Introduction

Before multilevel analysis was developed, the problem of correlated observations within, for instance, medical doctors was tackled in two ways: either ignoring the fact that the observations are correlated or combining the correlated observations into one value. In fact, both methods are still frequently used. Ignoring the fact that the observations are correlated assumes that all observations are independent. In Chapter 2, this method was called ‘naive’ analysis, the advantage of which is that ‘standard’ regression analysis can be used. This way of analysing clustered data is also referred to as the ‘disaggregation’ method. The other possibility is not to ignore the dependency of the observations, but to analyse the group observations (i.e. made by each medical doctor) instead of the individual observations. Therefore, some sort of average value of the observations for each group must first be calculated and then these averages can be used as outcome in a ‘standard’ regression analysis. This method is referred to as the ‘aggregation’ method. To answer the question: ‘What do we gain by using multilevel analysis?’, it is interesting to compare the results obtained from these three types of analysis: the ‘naive/disaggregation’ method, the ‘aggregation’ method, and the (more sophisticated) multilevel analysis.

Example with a balanced dataset

In this example we use a dataset from a randomised controlled trial. The outcome variable in this experimental study is a certain continuous health outcome. The total study population consists of 200 patients, randomly divided into an intervention group and a control group.