Generalizability theory, which subsumes classical measurement theory as a special case, provides a general model for estimating the reliability of observational rating data by estimating the variance components of the measurement design. Research data from the "Heart Smart" health intervention program were analyzed as a heuristic tool. Systolic blood pressure readings for 17 children were taken by two nurses in one of six pairs. Differentiating the sources of error variance was carried out in what is called a "G" study, which used the GENOVA program to estimate the variance of measurement facets. As expected, the variance of individuals' blood pressure was large relative to pairs of nurse observers, sets of measurement, and occasions of measurement within the sets. A generalizability coefficient is the expected squared correlation between the actual, observed scores and the full universe of scores defined by the researchers by specifying measurement facets or protocols. It allows the calculation of different coefficients, given interest in making either norm-referenced or criterion-referenced decisions. While the "G" study evaluated data quality for different decisions, a "D" study (i.e., a study estimating how many conditions are needed to obtain a certain level of generalizability) was also conducted to pinpoint sources of measurement error, so that observations could be increased on a particular facet to reduce error. Many levels for each protocol facet are needed to estimate this properly. (LPG)