Prevalence of Health Problems and Primary Care Physicians’ Specialty Referral Decisions

Primary care physicians make specialty referrals to obtain advice for clinically uncertain diagnostic evaluations or treatment plans, to obtain a specialized service that falls outside their scope of practice, because of patient or third-party requests, or because of a combination of these reasons.¹ The clinical reasons for these referral decisions include characteristics of the presenting health problem, the burden and severity of comorbidities, and patient preferences for various treatments and outcomes.

Previous research has shown that certain features ascribed to morbidities influence the likelihood of specialty referral. The type of diagnosis is the most obvious determinant. In one study,² adults with malignancies were 5 times more likely to be referred than those with respiratory illnesses. Similarly, in the Netherlands, Van Suijlekom-Smit and colleagues³ found more than an 8-fold variation among childhood diagnosis groups in the likelihood of referral. For patients with similar diagnoses, research has found that severe variants are more likely to be referred.^4-7 Specialty referral is also influenced by the array and complexity of comorbid conditions.⁸

The conceptual foundations of primary care provide further insight into how clinical factors may influence referral to specialty care. A defining feature of primary care is the provision of a comprehensive set of services that meets the majority of a population’s health needs.^9,10 Primary care physicians develop greater experience and expertise for health problems with which they are familiar than those that occur less often. It follows that they would seek specialist assistance for uncommon health problems. However, empirical evidence for this effect is currently lacking.

Our goal was to test the hypothesis that the frequency with which a condition is seen by primary care physicians (practice prevalence) influences the likelihood of referral from primary to specialty care. We use the term practice prevalence to mean the frequency of presentation to primary care physicians and to distinguish it from the frequency of occurrence in the community. Also, we examine the impact of other clinical factors on primary care physicians’ referral decisions, including patient age, sex, comorbidities, and the medical versus surgical nature of the target condition’s management.

Methods

Data Source and Study Sample

We used the 1989 to 1994 National Ambulatory Medical Care Surveys (NAMCS) to examine referrals made to specialist physicians during visits with primary care physicians. NAMCS is a nationally representative survey of office-based physician visits in the United States. Each year, a multistage probability sample of nonfederally funded US physicians who are engaged in patient care activities (excluding radiologists, anesthesiologists, and pathologists) is selected from the master files of the American Medical Association and the American Osteopathic Association. For 1 week each selected physician completes a questionnaire for a 20% to 100% systematic sample of patient visits. Details of the survey methodology and the survey instrument are presented elsewhere.¹¹ The distribution of patient age and sex remained consistent over the 6 years of data collection we used.¹² The 1995 to 1998 surveys were not used, because information on referral was not collected. Using the 1994 and 1998 NAMCS, Forrest and Whelan¹³ found that primary care practice patterns did not substantively differ over time. The pooled data set contained 219,830 visits, of which 78,107 (35.5%) were with generalists (self-reported specialty designation was family/general practice, general pediatrics, or general internal medicine).

Clinical Factors

To examine the referral characteristics of different types of conditions, International Classification of Diseases-ninth revision-Clinical Modification (ICD-9-CM) codes were grouped into clinically similar categories that we called expanded diagnosis clusters (EDCs). The EDCs were based on the diagnosis clusters originally developed by Schneeweiss and coworkers.¹⁴ EDCs were assigned to the vast majority of primary care visits in the data set, matching 93.6% of the unique diagnosis codes. EDCs were grouped into clinical domains based on the nature of the problems and the specialty most responsible for the care. We also classified each EDC according to whether the primary treatment approach was medical or surgical.

A practice-prevalence ratio was calculated for each condition. The numerator was the number of visits for the EDC under consideration, and the denominator was the total number of visits in the data set. These ratios were further divided into tertiles (ie, high-, medium-, and low-frequency EDCs.)

To account for the number and severity of comorbid conditions with which patients presented, we assigned a comorbidity index to each visit. The index was based on the aggregated diagnostic groups (ADGs), the building blocks of The Johns Hopkins Adjusted Clinical Groups Case-Mix system.¹⁵ The 32 ADGs represent morbidity groups that contain ICD-9-CM diagnosis codes that are similar with respect to likelihood of persistence, expected need for health care resources, and other clinical criteria. In the NAMCS, up to 3 ICD-9-CM codes were assigned per visit; thus, each visit was assigned 1 to 3 ADGs. A comorbidity index score was obtained by summing ADG-specific resource intensity weights.* Larger weights suggest more complex health problems and greater expected resource intensity. In previous research,¹³ the comorbidity index increased with patient age and distinguished type of primary care delivery site by morbidity burden. The comorbidity index was divided into tertiles representing high, medium, and low levels of comorbidity.

