LAUSR

Forty years ago, Weinstein and Stason published a seminal paper on the foundations for modeling cost-effectiveness in health care. This paved the way for how governments and health care organizations thought about setting priorities with scarce resources. As the growth of new technologies placed greater strain on health care budgets, some countries, such as the United Kingdom, New Zealand and Australia, invested in programs to use operations research methods to gain efficiencies and model the value of these new technologies. Fast forward forty years, Capan and colleagues have provided an excellent introduction to operations research in health care. It is ironic that operations research is new to many health care professionals, despite its deep and rich tradition in medicine. In a world where the average time to adopt and implement a new technology is 17 years, the adoption of operations research in medicine is clearly below average. The immediate questions are why has the adoption been so slow and is there more we can do to support it? One possible answer has been the lack of data. While the electronic medical record has plugged this void in the US, big data is not synonymous with complete data, and some types of missing data are particularly thorny for decision models. Consider a health care organization wanting to optimize care for patients with lower back pain. Every time the patient receives care for back pain, new clinical and administrative records are generated with details on each encounter, whether it is physical therapy, a new opioid prescription, or surgery. But there is no information when the patient does not receive care. Researchers have developed surrogate endpoints, such as a gap in care, to indicate that the problem has resolved. Unfortunately, we do not know why the gap happened, and it could be because the pain improved or it could be because the patient changed providers. Optimizing programs with imperfect data is likely to lead to errant policies or inefficiencies. And while operations researchers can suggest that more data should be collected, health care organizations are weakly motivated to invest in new data collection systems. They are already bogged down with the huge amount of administrative and clinical data that they currently generate. Another problem that emanates from relying on observational data is the endogeneity of the datagenerating process. Consider a health care organization wanting to develop a cost-effectiveness model for treating patients with depression. It could develop a Markov model with disease states, transition probabilities, and outcomes. The strongest evidence for transition probabilities comes from multisite, double-blinded clinical trials, but such trials are limited and few in number. Turning to the administrative or clinical data for these parameters may seem like a simple solution, but transition probabilities estimated from observational data are likely to be biased. We do not observe why a patient changed health states or why they received a treatment, and this creates a bias in the data generating process that is hard to eliminate. Luckily, data limitations do not plague all optimization models in health care. For example, Received 18 April 2017 from Health Economics Resource Center, VA Palo Alto, Menlo Park, CA (THW); Department of Surgery, Stanford University, Stanford, CA (THW); and Department of Surgery, Stanford University, Stanford, CA (JKJ). Financial support was provided by the Department of Veterans Affairs and the Department of Surgery, Stanford University. THW is employed by the Department of Veterans Affairs. THW and JKJ are employed by Stanford University, Department of Surgery. Neither author has a conflict of interest. Revision accepted for publication 1 May 2017.