Inference from Imprecisely Measured Exposures: Caution Is Needed When Using Health Insurance Data for Clinical Research: Commentary on an article by Karch M. Smith, BA, et al.: "The Effect of Sex Hormone Deficiency on the Incidence of Rotator Cuff Repair. Analysis of a Large Insurance Database".

Abstract

Commentary Rotator cuff injuries are common and can lead to disability and pain1,2. Smith et al. utilized health-care billing data to explore whether sex hormone deficiency (SHD) was associated with rotator cuff repair (RCR). Their work builds on several studies that have assessed the relationship between SHD and postoperative recovery. Understanding the possible link between SHD and rotator cuff disease could point to strategies for preventing the need for RCRs and improving outcomes following treatment. As the availability and use of billing data for research increase, the importance of understanding their strengths and limitations cannot be overemphasized. The longitudinal, large-scale, and comprehensive nature of health-care claims databases make the data particularly appealing for answering questions such as the one posed by Smith et al. The MarketScan database includes all services paid by a variety of health insurers across multiple care settings and over time. Billing data also include clinical information relevant to payment decisions, including diagnoses and procedure codes; however, it is important to bear in mind that billing data are requests for payment of services rather than a record of patient health. Many of the challenges associated with the use of administrative data arise because of their origin as payment claims. A diagnosis is recorded on a bill when it explains the need for the service or how its presence complicates the service delivery. Therefore, it is much easier to measure services in administrative billing data than diagnoses. In the case of tobacco use, a diagnosis code reporting a history of tobacco use may be recorded to explain a bill for services such as tobacco cessation treatment or counseling. Not all people who use tobacco will have claims-based evidence of tobacco use, only those whose tobacco use is important for payment. Administrative data do not detect the absence of a condition, and it is not possible to add a claims-based code for “never used tobacco,” which results in claims-based measures of tobacco use having high specificity but low sensitivity3. When interpreting an SHD diagnosis in the claims, we must remember that it will typically have been recorded for use explaining or informing the use of services. The challenge, then, is interpreting analyses performed with use of these imprecisely measured exposures. If the exposure is mismeasured in both the RCR and non-RCR groups, the bias is in favor of finding no association. We are most concerned if the measurement of the condition may be related to the measurement of the outcome. If surgeons use the opportunity of rotator cuff disease or recurrence to counsel their patients on tobacco use, a tobacco use diagnosis may appear at a higher rate for those receiving RCR. This same challenge may be seen with SHD if some part of the RCR work-up led to the diagnosis and treatment of SHD or if the diagnosis and work-up of SHD led to the recognition of a need for RCR. In this case, the association is identified but the underlying mechanism would be misinterpreted. It is encouraging to note, as the authors do, that the measured rate is similar to a non-claims-based estimate, but the differential between the RCR and non-RCR groups was previously not known. As Smith et al. note, there are limitations to measuring health conditions in administrative data. Spending time thinking about the proposed relationship and the factors associated with observing the exposure and outcome of interest in administrative billing data can be helpful. The strengths of administrative data—e.g., the number of patients, availability of controls, and data provided from a variety of health-care settings—seem to support the validity of the reported association between SHD and RCR. Further prospective studies are needed to determine if the relationship remains with direct measurement. If this association is confirmed, we will need to determine how to use this information in clinical settings for the prevention and treatment of rotator cuff injury.