Data Analysis

The unadjusted association of each clinical factor with the probability of referral was evaluated in bivariate analyses using the chi-square statistic. We further analyzed the relationship between a condition’s practice prevalence and its referral rate using scatter plots and Pearson correlation coefficients. Because both the practice prevalence and referral rate variables were right skewed, we used a logarithmic transformation to normalize both. For the scatter plots we excluded conditions with unstable referral rate estimates because of small sample sizes. Specifically, a condition was included if the ratio of the difference between the 95% confidence interval upper and lower limits of the referral rates divided by the condition-specific referral rate was less than 1. For example, multiple sclerosis was excluded because it presented at a rate of just 6 per 10,000 visits, had a referral rate of 7.1%, and the difference between the 95% confidence interval limits divided by the referral rate was 2.7.

We conducted a multiple logistic regression analysis with the patient visit as the unit of analysis for examining the independent effect of practice prevalence of the presenting problem on the chances of referral. Controlling variables in this analysis included age, sex, comorbidity index tertile, and the medical versus surgical nature of the principal diagnosis. To determine if the clinical complexity of a visit as assessed by the comorbidity index modified the effect of practice prevalence, we also added interaction terms for these variables to the final model. The goodness of fit of the final model was tested using the Hosmer-Lemeshow statistic, which suggested adequate model fit.¹⁶ The regression analyses used the generalized estimating equation¹⁷ to control for the intraphysician correlation of clustered visits.

Results

The distribution of clinical characteristics is shown in Table 1. The population was somewhat more represented by women (57.5%) than men, and by children/adolescents (33.3%) and seniors (39.4%) than middle-aged persons. The average comorbidity index value was 0.63, but there was considerable variability and the distribution was right skewed (median=0.31; interquartile range=0.05-0.83). The mean EDC practice-prevalence estimate for patient visits was 51 per 1000 primary care visits, but this distribution was also right skewed (median=25; interquartile range=10-106).

In the bivariate analyses, the probability of referral was significantly related to age, sex, practice prevalence of the condition, patient comorbidity, and whether the condition was a surgical problem Table 2. For the 65 condition categories with stable referral rate estimates Figure 1, the correlation between the logarithms of the practice-prevalence ratios and the referral rates was -0.87 (P <.001). Also, this relationship remained after stratifying for the 35 medical (r=-0.82; P <.001) and 24 surgical conditions (r=-0.88; P=.001).

Table 3 shows the multiple logistic regression results modeling the probability of referral during an office visit as a function of the clinical factors. There were no significant differences in the chances of referring patients aged 11 to 64 years once measures of morbidity, such as patient comorbidity and practice prevalence of the principal diagnosis, were controlled. Men were 19% more likely to be referred than women, and medical conditions were 39% less likely to be referred than surgical ones.

Only one of the practice prevalence/comorbidity interaction terms was significantly different from 0: commonly occurring conditions presenting among patients with high levels of comorbidity. This finding implies that the comorbidity has a stronger influence on the chances of referral for patients presenting with common problems than those presenting with less common problems.

Table 4 shows the estimated probabilities of referral based on differences in practice prevalence and comorbidity. These probabilities were obtained from the b coefficients in Table 3. The reference group for the probability estimates is women aged 18 to 44 years with health problems categorized as medical conditions. The chances of referral varied as much as 8-fold based on only the practice prevalence of the principal diagnosis and level of comorbidity.

Discussion

Our results support the hypothesis that the frequency with which patients’ health problems present to primary care physicians (practice prevalence) has a strong inverse relationship with the chances of referral to specialty care. Primary care physicians were more likely to send patients with uncommon problems to specialists and retain those with the most common conditions. This finding highlights the responsible judgment primary care physicians employ in recognizing the boundaries of their scope of practice. Practice prevalence is a defining feature of the primary care–specialty care interface.

Referring patients with uncommon problems to specialists is a rational way to organize medical care. Outcomes are related to the volume of patients managed with a specific condition.¹⁸ Specialists need to care for an adequate number of patients with uncommon problems to maintain clinical competence. Patient self-referral, however, which dilutes the prevalence of health problems presenting to specialists, may result in potentially invasive and expensive diagnostic approaches to patients more appropriately evaluated by primary care physicians.¹⁹

In addition to a condition’s practice prevalence, the number and severity of comorbidities managed during the visit influenced primary care physicians’ decisions to make specialty referrals. Also, we found an interaction effect between high practice prevalence and high levels of comorbidity. In other words, patients with uncommon conditions were commonly referred, regardless of the complexity of other conditions. The chances of referral markedly increased for patients with common conditions when they also presented with co-existing medically complex health problems. Thus, the rare presentations for which specialist assistance is sought may be a result of either the practice prevalence of the presenting problem or the overall complexity of a patient.

Men were more commonly referred than were women, after accounting for differences in the nature of their problems. A possible explanation for this finding is that because women make more office visits over a year than men,²⁰ their probability of referral during any given visit will be lower given roughly equal chances of referral between the 2 groups during the course of a year.

