In field surveys, it has long been accepted that not everyone who is in the 'target sample' will agree to be interviewed or will be available at the time the interviewer calls. It is common to find that only 70 to 90 per cent are actually assessed. Furthermore, those who refuse or are repeatedly not available are known to be more likely to have the mental disorder under investigation. For this reason, the prevalence that is found will often be an underestimate. A putative risk factor may itself increase the chances of a person's not being in a sample in the first place, of dropping out, or of dying during the study. Statistical methods have been available for many years for estimating how much error may have occurred due to refusals and how to correct for this in the conclusions drawn.
The other occasion when non-response is a problem is in longitudinal studies, where a sample is followed over several years. If a disorder with an increased mortality is the topic, such as dementia or schizophrenia, it is recognized that some cases will be lost at follow-up. This means that those who are successfully re-examined are a survival élite and are different in important ways from the original cohort. These distortions could lead to mistaken conclusions if the losses are not allowed for. Various techniques have been developed to handle these difficulties, including Bayesian methods which adjust final estimates on the basis of prior knowledge. (19>
Was this article helpful?