The Effect of Random Error on Diagnostic Accuracy Illustrated with the Anthropometric Diagnosis of Malnutrition

Emmanuel Grellety, Michael H. Golden
  • PLoS ONE, December 2016, Public Library of Science (PLoS)
  • DOI: 10.1371/journal.pone.0168585

Why is it important?

For many illnesses biochemical, anthropometric or other measurements are made and compared with the distribution of the parameter derived from a normal healthy population. To assess the prevalence of the illness an epidemiological survey is usually conducted and the proportion of the survey population that falls outside the “healthy” range is defined as having the condition. Epidemiologists, medical experts & researcher often consider that random measurement error has a relatively minor effect upon the results and increased accuracy of the estimate is mainly obtained by increasing the sample size. In theory errors of measurement should always increase the spread of a distribution increasing the relative size of the tails of the distribution. Thus, potentially inflating the reported prevalence of the condition being studied. Malnutrition is of major public health importance with between 25% and 50% of all childhood deaths having malnutrition as the underlying factor which determines whether a child will die or recover from associated illness. The profile of malnutrition has increased considerably since the two Lancet series and the use of anthropometric malnutrition as part of the Millennium Development Goals. Yet the surveys to ascertain the prevalence of malnutrition have very rarely rigorously assessed the quality of the data despite the fact that they are undertaken in difficult circumstances and are prone to many errors of measurement made by field staff. In order to examine the effect of measurement error we have conducted a Monte Carlo simulation of anthropometric surveys and imposed random errors of measurement on the data. The results show that there is an increase in the standard deviation with each of the errors, that the spread becomes exponentially greater with the magnitude of the sort of error that occur in real life situations and that the effect of an increase in SD that appears to be fairly trivial has a major effect upon the reported prevalence of the condition. This effect is general phenomenon which will apply to all illnesses defined with a cut-off point and should thus be of general interest to all those involved in public health, in the definition of disease and in clinical diagnosis. To our knowledge, this is the first time that a study determines the magnitude of the effect of measurement error on the quality of the data and the effect on the reported prevalence of a major illness. The following analyses show that the effect of measurement error is not trivial and can inflate the prevalence of a diagnostic parameter to an unacceptable degree and that this is not ameliorated by increasing sample size. Indeed, increasing the sample size and complexity of a survey in order to increase the accuracy (decrease the confidence intervals) of a survey may have the opposite effect entirely if the measurement errors increase. We conclude that an estimate of the measurement errors should always be made with any survey and the effect of these on the reported result evaluated. The present study provide greater clarity on how the estimates of the burden of diseases and the definition of normal ranges of biological variables may be misleading.

Read Publication

The following have contributed to this page: Emmanuel Grellety Bosviel