Further Research

We demonstrated that the potential need for surgical interventions was an important predictor of referral. Even after other clinical factors were controlled, medical conditions were 39% less likely to be referred than surgical ones. This is not surprising given that primary care physicians generally perform only minor office-based surgical procedures. But which surgical procedures should be in the scope of practice of primary care physicians? This question deserves further research and could be addressed in part by an analysis that is similar to the one presented here. Common outpatient procedures are candidates for inclusion as primary care services. Secondary considerations include the requirements and expense of necessary equipment, technical personnel, and training. Research that builds epidemiologic profiles of office-based procedures would be helpful in determining how responsibilities should be divided between generalists and specialists for these technical services.

Limitations

Several limitations in our study’s data source warrant consideration. First, the data set of visits provided information on primary care physicians’ referral decisions and did not elucidate whether patients actually received specialty care. Second, the sample was restricted to visits made to generalist physicians, excluding both obstetrician-gynecologists and medical subspecialists who may act as primary care physicians. Third, the NAMCS data set did not include hospital-based physicians, who are known to have higher referral rates than their office-based counterparts.¹³ Fourth, the unit of analysis was the visit rather than the patient. Patients with certain chronic conditions may have higher referral rates than suggested by our data if the measure used is the percentage of persons obtaining specialty care over a year. The advantage of focusing on the visit is that physician referral decisions can be examined rather than specialist use. Fifth, some conditions had lower than expected referral rates (eg, appendicitis had a referral rate of 46%), because the denominator for the referral rates was all visits made to generalists for the condition, which included both new presentations and follow-up visits. Finally, because of data limitations we did not assess the extent to which condition prevalence within an individual physician’s own practice affects his or her referral behavior.

Specialist visits can be initiated by primary care physician referral, patient self-referral, or specialist-to-specialist cross-referral. Although our database did not permit us to examine each of these pathways, other research suggests that primary care physician referral is the predominant route, particularly in health maintenance organizations.¹²

Conclusions

Our findings provide evidence that the boundaries between primary care physicians and specialists are defined in part by prevalence of health problems and the overall complexity of patients. Future research should focus on identifying modifiable characteristics of the physician-patient interaction, physicians, their practices, and the health system that influence referral decisions, after accounting for clinical factors. The appreciation of relevant clinical factors is critical to the fair application of administrative and financial constraints on physicians’ abilities to refer. Managed care plans that penalize physicians for high referral behavior, without adjusting for practice prevalence and comorbidity work, are contrary to the goal of providing quality patient care in the most appropriate settings. With more precise definitions of the clinical determinants of referral for populations, health systems can better gauge generalist and specialist workforce requirements.

Acknowledgments

This work was supported by the Agency for Healthcare Research and Quality grants #R01 and #HS09377. Barbara Starfield inspired this work and provided comments on the manuscript. We also thank Barbara Bartman, Norm Smith, MD, MPH, and Jonathan Weiner MD, MPH, for their review and comments on the manuscript. Mia Kang and Sarah von Schrader provided excellent technical assistance.

Related Resources

• Agency for Healthcare Research and Quality, Primary Care Subdirectory Page—includes research articles on primary care referral patterns and coordination of care among referring physicians and specialists. http://www.ahrq.gov/research/primarix.htm

Primary care physicians make specialty referrals to obtain advice for clinically uncertain diagnostic evaluations or treatment plans, to obtain a specialized service that falls outside their scope of practice, because of patient or third-party requests, or because of a combination of these reasons.¹ The clinical reasons for these referral decisions include characteristics of the presenting health problem, the burden and severity of comorbidities, and patient preferences for various treatments and outcomes.

Previous research has shown that certain features ascribed to morbidities influence the likelihood of specialty referral. The type of diagnosis is the most obvious determinant. In one study,² adults with malignancies were 5 times more likely to be referred than those with respiratory illnesses. Similarly, in the Netherlands, Van Suijlekom-Smit and colleagues³ found more than an 8-fold variation among childhood diagnosis groups in the likelihood of referral. For patients with similar diagnoses, research has found that severe variants are more likely to be referred.^4-7 Specialty referral is also influenced by the array and complexity of comorbid conditions.⁸

The conceptual foundations of primary care provide further insight into how clinical factors may influence referral to specialty care. A defining feature of primary care is the provision of a comprehensive set of services that meets the majority of a population’s health needs.^9,10 Primary care physicians develop greater experience and expertise for health problems with which they are familiar than those that occur less often. It follows that they would seek specialist assistance for uncommon health problems. However, empirical evidence for this effect is currently lacking.

Our goal was to test the hypothesis that the frequency with which a condition is seen by primary care physicians (practice prevalence) influences the likelihood of referral from primary to specialty care. We use the term practice prevalence to mean the frequency of presentation to primary care physicians and to distinguish it from the frequency of occurrence in the community. Also, we examine the impact of other clinical factors on primary care physicians’ referral decisions, including patient age, sex, comorbidities, and the medical versus surgical nature of the target condition’s management.

