Unit of Analysis

In a statistical model, the unit of analysis is the entity about which inference is being made. For example, in a clinical study, an investigator must decide if inference is to be made with regard to individual patient outcomes or with regard to the physicians treating the patients. If the former, then the unit of analysis is the patient, and the resulting odds ratios or relative risks (or other statistics) would be interpreted as reflecting changes in patient risk or differences in patient characteristics. Similarly, it may be desirable to make inference regarding the treating physicians, each of whom may have treated multiple patients. In this case, statistical methods should be chosen so as to address questions related to the physician.

In most instances, the selection of the unit of analysis is straight-forward. In a cross-sectional survey of patients in the emergency department waiting room, the unit of analysis would be the survey respondent, the patient. In a randomized clinical trial of the effectiveness of a new medication treatment, the unit of analysis would again be the individual patient.

Proper identification of the unit of analysis is critical. Failure to do so may result in biased or invalid results. In a clinical study of 100 patients treated by 5 different doctors, if the unit of analysis is the patient, we end up ignoring the fact that patients treated by one doctor will have certain characteristics in common compared with patients treated by another doctor. That is, a doctor is likely to approach different patients in a roughly similar way. Ignoring this ‘clustering’ by physician, or selecting analytic techniques that do not take this into account will yield incorrect estimates of variance, leading to erroneous confidence intervals or p values.

If we are interested in making inference with regard to the treating physicians, patient outcome measures can be summarized with means or proportions within treating physician. The analyses of these types of data require different statistical tests and have a different interpretation than if the patient was the unit of analysis. Analyses that take into consideration effects at these different levels (e.g., patient and physician) are often referred to as ‘multilevel’ or ‘hierarchical’ models.