What do statistics really tell us about the quality of the data from self-monitoring of blood glucose?

Recently a number of studies have examined the quality of the data obtained from various systems used in the self-monitoring of blood glucose. Many of these studies have used parametric statistical techniques such as the Pearson product-moment correlation (r) and linear regression to evaluate the errors associated with self-monitoring results. These statistical methods, while well known and easily computed on modern computers, are often inappropriate for evaluating either the amount of error associated with self-monitoring or the clinical significance of these errors. For example: 1. a correlation of 1.00 does not necessarily mean that the measurements from a self-monitoring system agree with the true values and are without error; 2. a correlation close to 0.00 does not necessarily mean that the measurements from self-monitoring differ widely from the true values and possess large amounts of error; 3. a slope of 1.0 and a y-intercept of 0.0 in a linear regression equation do not necessarily mean that the self-monitoring measurements agree with the true values; 4. a slope and y-intercept that deviate significantly from 1.0 and 0.0, respectively, do not necessarily mean that such measurements differ widely from the true values. The present paper illustrates some of the limitations and common misconceptions concerning these statistics, and shows that a reliance on these techniques alone can, in certain circumstances, lead to misleading estimates of the amount of error associated with self-monitoring systems and inappropriate descriptions of the clinical significance of these errors.(ABSTRACT TRUNCATED AT 250 WORDS)