Methods

Data Source and Study Sample

We used the 1989 to 1994 National Ambulatory Medical Care Surveys (NAMCS) to examine referrals made to specialist physicians during visits with primary care physicians. NAMCS is a nationally representative survey of office-based physician visits in the United States. Each year, a multistage probability sample of nonfederally funded US physicians who are engaged in patient care activities (excluding radiologists, anesthesiologists, and pathologists) is selected from the master files of the American Medical Association and the American Osteopathic Association. For 1 week each selected physician completes a questionnaire for a 20% to 100% systematic sample of patient visits. Details of the survey methodology and the survey instrument are presented elsewhere.¹¹ The distribution of patient age and sex remained consistent over the 6 years of data collection we used.¹² The 1995 to 1998 surveys were not used, because information on referral was not collected. Using the 1994 and 1998 NAMCS, Forrest and Whelan¹³ found that primary care practice patterns did not substantively differ over time. The pooled data set contained 219,830 visits, of which 78,107 (35.5%) were with generalists (self-reported specialty designation was family/general practice, general pediatrics, or general internal medicine).

Clinical Factors

To examine the referral characteristics of different types of conditions, International Classification of Diseases-ninth revision-Clinical Modification (ICD-9-CM) codes were grouped into clinically similar categories that we called expanded diagnosis clusters (EDCs). The EDCs were based on the diagnosis clusters originally developed by Schneeweiss and coworkers.¹⁴ EDCs were assigned to the vast majority of primary care visits in the data set, matching 93.6% of the unique diagnosis codes. EDCs were grouped into clinical domains based on the nature of the problems and the specialty most responsible for the care. We also classified each EDC according to whether the primary treatment approach was medical or surgical.

A practice-prevalence ratio was calculated for each condition. The numerator was the number of visits for the EDC under consideration, and the denominator was the total number of visits in the data set. These ratios were further divided into tertiles (ie, high-, medium-, and low-frequency EDCs.)

To account for the number and severity of comorbid conditions with which patients presented, we assigned a comorbidity index to each visit. The index was based on the aggregated diagnostic groups (ADGs), the building blocks of The Johns Hopkins Adjusted Clinical Groups Case-Mix system.¹⁵ The 32 ADGs represent morbidity groups that contain ICD-9-CM diagnosis codes that are similar with respect to likelihood of persistence, expected need for health care resources, and other clinical criteria. In the NAMCS, up to 3 ICD-9-CM codes were assigned per visit; thus, each visit was assigned 1 to 3 ADGs. A comorbidity index score was obtained by summing ADG-specific resource intensity weights.* Larger weights suggest more complex health problems and greater expected resource intensity. In previous research,¹³ the comorbidity index increased with patient age and distinguished type of primary care delivery site by morbidity burden. The comorbidity index was divided into tertiles representing high, medium, and low levels of comorbidity.

Data Analysis

The unadjusted association of each clinical factor with the probability of referral was evaluated in bivariate analyses using the chi-square statistic. We further analyzed the relationship between a condition’s practice prevalence and its referral rate using scatter plots and Pearson correlation coefficients. Because both the practice prevalence and referral rate variables were right skewed, we used a logarithmic transformation to normalize both. For the scatter plots we excluded conditions with unstable referral rate estimates because of small sample sizes. Specifically, a condition was included if the ratio of the difference between the 95% confidence interval upper and lower limits of the referral rates divided by the condition-specific referral rate was less than 1. For example, multiple sclerosis was excluded because it presented at a rate of just 6 per 10,000 visits, had a referral rate of 7.1%, and the difference between the 95% confidence interval limits divided by the referral rate was 2.7.

We conducted a multiple logistic regression analysis with the patient visit as the unit of analysis for examining the independent effect of practice prevalence of the presenting problem on the chances of referral. Controlling variables in this analysis included age, sex, comorbidity index tertile, and the medical versus surgical nature of the principal diagnosis. To determine if the clinical complexity of a visit as assessed by the comorbidity index modified the effect of practice prevalence, we also added interaction terms for these variables to the final model. The goodness of fit of the final model was tested using the Hosmer-Lemeshow statistic, which suggested adequate model fit.¹⁶ The regression analyses used the generalized estimating equation¹⁷ to control for the intraphysician correlation of clustered visits.

Results

The distribution of clinical characteristics is shown in Table 1. The population was somewhat more represented by women (57.5%) than men, and by children/adolescents (33.3%) and seniors (39.4%) than middle-aged persons. The average comorbidity index value was 0.63, but there was considerable variability and the distribution was right skewed (median=0.31; interquartile range=0.05-0.83). The mean EDC practice-prevalence estimate for patient visits was 51 per 1000 primary care visits, but this distribution was also right skewed (median=25; interquartile range=10-106).

