Experimental Design Details
We will employ a vignette survey experiment to assess treatment decisions by presenting physicians with randomized patient scenarios. Each vignette represents a hypothetical patient case, including background information and symptoms. First, we elicit the prior disease probability based on the patient characteristics and the prescription decision. On the next page, physicians observe a randomly drawn signal from either an AI support tool (AI tool treatment) or from a dipstick test (Dipstick treatment). After seeing the signals, we elicit their posterior disease probability again, as well as their final treatment decision. This choice situation is repeated for six vignettes. Afterwards, participants see another three vignettes, but this time they receive both the AI support tool and the dipstick signal. We vary experimentally whether the dipstick signal is included in the AI signal (Correlated treatment) or whether it is not included (Uncorrelated treatment). Again, we ask for their diagnostic beliefs and treatment decisions.
We test the following experimental hypotheses:
H1: Physicians update differently to the AI signal compared to the dipstick signal.
H2: Physicians update differently in the Correlated treatment compared to the Uncorrelated treatment.
We will study heterogeneous treatment effects for hypotheses H1 and H2 by age, gender, education (university attended), prescription intensity (based on admin data; individual and clinic-level prescribing), clinic characteristics (such as clinic size, number of physicians working at the clinic, consultations per physician, rural/urban, based on admin data), diagnostic intensity (based on admin data; individual and clinic-level claims), and time preferences (based on survey data).
To test the hypotheses, we will compare the average absolute updating using parametric and non-parametric tests. Moreover, we will estimate Bayesian updating regressions following Grether (1992, JEBO) and compare the updating coefficients. In further analyses, we will compare updating to positive vs. negative signals.
After the vignettes, we ask for physician attitudes towards the AI support tool and the dipstick test. We will analyze whether these attitudes predict average belief updating to the respective signal. Moreover, we study whether general attitudes toward AI predict belief updating to the AI signal.
We will exploratively investigate whether differences in prescription and diagnostic intensity (in the survey and the admin data) are associated with differences in prior beliefs, reaction to signals, and/or different prescription thresholds (as measured in the experiment). Similarly, we will investigate whether individuals with different concern about AMR externality differ with respect to these factors.
Moreover, we elicit a range of physician characteristics (risk, time, and social preferences, concern about antimicrobial resistance, belief about own prescription intensity) and will exploratively analyze whether these characteristics predict antibiotic prescription decisions in the survey as well as prescribing and diagnostic intensities in the administrative data. Furthermore, we will exploratively analyze the determinants of technological and AI adoption.
For robustness, we will exclude participants who repeatedly updated in the wrong direction (i.e., who updated positively to a negative signal, or negatively to a positive signal).