates of intervention effects?
2. Do the five approaches provide different or similar estimates of uncertainty (tests and confidence
intervals)?
3. Investigate ways in which model selection can be done here!

More details are available from Professor Böhning.

Prerequisites: The project needs interest and motivation to work with the package STATA. All
the required analysis and modelling can easily done within STATA. The point of the project is that
in statistical practice baseline heterogeneity is often ignored and coped with inappropriately.
For making more in-depth comparisons, simulation work has to be included which requires solid
knowledge of the statistical programming language R.