In the bivariate analyses, the probability of referral was significantly related to age, sex, practice prevalence of the condition, patient comorbidity, and whether the condition was a surgical problem Table 2. For the 65 condition categories with stable referral rate estimates Figure 1, the correlation between the logarithms of the practice-prevalence ratios and the referral rates was -0.87 (P <.001). Also, this relationship remained after stratifying for the 35 medical (r=-0.82; P <.001) and 24 surgical conditions (r=-0.88; P=.001).

Table 3 shows the multiple logistic regression results modeling the probability of referral during an office visit as a function of the clinical factors. There were no significant differences in the chances of referring patients aged 11 to 64 years once measures of morbidity, such as patient comorbidity and practice prevalence of the principal diagnosis, were controlled. Men were 19% more likely to be referred than women, and medical conditions were 39% less likely to be referred than surgical ones.

Only one of the practice prevalence/comorbidity interaction terms was significantly different from 0: commonly occurring conditions presenting among patients with high levels of comorbidity. This finding implies that the comorbidity has a stronger influence on the chances of referral for patients presenting with common problems than those presenting with less common problems.

Table 4 shows the estimated probabilities of referral based on differences in practice prevalence and comorbidity. These probabilities were obtained from the b coefficients in Table 3. The reference group for the probability estimates is women aged 18 to 44 years with health problems categorized as medical conditions. The chances of referral varied as much as 8-fold based on only the practice prevalence of the principal diagnosis and level of comorbidity.

Discussion

Our results support the hypothesis that the frequency with which patients’ health problems present to primary care physicians (practice prevalence) has a strong inverse relationship with the chances of referral to specialty care. Primary care physicians were more likely to send patients with uncommon problems to specialists and retain those with the most common conditions. This finding highlights the responsible judgment primary care physicians employ in recognizing the boundaries of their scope of practice. Practice prevalence is a defining feature of the primary care–specialty care interface.

Referring patients with uncommon problems to specialists is a rational way to organize medical care. Outcomes are related to the volume of patients managed with a specific condition.¹⁸ Specialists need to care for an adequate number of patients with uncommon problems to maintain clinical competence. Patient self-referral, however, which dilutes the prevalence of health problems presenting to specialists, may result in potentially invasive and expensive diagnostic approaches to patients more appropriately evaluated by primary care physicians.¹⁹

In addition to a condition’s practice prevalence, the number and severity of comorbidities managed during the visit influenced primary care physicians’ decisions to make specialty referrals. Also, we found an interaction effect between high practice prevalence and high levels of comorbidity. In other words, patients with uncommon conditions were commonly referred, regardless of the complexity of other conditions. The chances of referral markedly increased for patients with common conditions when they also presented with co-existing medically complex health problems. Thus, the rare presentations for which specialist assistance is sought may be a result of either the practice prevalence of the presenting problem or the overall complexity of a patient.

Men were more commonly referred than were women, after accounting for differences in the nature of their problems. A possible explanation for this finding is that because women make more office visits over a year than men,²⁰ their probability of referral during any given visit will be lower given roughly equal chances of referral between the 2 groups during the course of a year.

Further Research

We demonstrated that the potential need for surgical interventions was an important predictor of referral. Even after other clinical factors were controlled, medical conditions were 39% less likely to be referred than surgical ones. This is not surprising given that primary care physicians generally perform only minor office-based surgical procedures. But which surgical procedures should be in the scope of practice of primary care physicians? This question deserves further research and could be addressed in part by an analysis that is similar to the one presented here. Common outpatient procedures are candidates for inclusion as primary care services. Secondary considerations include the requirements and expense of necessary equipment, technical personnel, and training. Research that builds epidemiologic profiles of office-based procedures would be helpful in determining how responsibilities should be divided between generalists and specialists for these technical services.

Limitations

Several limitations in our study’s data source warrant consideration. First, the data set of visits provided information on primary care physicians’ referral decisions and did not elucidate whether patients actually received specialty care. Second, the sample was restricted to visits made to generalist physicians, excluding both obstetrician-gynecologists and medical subspecialists who may act as primary care physicians. Third, the NAMCS data set did not include hospital-based physicians, who are known to have higher referral rates than their office-based counterparts.¹³ Fourth, the unit of analysis was the visit rather than the patient. Patients with certain chronic conditions may have higher referral rates than suggested by our data if the measure used is the percentage of persons obtaining specialty care over a year. The advantage of focusing on the visit is that physician referral decisions can be examined rather than specialist use. Fifth, some conditions had lower than expected referral rates (eg, appendicitis had a referral rate of 46%), because the denominator for the referral rates was all visits made to generalists for the condition, which included both new presentations and follow-up visits. Finally, because of data limitations we did not assess the extent to which condition prevalence within an individual physician’s own practice affects his or her referral behavior.

