Alternative data sources and discrepant results in case-control studies of estrogens and endometrial cancer

Abstract

In most case-control studies, little attention has been given to two features that can affect the odds ratio: 1) self-selection of the people who are available for interviews, and 2) disagreements between the data obtained by medical records and by direct conversation. The authors investigated these problems during two separate case-control studies of estrogens and endometrial cancer. When interviews were attempted with all patients, fewer controls than cases were available. In samples of patients from a tumor registry, more controls than cases had died before the interview could be solicited, and fewer control patients could be located. In patients sampled from a diagnostic test registry, more controls than cases refused participation. The interview data about estrogen usage disagreed with previously recorded medical information in 16% (52/324) of patients). Among persons marked as "non-users" in medical record data, more cases than controls were available for interview and thus for redesignation as estrogen users. The bias produces a possible artificial increase in the odds ratio when interview and medical record data are combined. Since patients available for interview were systematically biased in favor of having both estrogen use and endometrial cancer, the odds ratio for this association should be calculated separately for record and for interview data.

PIP: Various retrospective case-control studies were performed in an attempt to elucidate problems with data collection bias and interpretation bias these authors felt were present in earlier studies. The 2 features of case-control studies to which little attention has been paid in previous studies and that can affect the odds ratio are: 1) self-selection of the people available for interviews, and 2) disagreement between the data obtained by medical records and direct conversation. To assess these biases and their effects on the odds ratio, 2 case-control studies of estrogen use and endometrial cancer were carried out. The following findings confirm the odds ratio bias problems in case-control study methodology: 1) when interviews were attempted with all patients, fewer controls than cases were available; 2) in samples of cases from tumor registries, more controls than cases had died before soliciting interviews, and fewer control patients in general could be located; 3) when patients are sampled from a diagnostic data test registry, more controls than cases refused to participate; 4) interview data concerning use of estrogen disagreed with medical records in 16% of patients--and among persons marked as nonusers in medical records, more cases than controls were available for interview and thus for possible redesignation as users. All of these biases produce an artificial increase in the odds ratio when interview data and medical record data are combined; hence, the odds radio for the association between estrogen use and endometrial cancer should be calculated separately for record and for interview data.