Every day, we are confronted with new numbers on SARS-CoV-2 infection cases. We eagerly use these numbers to analyze recent developments (“Social distancing appears to be working”) and draw policy conclusions by making cross-country comparisons (“Which country has flattened the curve”) and learning from the international experience (“Fatality rates – deaths per infection cases – are initially higher in countries with more intergenerational interactions”). By collecting and publishing case numbers, the Johns Hopkins Coronavirus Resource Center, like several other initiatives, has turned into an important resource for information on the spread of SARS-CoV-2. What these data fail to document is under which conditions patients are being tested for the virus.
The economics literature provides some insights into how patients are selected into medical testing. The structural assumption is that physicians try to allocate costly tests to those patients with the highest expected returns to testing. The value of a test result is determined by its potential to affect subsequent decisions. Typically, only a positive test result requires further interventions (e.g. medical treatments, quarantine). Therefore, patients with a higher probability of being tested positive have higher expected returns to testing. In this model, a physician first assesses a patient’s probability of being tested positive and then, based on her assessment, decides whether to perform testing. Analogously, many countries have adopted the strategy to selectively test individuals only when they report symptoms associated with COVID-19.
Abaluck et al. (2016) apply the selective testing model to examine heterogeneity in physicians’ decisions to perform diagnostic imaging on US Medicare patients. Their study makes three observations. First, physicians’ testing rates range widely from 1.7% to 8.2% of a physician’s patients. Second, physicians’ testing rates are correlated negatively with their average test yields, defined as the probability of a positive result conditional on testing. While physicians attempt to maximize the number of diagnosed positive cases, they do so with varying success. Physicians with a 1 percentage point higher test rate have a 0.39 percentage point lower test yield. Third, for several comorbidities present in risk scoring systems, patients are less likely to be tested, although their test yield is higher than for patients without those conditions.
The authors discuss multiple sources of underlying heterogeneity to explain these empirical observations: patient populations, testing thresholds, and diagnostic skills. Differences in physicians’ patient populations affect the prevalence of a condition. These differences could explain the observed variation in test rates, but not the variation in test yields. For that, heterogeneity in physicians’ testing thresholds is required. Physicians who apply relatively lower testing thresholds are willing to test patients with a lower expected probability of testing positive. These physicians might test more, but end up with lower average test yields. Lastly, the authors show that physicians may make systematic mistakes in assessing patients prior to testing decisions.
A key takeaway from this study is that selective testing makes it difficult to interpret variation in the number of positively tested SARS-CoV-2 infections. Without further information, differences in case numbers could reflect variation in prevalence just as well as different testing thresholds – driven by costs, capacity constraints, and preferences – or diagnostic skills. These elements might even vary over time: Testing guidelines provided by the Robert Koch Institute change continually and we do not know how test yields are affected.
With this background, a planned study to test non-selectively for SARS-CoV-2 antibodies in a large random sample of the German population is an exciting development. Such a study will be key to understand how the pandemic evolves in Germany and to devise informed, effective policy options. For international comparisons, at a minimum information on testing criteria and test numbers is required to base meaningful interpretations on published infection case data.
Shan Huang (DIW Berlin and University of Copenhagen)
Abaluck, Jason, Leila Agha, Chris Kabrhel, Ali Raja, and Arjun Venkatesh. 2016. The Determinants of Productivity in Medical Testing: Intensity and Allocation of Care. American Economic Review, 106 (12): 3730-64. DOI: 10.1257/aer.20140260