Specialist visits can be initiated by primary care physician referral, patient self-referral, or specialist-to-specialist cross-referral. Although our database did not permit us to examine each of these pathways, other research suggests that primary care physician referral is the predominant route, particularly in health maintenance organizations.¹²

Conclusions

Our findings provide evidence that the boundaries between primary care physicians and specialists are defined in part by prevalence of health problems and the overall complexity of patients. Future research should focus on identifying modifiable characteristics of the physician-patient interaction, physicians, their practices, and the health system that influence referral decisions, after accounting for clinical factors. The appreciation of relevant clinical factors is critical to the fair application of administrative and financial constraints on physicians’ abilities to refer. Managed care plans that penalize physicians for high referral behavior, without adjusting for practice prevalence and comorbidity work, are contrary to the goal of providing quality patient care in the most appropriate settings. With more precise definitions of the clinical determinants of referral for populations, health systems can better gauge generalist and specialist workforce requirements.

Acknowledgments

This work was supported by the Agency for Healthcare Research and Quality grants #R01 and #HS09377. Barbara Starfield inspired this work and provided comments on the manuscript. We also thank Barbara Bartman, Norm Smith, MD, MPH, and Jonathan Weiner MD, MPH, for their review and comments on the manuscript. Mia Kang and Sarah von Schrader provided excellent technical assistance.

Related Resources

• Agency for Healthcare Research and Quality, Primary Care Subdirectory Page—includes research articles on primary care referral patterns and coordination of care among referring physicians and specialists. http://www.ahrq.gov/research/primarix.htm

Primary care physicians make specialty referrals to obtain advice for clinically uncertain diagnostic evaluations or treatment plans, to obtain a specialized service that falls outside their scope of practice, because of patient or third-party requests, or because of a combination of these reasons.¹ The clinical reasons for these referral decisions include characteristics of the presenting health problem, the burden and severity of comorbidities, and patient preferences for various treatments and outcomes.

Previous research has shown that certain features ascribed to morbidities influence the likelihood of specialty referral. The type of diagnosis is the most obvious determinant. In one study,² adults with malignancies were 5 times more likely to be referred than those with respiratory illnesses. Similarly, in the Netherlands, Van Suijlekom-Smit and colleagues³ found more than an 8-fold variation among childhood diagnosis groups in the likelihood of referral. For patients with similar diagnoses, research has found that severe variants are more likely to be referred.^4-7 Specialty referral is also influenced by the array and complexity of comorbid conditions.⁸

The conceptual foundations of primary care provide further insight into how clinical factors may influence referral to specialty care. A defining feature of primary care is the provision of a comprehensive set of services that meets the majority of a population’s health needs.^9,10 Primary care physicians develop greater experience and expertise for health problems with which they are familiar than those that occur less often. It follows that they would seek specialist assistance for uncommon health problems. However, empirical evidence for this effect is currently lacking.

Our goal was to test the hypothesis that the frequency with which a condition is seen by primary care physicians (practice prevalence) influences the likelihood of referral from primary to specialty care. We use the term practice prevalence to mean the frequency of presentation to primary care physicians and to distinguish it from the frequency of occurrence in the community. Also, we examine the impact of other clinical factors on primary care physicians’ referral decisions, including patient age, sex, comorbidities, and the medical versus surgical nature of the target condition’s management.

Methods

Data Source and Study Sample

We used the 1989 to 1994 National Ambulatory Medical Care Surveys (NAMCS) to examine referrals made to specialist physicians during visits with primary care physicians. NAMCS is a nationally representative survey of office-based physician visits in the United States. Each year, a multistage probability sample of nonfederally funded US physicians who are engaged in patient care activities (excluding radiologists, anesthesiologists, and pathologists) is selected from the master files of the American Medical Association and the American Osteopathic Association. For 1 week each selected physician completes a questionnaire for a 20% to 100% systematic sample of patient visits. Details of the survey methodology and the survey instrument are presented elsewhere.¹¹ The distribution of patient age and sex remained consistent over the 6 years of data collection we used.¹² The 1995 to 1998 surveys were not used, because information on referral was not collected. Using the 1994 and 1998 NAMCS, Forrest and Whelan¹³ found that primary care practice patterns did not substantively differ over time. The pooled data set contained 219,830 visits, of which 78,107 (35.5%) were with generalists (self-reported specialty designation was family/general practice, general pediatrics, or general internal medicine).

Clinical Factors

To examine the referral characteristics of different types of conditions, International Classification of Diseases-ninth revision-Clinical Modification (ICD-9-CM) codes were grouped into clinically similar categories that we called expanded diagnosis clusters (EDCs). The EDCs were based on the diagnosis clusters originally developed by Schneeweiss and coworkers.¹⁴ EDCs were assigned to the vast majority of primary care visits in the data set, matching 93.6% of the unique diagnosis codes. EDCs were grouped into clinical domains based on the nature of the problems and the specialty most responsible for the care. We also classified each EDC according to whether the primary treatment approach was medical or surgical.

