Abstract:
When a response having two outcomes is modelled
using a logistic model the responses on each observation
are often considered to be independent of each other. This
assumption may not always be valid and the responses may
be correlated with each other as in the case of clustered data,
which can occur especially in the case of survey data. When
responses are correlated as explained, then the ordinary
logistic regression model is unsuitable as the standard errors
will be biased, and therefore this model should be adjusted for
the cluster effect. In this paper one of the many methods of
adjustment suggested in the literature, which is based on robust
standard error estimation for cluster sampling data is examined.
The objective of this paper is to illustrate the theory and mode of
application of this theory by way of using a survey on Paraquet
poisoned patients in Sri Lanka. Here, the patients are clustered
within hospitals and therefore adjustment of the logistic model
for the clustering effect is discussed. Adjusting for the hospital
(cluster) effect significantly reduced the standard error of
four out of thirty three odds ratios given by the model. That
is four odds ratios that were not significant before adjustment
became significant after adjustment. There was no change in
significance in the other twenty nine odds ratios. On average
there is a reduction of 2.29% in the standard error of the odds
ratios after adjustment for the cluster effect. This indicates that
it is effective and important to adjust for the cluster effect.
