Two thousand physicians were randomly selected from the national list of physicians provided by the Medical Marketing Service, Inc (MMS). The Medical Marketing Service list is based on the national database of the American Medical Association (AMA) which includes both members and nonmembers of the American Medical Association, and is updated weekly. The American Medical Association database contains over 650000 physicians, and is the most-complete list of physicians available in the United States. Physicians who currently spent over 20 hours a week on direct patient care were included in the survey. The sample was stratified by specialty: primary care, medical specialty, or surgical specialty. Primary care included family practice, general practice, internal medicine, and pediatrics. Ob-Gyn was classified as a surgical specialty.

The questionnaire was developed following literature review and focus-group discussions. It was pretested to ensure that the instrument was easy to complete, all areas of interest were covered, and no questions were ambiguous. It consisted of closed-end questions, took approximately 12 minutes to complete, and was in 3 parts. The entire sample received Part 1 of the questionnaire, which elicited general information about views on health information on the Internet and direct-to-consumer advertising (DTCA). Questions included general views on accuracy and effects of such information, and personal use of the Internet at work. Part 2 was sent to a random 50% of the sample, and requested information about the last time a patient brought in information from the Internet. "Last-time" methodology was used to minimize recall bias. Areas explored were the relevance and accuracy of the information, physicians' perceptions of why the patient had brought the information, physicians' responses to the patient, and their views about the impact on health care, health outcomes, and the physician-patient relationship. The other 50% of the sample received a different Part 2, which explored these same areas but with regard to the last time a patient brought in information from direct-to-consumer advertising. The direct-to-consumer advertising data are presented elsewhere [9]. Part 3 was received by the entire sample and obtained demographic and workload information: hours per week on face-to-face consultations, on other tasks related to patient care, and on administrative tasks; numbers of patients seen per week; practice income; proportions of patients on Medicaid, from minority groups, having household incomes of less than $20000 per annum, and without health insurance; geographic setting of practice; age and racial origin of respondent. This was supplemented with information from the Medical Marketing Service database including specialty, year of graduation from medical school, geographic region (East, South, Midwest, West), whether hospital-based or office-based, and whether trained in the United States or overseas.

Data collection was undertaken between November 2000 and February 2001. The questionnaire was mailed to the selected physicians with a check for US $35 as a token of appreciation for completing the questionnaire. Up to 3 reminders were sent and additional telephone contact made with nonresponders. Of the original 2000 physicians sent the survey, 38 were ineligible because they were deceased, retired, or no longer in practice; and 1050 physicians completed the questionnaire (response rate 53%). Of these, 515 received the Internet version of the questionnaire, and 535 the direct-to-consumer advertising version.

Data were weighted to represent the national population of physicians in the Medical Marketing Service database who spend 20 or more hours per week on direct patient care, using the Medical Marketing Service variables mentioned above. As can be seen in Table 1, there is little difference between weighted and unweighted data, confirming that respondents were representative of US (United States) physicians.

The analytic approach focused on evaluating univariate and multivariate relationships with 4 clinically-important outcomes — change in physician-patient relationship; time efficiency; quality of care; and patient health outcome — each of which was assessed on a 3-point scale (improved vs no difference vs worsened). All the demographic, workload, and practice variables listed in Table 1 were run against each of these 4 outcome variables. Univariate relationships were calculated using the chi-square statistic or Fisher exact test as appropriate. In addition, univariate relationships were also investigated for an intermediate outcome: whether or not the physician did what the patient requested (yes completely vs yes partially vs no), a variable which in turn is evaluated for its relationship with the 4 main outcome variables.

Although several of the workload and practice characteristics were assessed as continuous variables (eg, percentage of patients who were uninsured, average number of patients seen per week), most were highly skewed, so medians and interquartile ranges are reported for these data. These variables were split at the 75th percentile for analysis of univariate relationships to test for the influence of these factors. This split was chosen over a median split to maximize the opportunity for an effect to be visible.

Separately for each outcome variable, correlates with chi-square statistics achieving P< .20 were analyzed using a stepwise multiple-logistic regression procedure to determine the "most-important" correlates, where importance is defined solely by statistical criteria. Each analysis went through several iterations, with each new iteration employing successively more-stringent statistical criteria for inclusion in the model. Each iteration included consideration of a model yielded by a forward-stepwise procedure and a model yielded by a backward-stepwise procedure. Final models include all correlates with a significant ( P< .05) or near-significant (.05 < P<.10) likelihood ratio test while still achieving adequate fit, operationalized as P> .20 on the Hosmer-Lemeshow goodness-of-fit test.

As all data were weighted (except where specified), the appropriate procedures to correct P values and standard errors were undertaken. We used the SVYTAB procedure in STATA to obtain the Rao and Scott F-test P-values [10], and the SVYLOGIT procedure in STATA to obtain corrected standard errors for parameter estimates.

The characteristics of the respondents before and after weighting are presented in Table 1. Weighting made only minimal difference to the characteristics of the sample, confirming that respondents were representative of US physicians. From this point on, all data presented are weighted.

Sixty-one percent (n = 639; 95% CI, 58%-64%) of all respondents used the Internet in their own practice. In this group, the most-frequent uses were to obtain scientific information such as articles or guidelines (88%; 95% CI, 86%-91%) or to e-mail colleagues (63%; 95% CI, 59%-67%). Obtaining clinical information about patients, such as lab results (28%; 95% CI, 25%-32%), and e-mailing patients (16%; 95% CI, 13%-18%) were much less common uses of the Internet by physicians.

Overall, respondents were positive about the recent increase in health information on the Internet, with 75% (95% CI, 72%-77%) of the total sample thinking that it was a good or very-good thing. Only 15% (95% CI, 13%-17%) believed that it was a bad thing, and the remainder were neutral. Similarly, most physicians (77%; 95% CI, 74%-79%) stated that they had encouraged patients to look for information, although only 35% (95% CI, 32%-38%) had referred patients to Web sites.

Eighty-five percent (95% CI, 82%-87%) of all respondents had experienced an occasion when a patient brought information from the Internet to a visit. For most physicians this is still a relatively-rare event; 59% (95% CI, 56%-62%) of respondents stated that less than one fifth of their patients had done this. 87% (95% CI, 85%-89%) of physicians perceived their patients as being concerned about the quality of information on the Internet, and 84% (95% CI, 82%-86%) of respondents rated their patients as only fair or poor (rather than good, very good, or excellent) at appraising the quality of information on a Web site .

Although our response rate is only moderate at 53%, it compares well to other surveys of Internet use by physicians. Because our sample was representative of all US physicians in terms of age, gender, specialty, location of practice, and practice income our results are likely to generalize to all US physicians. In contrast, previous surveys have examined specific branches of medicine [21], used convenience samples [22] or Internet-literate samples [23], had unacceptably-low (21%) response rates [24], or had very-small samples [25]. Response rates in other recent surveys of US physicians are lower than ours [26- 29], and the absence of substantive differences between responders and nonresponders argues against the presence of systematic selection bias.

As with all cross-sectional studies, we cannot determine causality, nor do we have objective data on whether patient requests were truly inappropriate or on quality of care or health outcomes. However, our measures are plausible because physicians address the appropriateness of care and outcomes daily on a professional basis. Patient perceptions of these consultations may have been different, but our results from a population survey of public perceptions of the effects of health information on the Internet are not dissimilar [30].

Health care organizations, payers, and providers have a strong interest in ensuring both that health information on the Internet is accurate and that physicians have the necessary skills to respond to patients who bring in such information. Vigorous leadership in these areas will be needed if the effect of the Internet on medicine is to be truly beneficial.