A practice-prevalence ratio was calculated for each condition. The numerator was the number of visits for the EDC under consideration, and the denominator was the total number of visits in the data set. These ratios were further divided into tertiles (ie, high-, medium-, and low-frequency EDCs.)

To account for the number and severity of comorbid conditions with which patients presented, we assigned a comorbidity index to each visit. The index was based on the aggregated diagnostic groups (ADGs), the building blocks of The Johns Hopkins Adjusted Clinical Groups Case-Mix system.¹⁵ The 32 ADGs represent morbidity groups that contain ICD-9-CM diagnosis codes that are similar with respect to likelihood of persistence, expected need for health care resources, and other clinical criteria. In the NAMCS, up to 3 ICD-9-CM codes were assigned per visit; thus, each visit was assigned 1 to 3 ADGs. A comorbidity index score was obtained by summing ADG-specific resource intensity weights.* Larger weights suggest more complex health problems and greater expected resource intensity. In previous research,¹³ the comorbidity index increased with patient age and distinguished type of primary care delivery site by morbidity burden. The comorbidity index was divided into tertiles representing high, medium, and low levels of comorbidity.

Data Analysis

The unadjusted association of each clinical factor with the probability of referral was evaluated in bivariate analyses using the chi-square statistic. We further analyzed the relationship between a condition’s practice prevalence and its referral rate using scatter plots and Pearson correlation coefficients. Because both the practice prevalence and referral rate variables were right skewed, we used a logarithmic transformation to normalize both. For the scatter plots we excluded conditions with unstable referral rate estimates because of small sample sizes. Specifically, a condition was included if the ratio of the difference between the 95% confidence interval upper and lower limits of the referral rates divided by the condition-specific referral rate was less than 1. For example, multiple sclerosis was excluded because it presented at a rate of just 6 per 10,000 visits, had a referral rate of 7.1%, and the difference between the 95% confidence interval limits divided by the referral rate was 2.7.

We conducted a multiple logistic regression analysis with the patient visit as the unit of analysis for examining the independent effect of practice prevalence of the presenting problem on the chances of referral. Controlling variables in this analysis included age, sex, comorbidity index tertile, and the medical versus surgical nature of the principal diagnosis. To determine if the clinical complexity of a visit as assessed by the comorbidity index modified the effect of practice prevalence, we also added interaction terms for these variables to the final model. The goodness of fit of the final model was tested using the Hosmer-Lemeshow statistic, which suggested adequate model fit.¹⁶ The regression analyses used the generalized estimating equation¹⁷ to control for the intraphysician correlation of clustered visits.

Results

The distribution of clinical characteristics is shown in Table 1. The population was somewhat more represented by women (57.5%) than men, and by children/adolescents (33.3%) and seniors (39.4%) than middle-aged persons. The average comorbidity index value was 0.63, but there was considerable variability and the distribution was right skewed (median=0.31; interquartile range=0.05-0.83). The mean EDC practice-prevalence estimate for patient visits was 51 per 1000 primary care visits, but this distribution was also right skewed (median=25; interquartile range=10-106).

In the bivariate analyses, the probability of referral was significantly related to age, sex, practice prevalence of the condition, patient comorbidity, and whether the condition was a surgical problem Table 2. For the 65 condition categories with stable referral rate estimates Figure 1, the correlation between the logarithms of the practice-prevalence ratios and the referral rates was -0.87 (P <.001). Also, this relationship remained after stratifying for the 35 medical (r=-0.82; P <.001) and 24 surgical conditions (r=-0.88; P=.001).

Table 3 shows the multiple logistic regression results modeling the probability of referral during an office visit as a function of the clinical factors. There were no significant differences in the chances of referring patients aged 11 to 64 years once measures of morbidity, such as patient comorbidity and practice prevalence of the principal diagnosis, were controlled. Men were 19% more likely to be referred than women, and medical conditions were 39% less likely to be referred than surgical ones.

Only one of the practice prevalence/comorbidity interaction terms was significantly different from 0: commonly occurring conditions presenting among patients with high levels of comorbidity. This finding implies that the comorbidity has a stronger influence on the chances of referral for patients presenting with common problems than those presenting with less common problems.

Table 4 shows the estimated probabilities of referral based on differences in practice prevalence and comorbidity. These probabilities were obtained from the b coefficients in Table 3. The reference group for the probability estimates is women aged 18 to 44 years with health problems categorized as medical conditions. The chances of referral varied as much as 8-fold based on only the practice prevalence of the principal diagnosis and level of comorbidity.

