The rationale of PS matching is that once the PS is accounted for, selection bias has been eliminated. This presupposes that the variables in the equation are the most important ones that determine group membership. If any crucial variables have been omitted, then the groups may remain unbalanced and the results of the study can be seriously biased.7 However, especially when the study uses a preexisting database, it is limited to those variables that have been collected, which may not be the ones that account for who is in one group and who is in the other. Very often, the database was set up for reasons other than research (eg, billings) and the variables we want are not those we have. For example, in the article cited earlier, Nakajima et al6 compared the prognoses of patients with non-small cell lung cancer whose preoperative diagnostic transbronchial biopsies were successful with those whose diagnoses were not (the latter group actually did better). A PS was derived using the patient’s age, sex, pathologic type and stage of cancer, use of thoracoscopy during surgery, and other clinical variables. This is fine as far as it goes, but the question remains as to whether other factors that did not make it into the database may have affected the outcome (eg, comorbidities or surgeon-related factors). In cohort studies in which group membership is determined by the patient (eg, who does or does not smoke; who opts to use medication or elects not to), psychologic factors such as beliefs about health, compliance with medical advice, and the like are prime determinants of both the behavior and the outcome but they are not often measured and are rarely available in administrative databases.