There is question that infant mortality rates (IMRs) must be understood as random variables, iftrends and differences are to be properly interpreted. The paper by Kleinman' proposes a mathematical model for comparing rates to long-term trends, with the suggestion that the methodology would be useful at the state level in sorting out adverse effects from random fluctuations. The model is said to be merely descriptive. It serves the purpose of summarizing the trend over several years to allow assessment of whether recent IMRs represent a departure from previous 'average' experience . . . as a first step in a more detailed analysis of the data IMRs are particularly difficult to handle statistically because the range of values of interest is so narrow; hence large denominators (numbers of live births) are needed to detect differences and changes through statistical procedures. It is certainly useful to explore new techniques of analysis and to demonstrate their statistical characteristics, as Dr. Kleinman has done. Potential applications, however, must also be evaluated from a practical perspective. The author recommends that these analyses be used by state health officials to help decide whether there is an infant mortality problem. Without considering the merits of the mathematical model as such, I have several concerns with regard to this proposed use: * Given the tendency to take nonsignificant findings as proof of no problem, the use of significance tests could, in general, lead to serious misinterpretation; * Although the need for follow-up study is emphasized, any further stratification runs into the problem of small denominators; * This is not the way state health officials study and use IMRs. The paper begins by describing the concern that resulted from a finding that IMRs in 11 states increased between 1981 and 1982,2 and points out that random error and long-term trend were not considered. From 1969 to 1981, is noted, the number of states in which the IMR increased from the previous year ranged from 5 to 21, with a mean of 11.5. The implication is that at least some of these increases were the result of chance variability, without an increase in the true rates of infant death. Not mentioned is the fact that we might equally consider that the IMR did increase in some states where the observed rate showed change or even a decrease. The proposed statistical procedure, rather than using year-to-year comparisons, determines whether an IMR in a given year represents a statistically significant departure from the long-term trend as represented by a log-linear model. For each state, and separately for Whites and non-Whites, a 5 per cent test of significance is used. Further investigation is recommended if there are significant results. The author defends the use of 79 separate 5 per cent level tests, rather than using a multiple comparison technique with overall 5 per cent level, with the statement that it is more important to detect the possibility of an adverse trend in a state than to minimize Type I errors. I But even with the greater leeway