In cohort studies, to avoid selection bias, a fundamentally important criterion is that subjects recruited to the unexposed group are chosen from the same population as those chosen with an exposure. This criterion presupposes that, at least theoretically, the distribution of covariates, on average, is similar between those exposed and unexposed to a treatment or intervention. An analogous argument extends to case-control studies, which often use historical controls for comparisons. Historical controls are identified from a group of subjects observed in the past close enough in time with those exposed. The implicit assumption is that bias because of temporal changes in population health or clinical practice between the groups can be avoided or minimized.

In its guidance to industry, clinical investigators, and the Food and Drug Administration, the US Department of Health and Human Services defines historical controls as “comparable patients or populations” with adequately documented natural history of the disease or condition who, in the past, had received no treatment or had undergone a standard management of either therapeutic, diagnostic or prophylactic. However, the definition of “comparable patients or populations” was not provided and has remained murky; as a result, the ideal control group for comparison observational studies remains poorly characterized.

Comparability or similarity between the control and study groups is extremely important to ensure valid results. However, perfect or almost perfect similarity in the baseline characteristics and medical management can be expected only in randomzied controlled trials. In observational studies, which represent real-life scenarios, it may be impossible to achieve an ideal or perfect control group. In such instances, any differences in the baseline characteristics or medical management between the 2 groups are usually handled with the use of appropriate statistical analytic techniques (ie, adjustment for confounders). However, such statistical analyses can take into consideration only those confounders that are collected and reported. Unfortunately, some of the important confounding variables may be related to differences in the medical management that are impossible to uncover unless the circumstances of health care are known for the 2 comparison groups.

It is unusual for investigators to report complete data about the circumstances of health care for the groups studied. Important circumstances of health care include the geographic setting, health care setting, type of health care providers, and likelihood in having confounding introduced by differences in the medical management between the 2 groups. The circumstances of care should be taken into consideration in determining comparability (or degree of similarity) between the groups being compared.

We hypothesized that some of the inherent selection biases of observational studies can be minimized if certain criteria (regarding the circumstances of health care) are used to determine comparability regarding the already provided medical care of the comparison groups. However, there are multiple types of circumstances of care along with different degrees of similarity or dissimilarity within each circumstance of care, thus leading to different degrees of comparability that are difficult to untangle.

The use of a comparability scoring system can provide a reasonable assessment of group similarities but, perhaps more importantly, can provide a discrete variable that can be used during data analysis. The comparability score can be incorporated in logistic regression models to adjust for differences in the circumstances of care.

The aim of this commentary is to propose certain comparability criteria, as they relate to circumstances of care, through the development of a scoring system that can be used to determine the degree of comparability or similarity between the study and control groups ( Table 1 ).