We analyzed data from all four rounds of the Community Tracking Study (CTS) Physician Survey, a telephone survey of nationally representative samples of U.S. physicians conducted in 1996–1997, 1998–1999, 2000–2001, and 2004–2005. The physician sample was drawn from the American Medical Association and the American Osteopathic Association master files and included active, nonfederal, office-based, and hospital-based physicians who spent at least 20 hours a week in direct patient care. Residents and fellows, as well as radiologists, anesthesiologists, and pathologists were excluded.Rounds 1–3 each included approximately 12,000 physicians. Because of funding cuts, round 4 included a smaller, but more statistically efficient and still nationally representative sample of 6,628 physicians. Cross-sectional survey response rates ranged from 52% to 65%, which are relatively high for surveys of physicians, and the lowest of these (52% in round 4) has been demonstrated to not bias estimates relative to the highest response rate of 65% from round 1.9
The second, third, and fourth rounds of the survey included physicians sampled in the previous round as well as new physicians. This approach resulted in three panels of physicians, spanning two consecutive rounds each, allowing us to track changes in responses from individual physicians between rounds. To create the first panel, 9,353 of the 12,385 respondents (76%) to the first round were contacted in the second round. Of these, a total of 915 had become ineligible because they had retired, worked fewer than 20 hours per week, or were not locatable. Of these, 7,092 completed the round 2 survey. After excluding those physicians whose specialty changed between rounds or who had missing data for key variables, there were 7,057 remaining in panel 1 for our analysis. Applying a similar process in later rounds resulted in 8,487 physicians in panel 2 (75% response rate) and 4,401 physicians in panel 3 (76% response rate). Item nonresponse for each round was typically less than 3%, and less than 0.1% for our dependent variable. Further details about the survey have been previously published and are available at http://www.hschange.org/index.cgi?data=04
Outcome Variables Our outcome was the reported effect of CPGs on the physician’s practice of medicine. In each round of the survey, physicians were asked: “How large an effect does your use of formal, written practice guidelines such as those generated by physician organizations, insurance companies or HMOs, or government agencies have on your practice of medicine? Would you say that the effect is very large, large, moderate, small, very small, or no effect at all?” These responses were coded on a scale from 6 to 1. (If a physician said that he/she uses his/her own guidelines, then the interviewer would say: “In this question, we are only interested in the use of formal, written guidelines such as those generated by physician organizations, insurance companies or HMOs, or other such groups.)We analyzed the original six-category variable for cross-sectional analyses. For longitudinal analyses, the outcome variable is the “change in effect of guidelines” from one round to the next. We calculated values for this change for panel respondents by taking the difference in reported effect of guidelines for the same respondent in consecutive rounds of the survey, with the resulting value ranging from −5 to +5.
We evaluated the extent to which changes in physicians’ practice characteristics were associated with the effect that they reported CPGs had on the care they provided. We considered four types of factors that prior studies have suggested play a role in guideline adoption: (1) the practice environment, (2) relevant IT, (3) exposure to financial incentives based on performance, and (4) revenue sources.Practice environment included practice size (<10 vs 10 or more physicians), practice ownership (physician is full or part owner of practice vs an employee), and percentage of practice revenue paid on a capitated or other prepaid basis. The question on IT addressed the availability of a computer or other forms of IT in the practice to access CPGs. We also asked physicians whether “specific measures of quality of care, such as rates of preventive care services delivered,” and “results of practice profiling comparing their pattern of using medical resources to treat patients with that of other physicians” were explicitly considered in determining their compensation. Response options to both compensation items were yes/no.Finally, respondents indicated the percentage of their practice revenue derived from Medicaid and Medicare. Because managed care penetration varies for Medicaid and Medicare populations, we felt it was important to control for these types of insurance in the multivariate models. In 2004, approximately 12% of Medicare beneficiaries were enrolled in managed care plans versus 60% of Medicaid enrollees in managed care (http://www.cms.hhs.gov/apps/media/press/release.asp?Counter=1783
). Changes in the proportion of revenue from these sources might alter the proportion of one’s patient panel under managed care processes and hence influence a practitioner’s need to be responsive to CPGs. Whereas the CTS survey has items on specific practice type (physician group practice, hospital practice, group model HMO, etc.) there was little movement of physicians between types of settings from one round to the next, and thus too little variation for this characteristic to be an informative independent variable.We defined independent variables using a set of dummy variables for each practice characteristic, reflecting changes in responses to the relevant survey question between two consecutive rounds, for individual physicians in each panel. For continuous variables (e.g., percentage of revenue from Medicaid), the difference in percentage was calculated. If the difference from the earlier to later round exceeded 5 percentage points, it was categorized as either an increase
depending on its direction; otherwise, the value was categorized as no change
. For categorical variables (e.g., availability of IT to access guidelines, Yes/No) the change from earlier to subsequent round was noted as, either an increase (if one gained IT access), a decrease (if one lost IT access), or no change (if one’s IT access remained the same) from the first round of the panel to the next.We also characterized physicians in cross-sectional analyses in terms of their age, sex, years in practice, board certification status, medical education (United States/Canada vs elsewhere), and primary specialty. When we distinguish primary care physicians (PCPs) from specialists, the PCP category includes general internal medicine, family practice, general practice, geriatrics, and pediatricians. Specialists include medical or surgical specialists.
Because PCPs and specialists may operate under different practice characteristics, face different degrees of pay-for-performance pressures, and be exposed to different numbers of CPGs,10
we stratify analyses by PCP versus specialists. Percentages were weighted to be nationally representative and to account for the complex sample design.
Longitudinal Analyses We used the panel data to construct a first differences model. This allowed us to examine whether changes in practice characteristics were associated with changes in the reported effect of CPGs on physicians’ clinical practice over time. These longitudinal analyses have the advantage over cross-sectional analyses of avoiding bias because unobserved physician preferences and characteristics are held constant, with each physician in effect acting as his or her own control. We also included baseline ratings of the “effect of CPGs on one’s clinical practice” in the regression model to control for possible bias because of floor and ceiling effects, and the possibility that different physicians may tend to rate the CPG effect more or less highly under similar conditions. Similar baseline indicators were used for the IT variable and the profiling variable to account for the presence of those characteristics at baseline. (A physician who already had IT in the earlier round and kept it in the latter round would be classified as no change as would a physician who lacked IT in both rounds. The inclusion of the baseline indicator variable accounts for the fact that these two subgroups with no change are not equivalent.)Because bivariate relationships between the independent and dependent variables were consistent across the three panels, and to maximize statistical power, we combined all three panels in multivariate analyses. Multivariate ordinary least squares regression allowed us to examine the independent effect of change in each practice characteristic, holding all other factors constant. Whereas we were not powered to run separate change models for PCPs versus specialists, we did include a dummy variable to indicate whether the respondent was a PCP or a specialist. We also controlled for secular trends by including dummy variables indicating each panel. All analyses were conducted using SAS version 9.1 (SAS Institute Inc, Cary, NC, USA) and SUDAAN version 9.0 analytic software (Research Triangle Park, NC, USA), which used the appropriate weights and accounted for the complex sample design as well as for the nonindependence of observations when the same physician was included in multiple panels.