Should Deidentified Case Data Be Treated as Independent Data Points?

We read with interest Whitlock et al.’s¹  review of perioperative mortality derived from the National Anesthesia Clinical Outcomes Registry database. One shortcoming of the National Anesthesia Clinical Outcomes Registry database they briefly allude to is “the inability to eliminate patients who had multiple anesthetics” from the analysis. We suggest that this can be a major confounder in the retrospective analysis of such databases. For example, analysis of case data for the past 3 yr from our tertiary pediatric center showed a 48-h mortality of 85 per 70,194 cases (0.12%) or 64 per 42,808 patients (0.15%). Fifty patients had one procedure during this 48-h period, nine patients had two procedures, four patients had three procedures, and one patient underwent five procedures. The share of patients who had multiple procedures was even higher for 30-day mortality. This is not unexpected , as the sickest, most-likely-to-die patients are likely to have multiple procedures before their death.

Large databases allow collection of a sufficient number of relatively rare events (e.g., perioperative death) to identify statistically meaningful associations between outcomes of interest (e.g., perioperative mortality) and risk factors. Despite the large numbers of cases included in such analyses, e.g., nearly 3,000,000 in the study by Whitlock et al.,¹  or approximately 244,400 in the study by Mathis et al.,²  the actual numbers of index cases are still rather small, e.g., 944 and 232, respectively. Therefore, counting multiple procedures in the same patient as independent data points may introduce significant bias toward attributes found in those patients. This problem is not limited to mortality alone, but may apply to any infrequent serious outcome.

A potential remedy would be for national and institutional databases to assign a unique number to each patient with the key held by the institutional administrator. Patient data would have to be linked to this unique number before being submitted in a deidentified manner to the national database. Alternatively, the incidence of “double counting” should be determined first with identifiable data at the institutional level, and the potential effects of such bias should be discussed in each publication that relies on large deidentified databases.

Support was provided solely from institutional and/or departmental sources.

The authors declare no competing interests.