Ignoring heterogeneity in health may bias measurement of intervention outcomes through confounding with intervention of interest. If repeated observations on each subject are available, heterogeneity may be usefully included in outcomes studies. We assume heterogeneous health status as a latent index and multiple health proxies (and their correlations) are used to estimate heterogeneous health grouping from the latent index. For example, in a treatment effect study with longitudinal data: 1) estimate K, the number of heterogeneous groups, by latent class cluster analysis (LCCA) using health proxies of each subject at each period, such as comorbidity indices, length of hospitalization, total health care cost and so on; 2) if K >1 (heterogeneity), estimate a treatment effect for each group and compare the results across the groups; 3) if the effects vary over the groups, heterogeneity can be translated by each group's health profile (e.g. higher effectiveness found in sick but less hospitalized group). This approach is relatively conservative and combines multiple proxies objectively. Estimating K implies a near consensus of model selection criteria such as Bayesian Information Criteria (BIC), adjusted BIC, Akaike Information Criteria (AIC), and consistent AIC; and bootstrap likelihood ratio test (BLRT). Furthermore, it is difficult to find a practically useful K (say <5) because K tends to diverge to N (i.e. each subject is a group), for a large enough sample size N. Applying heterogeneity estimation to a claims data of 3260 subjects for two years found two heterogeneous groups (BIC, adjusted BIC, consistent AIC, and BLRT all supported K=2 except AIC). One group (N=2841) was significantly sicker than the other group (N=419) in Year 1 (and in Year 2) at 5%: Charlson Comorbidity Index 3.91 vs 0.11 (4.49 vs 0.14); length of stay 0.87 vs 0.03 (1.04 vs 0); total cost $10690 vs $245 ($11149 vs $184).