This paper discusses the need t o account for the sampling scheme in an analysis of epidemiological population data which were collected by a two-stage cluster sample. An example is presented where for validity reasons one should generally account for the sampling scheme, though a rule-of-thumb is given to estimate the effect of not doing 80. The example concerns one of the centers (Augsburg, F. R. G.) participating in the WHO MONICA Project, which was designed to study the relationship between changes in risk factor levels, 88 measured by several surveys in each center, and changes in cardiovascular incidence rates, 8s measured by a registry system for each center. Variance estimation methods which either account for or ignore the sampling scheme are compared for a part.icular sampling scheme. Easily computable upper bounds on the effect on variances of ignoring the s a m p h g . scheme are presented, both over all possible variables, and for a particular variable.' K e y m u % : Design effects ; Epidemiological population studiee ;' Horvitz-Thompson estimators; Systematic random sampling; Trends in coronary heart disease ; Two-stage cluster sampling. ' +(I -(H/N)z(Vi/V1)) (1 +EFn-l)-I = ( I +kPn-I)-l (l+(iK/N)z (Vi/Vi) (2 (a-d) F')) =(I+EFn-1)-1 (1 +(SSE/SSE) (2 (u-2) F'))
where SSE', SSE are the error sum of squarre8 from ANOVA (with respect to PSU effect) for the centers not chosen with certainty and all centers, respectively.