To investigate whether cancer patients who sought a second opinion received better medical care.

 A total of 1358 newly diagnosed colorectal cancer patients undergoing resection were identified from Taiwan’s National Health Insurance Research Database between 2004 and 2008. The frequency of doctor shopping and hospital shopping in the 6 months before resection was used to define “seeking a second opinion.”

 A generalized hierarchical linear model was used to determine the influence of doctor shopping and hospital shopping on in-hospital complications and prolonged hospitalization after colorectal resection.

The risk of in-hospital complications for heavy doctor shoppers was significantly higher than that for patients who were not doctor shoppers (odds ratio [OR] = 1.675, P = .037). However, the risk was significantly lower for heavy hospital shoppers compared with those who were not hospital shoppers (OR = 0.272, 

 = .007). The frequency of doctor shopping and hospital shopping was not significantly associated with prolonged hospitalization.

For colorectal resection patients, the selection of a proper hospital for surgery resulted in better surgical care. The quality of surgical care was worse with heavy doctor shopping. We suggest that healthcare authorities disclose data about the quality of a hospital’s cancer treatment to increase patient access to such information. This may help patients find quality healthcare providers more quickly and reduce the waste of medical resources resulting from the long process of seeking medical care.

Seeking second opinions increases the cost of medical care. However, seeking second opinions might be recommended for cancer patients due to the variety of cancer treatments.

		Among colorectal resection patients in Taiwan, those who were heavy hospital shoppers had a lower risk of in-hospital complications, whereas those who were heavy doctorshoppers had a higher risk of in-hospital complications.

		Prolonged hospitalization was not significantly affected by frequencies of hospital shopping and doctor shopping.

		Information on quality of cancer care should be disclosed to help patients shorten their process of seeking medical care.

The development of public insurance and the availability of media for medical information have reduced the obstacles that hinder patients from seeking medical care and raised patients’ consumer awareness.

 With the ability to choose a preferred treatment, patients can seek a second opinion for the same illness to receive better quality of medical care.

 Unlike the 1-way medical care of the past, patients can now not only spontaneously seek the medical care that they need, but also change their healthcare provider if they are dissatisfied with their first provider.

However, in view of increasing medical expenses, many researchers have examined hospital shopping or doctor shopping to test the hypothesis that seeking a second opinion could cause a financial burden on the medical care system.

Seeking a second opinion is defined as seeking medical care from more than 1 physician for the same condition. Compared with the judgment made by only 1 physician, a diagnosis made by multiple doctors is considered less likely to be inaccurate and can reduce a patient’s anxiety. 

 Most previous studies have evaluated the seeking of a second opinion by patients with specific conditions such as upper respiratory diseases, pulmonary tuberculosis, cancers, neurologic disorders, or polypharmacy. 

 A patient’s targets for comparison include both physicians and hospitals. Factors influencing a patient’s choice of a hospital include reputation, service volume, location, cost, and the patient’s own previous experiences.

 Factors important in the choice of a physician include the physician’s attitude, treatment recommendations, and patient satisfaction with  that physician’s care.

 Patients select a preferred care provider after a thorough comparison of all these factors.

There are 3 broad areas in the research on seeking a second opinion. One area of research has been to study the  characteristics of patients who are on the receiving end of poor communication.

 A second area involves investigation into the motives for seeking a second opinion.

 A few studies have evaluated the results after seeking a second opinion.

 Some earlier studies found that improperly seeking a second opinion could waste medical resources and should be avoided.

 In addition, patients tended to feel even more lost if the second physician made the same diagnosis. Conversely, some researchers suggested that when there are diagnostic differences between the first and second opinions,the second opinion could be helpful for providing proper medical care and avoiding excessive or insufficient treatment.

Due to the nature of their disease and the potential harm from toxic treatment modalities, cancer patients are more likely to seek a second opinion. According to studies, nearly 60% of cancer patients have sought a second opinion.

 A diagnosis of cancer is a great shock to a patient, and cancer patients often have many questions about their diagnosis and treatment. If they do not obtain satisfactory answers from the first physician, patients go to other doctors for reassurance about their disease.

 In some instances, seeking a second opinion is beneficial for cancer patients. Reassurance from another physician could increase a patient’s confidence in the treatment. Furthermore, there may be differences in testing methods, medications, therapeutic techniques, and medical teams in different hospitals. During the process of seeking a second opinion, patients can choose a medical technology that is newer, more effective, or more suitable for them.

 Because of patient psychological factors and the development of therapeutic technologies, perhaps second opinions should be advised for cancer patients.

Although seeking a second opinion is increasingly common for cancer patients, little is known about the relationship between seeking a second opinion and the outcomes of medical care. Therefore, we investigated whether cancer patients who sought second opinions received better medical care. Through the process of hospital or surgeon selection, we assumed that patients had more chances to choose a better medical team with better therapeutic skills. Therefore, we inferred that seeking a second opinion could be helpful in getting better care. Using colorectal resection patients as an example, the aim of this study was to determine the relationship between seeking a second opinion and therapeutic outcome.

Study data came from Taiwan’s National Health Insurance Research Database (NHIRD), which was provided by the Bureau of National Health Insurance in Taiwan for research purposes. The National Health Insurance program in Taiwan contains data on  more than 99% of the entire population of 23.74 million people. The NHIRD is a national and representative database that contains comprehensive claim records of outpatient, inpatient, and emergency care. These data were cross-checked and validated to ensure accuracy.26,27 We used the database for a sampled cohort of 1 million people from 2004 to 2008. We also used admission files from the population-based inpatient expenditures database from 2004 to 2005.

