The Use of Multiple Measurements to Monitor Protocol Adherence in Epidemiological Studies

We describe a method for detecting forged data that arise in multiple readings of a single variable when some of the reported values are not true readings but rather copies of one or two genuine values. The method exploits the fact that zero differences between pairs of readings will occur with large frequency in such forged data. The identification of those technicians in whose data zero differences occur often enough to raise suspicions of forgery is accomplished by eye from a frequency plot of zero differences, aided by a simple application of cluster analysis. The method is illustrated using the multiple blood pressure measurements made by seventyeight clinicians in the Lipids Research Prevention Trial. Introduction t In many epidemiological studies multiple measurements of certain variab1es are taken on a subject at one sitting because the mean of several readings provides a measure with less variability than that of a single reading. It sometimes happens, however, that technicians will only make one or two genuine readings. and substitute copies of them in place of the remaining readings of the series. Obviously some procedure must be set up to detect forgeries of this kind. One detection scheme is suggested by the fact that in any series of numbers, some of which have been manufactured by copying, an anomalously large number of values will be identical, and in a frequency plot one value will occur too often. Because the value which appears too frequently will no doubt differ from case to case, it is advantageous to standardize this procedure by examining differences among the reported readings. A frequency plot of such differences will then expose the forgery by showing an abnormally large number of zero differences. If this scheme of examining the frequency of zero differences is adopted, one then must decide what constitutes too large a number of zero differences. It is often the case that field data are not especially uniform, particularly if they are taken by different groups of technicians, using different instruments, working in different places on different populations, and under different conditions. When this is the case it is difficult or impossible to establish any general underlying distribution which can be utilized in determining how large a fraction of zero difference frequencies is to be taken as evidence of forgery. Under these circumstances it is possible that a separation of those zero differences which occur often enough to raise suspicions of forgery from those which occur with acceptably small frequencies can be done by eye from a frequency plot of zero differences~ Simple cluster analysis may provide some