Discussion

Our results support the hypothesis that the frequency with which patients’ health problems present to primary care physicians (practice prevalence) has a strong inverse relationship with the chances of referral to specialty care. Primary care physicians were more likely to send patients with uncommon problems to specialists and retain those with the most common conditions. This finding highlights the responsible judgment primary care physicians employ in recognizing the boundaries of their scope of practice. Practice prevalence is a defining feature of the primary care–specialty care interface.

Referring patients with uncommon problems to specialists is a rational way to organize medical care. Outcomes are related to the volume of patients managed with a specific condition.¹⁸ Specialists need to care for an adequate number of patients with uncommon problems to maintain clinical competence. Patient self-referral, however, which dilutes the prevalence of health problems presenting to specialists, may result in potentially invasive and expensive diagnostic approaches to patients more appropriately evaluated by primary care physicians.¹⁹

In addition to a condition’s practice prevalence, the number and severity of comorbidities managed during the visit influenced primary care physicians’ decisions to make specialty referrals. Also, we found an interaction effect between high practice prevalence and high levels of comorbidity. In other words, patients with uncommon conditions were commonly referred, regardless of the complexity of other conditions. The chances of referral markedly increased for patients with common conditions when they also presented with co-existing medically complex health problems. Thus, the rare presentations for which specialist assistance is sought may be a result of either the practice prevalence of the presenting problem or the overall complexity of a patient.

Men were more commonly referred than were women, after accounting for differences in the nature of their problems. A possible explanation for this finding is that because women make more office visits over a year than men,²⁰ their probability of referral during any given visit will be lower given roughly equal chances of referral between the 2 groups during the course of a year.

Further Research

We demonstrated that the potential need for surgical interventions was an important predictor of referral. Even after other clinical factors were controlled, medical conditions were 39% less likely to be referred than surgical ones. This is not surprising given that primary care physicians generally perform only minor office-based surgical procedures. But which surgical procedures should be in the scope of practice of primary care physicians? This question deserves further research and could be addressed in part by an analysis that is similar to the one presented here. Common outpatient procedures are candidates for inclusion as primary care services. Secondary considerations include the requirements and expense of necessary equipment, technical personnel, and training. Research that builds epidemiologic profiles of office-based procedures would be helpful in determining how responsibilities should be divided between generalists and specialists for these technical services.

Limitations

Several limitations in our study’s data source warrant consideration. First, the data set of visits provided information on primary care physicians’ referral decisions and did not elucidate whether patients actually received specialty care. Second, the sample was restricted to visits made to generalist physicians, excluding both obstetrician-gynecologists and medical subspecialists who may act as primary care physicians. Third, the NAMCS data set did not include hospital-based physicians, who are known to have higher referral rates than their office-based counterparts.¹³ Fourth, the unit of analysis was the visit rather than the patient. Patients with certain chronic conditions may have higher referral rates than suggested by our data if the measure used is the percentage of persons obtaining specialty care over a year. The advantage of focusing on the visit is that physician referral decisions can be examined rather than specialist use. Fifth, some conditions had lower than expected referral rates (eg, appendicitis had a referral rate of 46%), because the denominator for the referral rates was all visits made to generalists for the condition, which included both new presentations and follow-up visits. Finally, because of data limitations we did not assess the extent to which condition prevalence within an individual physician’s own practice affects his or her referral behavior.

Specialist visits can be initiated by primary care physician referral, patient self-referral, or specialist-to-specialist cross-referral. Although our database did not permit us to examine each of these pathways, other research suggests that primary care physician referral is the predominant route, particularly in health maintenance organizations.¹²

Conclusions

Our findings provide evidence that the boundaries between primary care physicians and specialists are defined in part by prevalence of health problems and the overall complexity of patients. Future research should focus on identifying modifiable characteristics of the physician-patient interaction, physicians, their practices, and the health system that influence referral decisions, after accounting for clinical factors. The appreciation of relevant clinical factors is critical to the fair application of administrative and financial constraints on physicians’ abilities to refer. Managed care plans that penalize physicians for high referral behavior, without adjusting for practice prevalence and comorbidity work, are contrary to the goal of providing quality patient care in the most appropriate settings. With more precise definitions of the clinical determinants of referral for populations, health systems can better gauge generalist and specialist workforce requirements.

Acknowledgments

This work was supported by the Agency for Healthcare Research and Quality grants #R01 and #HS09377. Barbara Starfield inspired this work and provided comments on the manuscript. We also thank Barbara Bartman, Norm Smith, MD, MPH, and Jonathan Weiner MD, MPH, for their review and comments on the manuscript. Mia Kang and Sarah von Schrader provided excellent technical assistance.

Related Resources

• Agency for Healthcare Research and Quality, Primary Care Subdirectory Page—includes research articles on primary care referral patterns and coordination of care among referring physicians and specialists. http://www.ahrq.gov/research/primarix.htm