Purpose: To determine the implicit treatment thresholds of physicians when making recommendations in cases of ocular hypertension, a risk factor for glaucoma. Thresholds were determined using a new method that accommodates uncertainty in recommendations.

Methods: A study of treatment actions in simulated cases without the aid of a risk calculator. We generated simulated cases of ocular hypertension using the results of the Ocular Hypertension Treatment Study. Each case consisted of values for six risk factors related to the development of glaucoma. We then calculated the true risk of developing glaucoma for each case using the results of a published, validated calculator. Recruited glaucoma specialists were asked to make a treatment recommendation for each of 50 scenarios using a seven point scale from "Definitely No" to "Definitely Yes". To estimate the five year risk at which each physician would recommend treatment to prevent glaucoma, we first reduced the treatment recommendations to a 3-point scale (No, Unsure, Yes). We then used used ordinal regression to predict the treatment recommendations using the true risk of developing glaucoma as the independent variable. We derived an expression for the value of risk at which a physician was equally likely to make a Yes or No recommendation (i.e., was maximally uncertain) and defined this point as the threshold for treatment. For those physicians who only used two of the three recommendation categories, their treatment threshold was similarly estimated using logistic regression.

Results: Regression models were successfully created for all 56 of the physicians participating in the study. Risk was a statistically significant predictor of treatment in 51 (91%) of the models. The average risk threshold for these 51 physicians was 21 +/- 13%.

Conclusions: We have defined and tested a new method for estimating implicit treatment thresholds using case scenarios and a graded treatment recommendation scale. In the case of ocular hypertension, this method produced threshold values that were within the expected range, although the physicians in the study tolerated higher risk than either expert opinion (15%) or cost-benefit analysis (10%) suggest. This method has the advantage of using physician behavior to determine a threshold rather than asking explicitly for a value not used in practice.