Nonmedical influences on medical decision making: an experimental technique using videotapes, factorial design, and survey sampling

Health Services Research, August, 1997 by Henry A. Feldman, John B. McKinlay, Deborah A. Potter, Karen M. Freund, Risa B. Burns, Mark A. Moskowitz, Linda E. Kasten

Interviews. To standardize the setting and to place the subject in his usual context for decision making, the interview was conducted in the physician's office during normal clinic hours. The format was a semi-structured interview, conducted by senior NERI staff with prior experience interviewing physicians. One interviewer was male and the other female; half of the physicians and half of the videotapes were assigned to each. The interviews were conducted between August 1993 and June 1994.

After watching each scenario, the physician was invited to order further diagnostic evaluation. Requests for a specific test were answered with a simulated laboratory report. The physician was allowed to act on the results by ordering further tests. After receiving test results, the physician was asked what recommendation for evaluation and follow-up he would make and what information on alternatives he would offer to the patient.

End Points. Because the results of the breast cancer study are not presented in this report, the outcomes will be sketched very briefly; full details can be found elsewhere (Freund, Burns, Moskowitz, et al. 1995; Burns, Freund, Moskowitz, et al. submitted; McKinlay, Burns, Durante, et al. 1997; Kasten, McKinlay, Freund, et al. submitted). For the prediagnosis scenario, interest centered on the physician's estimate of the likelihood of breast cancer and his diagnostic strategy. An important issue was whether tissue analysis was planned. For the postdiagnosis scenario, the principal issues were the extent of staging evaluation (axillary node dissection, metastatic evaluation), alternatives for primary therapy (breast-conserving surgery, lumpectomy with radiation, mastectomy with or without reconstructive surgery), and options for adjuvant therapy (tamoxifen, other chemotherapy, or no further treatment).

Table 6: Precision of Prevalence Estimates from Breast Cancer Study

Response Probability         Standard Error(*)

.100                               .027
.250                               .038
.400                               .043
.500                               .044
.600                               .043
.750                               .038
.900                               .027

* Based on binomial distribution, n = 128 physicians.

The Monte Carlo simulation was conducted for a representative variable from each group of formally equivalent factors in the design: a fully balanced patient factor (assertiveness), a partially balanced patient factor (physical condition), and a physician factor (experience). Power was calculated in each case for a variety of baseline probabilities and effect magnitudes (50 percent versus 75 percent, 80 percent versus 90 percent, etc.). The results are displayed in Table 7.

Table 7 shows that power was excellent as long as the prevalence of the binary response differed by at least 25 percent between the groups being compared. For example, if the effect of the patient's assertiveness were to raise from .50 to .75 the physician's likelihood of recommending biopsy, the estimated power of the experiment to detect that effect would be 92 percent. Power was near-perfect (100 percent) for comparing underlying rates any farther apart. The power estimates for effects of assertiveness also apply to effects of age, race, or SES because of the symmetry of the design. Only slightly lower power, attributable to the fractional design, was estimated for effects of the patient's physical and medical condition and the physicians' strata. Again, the power was sufficient (80 percent or greater) as long as the difference in underlying rates was at least .25.