From the hospitalization claims data of 1 million randomly chosen patients in NHIRD from July 2004 to November 2008, we identified patients who were newly diagnosed with colorectal cancer as a primary or secondary diagnosis 6 months before surgery 

(International Classification of Diseases, Ninth Revision, Clinical Modification

] codes 153.XX or 154.XX). We excluded patients with anal tumors (

 code 154.2) because surgical resection is not usually considered first-line treatment for this disease. For patients who underwent multiple surgical resections, data from only the first operation were analyzed.

We used doctor shopping and hospital shopping in the 6 months before surgery as second opinion—seeking behaviors. The number of physicians or hospitals that patients consulted about their colorectal cancer was calculated after the operating physician or hospital was excluded. Patients who shopped for 0, 1, and 2 or more physicians were classified as non–doctor shoppers, light doctor shoppers, and heavy doctor shoppers, respectively. Hospital shoppers were classified using the same criteria. Sex, age, Charlson Comorbidity Index (CCI) score, urbanization level, site of the tumor, use of preoperative screening colonoscopy, and use of preoperative chemotherapy were controlled for in the analyses. The urbanization level, which was developed by the Taiwan National Health Research Institute, was stratified into 7 classifications ( I, II, III, IV, and 

V)and ranged from greater to lower degrees of urbanization in the analyses.

Hospital characteristics obtained from the population-based inpatient expenditures database included in the analysis were accreditation level (academic medical center, regional, district), ownership (public, private), and average annual hospital volume for colorectal cancer surgery in 2004 and 2005.

Based on the study objective and the limitations of secondary data, we used in-hospital complications and prolonged  hospitalization as the indicators of surgical outcome. Mortality has been used to measure the quality of operations, and many studies have used mortality as a quality indicator following colorectal resection.

 The low number of deaths during a short surgical period, however, may result in a poor level of statistical confidence. Zheng and colleagues

 suggested overcoming this problem by improving the selection of indicators, extending the study period, or enlarging the sample size. Therefore, we used in-hospital complications and prolonged hospitalization, which are intermediate clinical indicators, as the outcome measures. In-hospital complications were defined as infections and cardiovascular, respiratory, gastrointestinal, urologic, or other complications.

 Patients who developed any of these conditions were classified as having in-hospital complications. To avoid overestimating the incidence of complications, patients presenting with comorbidities 1 month before surgery were excluded. Not all complications are the result of bad care or lower technical skills, and the correlation between comorbidities and complications has been observed in colorectal cancer patients.

 Therefore, prognostic factors, including patient characteristics and comorbid conditions, were controlled for in our analysis to reduce estimation bias. Furthermore, prolonged hospitalization was defined as a length of stay of more than 14 days.

 test (or Fisher exact test, as appropriate) was used to explore the relationship between seeking a second opinion and outcomes of surgical care. In addition, data in the present study had a hierarchical structure, with patients nested within physicians and physicians nested within hospitals because the clustering effect and the dependence of observations within groups made the traditional regression model unsuitable. Considering the same size and that the dependent variables were binary outcomes, the 2-level hierarchical generalized linear model was applied (level 1 was the patient level; level 2 was the hospital level). The analyses were performed using the SPSS 17.0 (SPSS Inc, Chicago, Illinois) and HLM 6.02 (Scientific  Software International, Inc, Skokie, IL) software packages. Significance was defined as a 2-sided 

Along with the hierarchical linear model, a null model with no predictor variables at any level was first used to verify suitability. A null model can measure the magnitude of variation in an outcome measure across the different levels. The proportion of the variance across different levels can be expressed as an inter-class correlation coefficient (ICC). For the 2-level dichotomous  outcome model, the ICC between level 2 and the total variation was represented as σ2/(σ2 + π2/3).

After confirming the existence of significant variances among groups, further mean-as-outcomes models were applied to predict the intercept in level 1.

The study subjects included 1358 patients from 77 hospitals.Patient characteristics and surgical outcomes are shown in 

. There were 767 (56.5%), 434 (32.0%), and 157 (11.6%) non—doctor shoppers, light doctor shoppers, and heavy doctor shoppers, respectively. The non–hospital shoppers, light hospital shoppers, and heavy hospital shoppers totalled 978 (72.0%), 334 (24.6%), and 46 (3.4%), respectively. There were 77 hospitals where patients underwent surgery. Among hospitals, 24.7%, 61.0%, and 14.3% were academic medical centers, regional hospitals, and district hospitals, respectively. There were 22 (28.6%) public hospitals and 55 (71.4%) private hospitals. The mean annual hospital volume for colorectal surgery was 99.3 (±128.8) patients. Complications during hospitalization occurred in 255 (18.8%) of patients, and prolonged hospitalization occurred in 631 (47.3%) of the 1333 patients who survived at least 14 days after resection.

 shows the relationships between seeking a second opinion and surgical outcomes. The incidence of in-hospital complications was 19.6%, 17.1%, and 19.7% for non—doctor shoppers, light doctor shoppers, and heavy doctor shoppers,respectively. The incidence of prolonged hospitalization was 48.8%, 46.7%, and 41.9% for non–doctor shoppers, light doctor shoppers, and heavy doctor shoppers, respectively. The frequency of doctor shopping was not significantly associated with surgical outcomes after a bivariate analysis. The incidence of in-hospital complications was 19.3%, 18.6%, and 8.2% for non–hospital shoppers, light hospital shoppers, and heavy hospital shoppers, respectively. Heavy shoppers had the lowest incidence of in-hospital complications, but this difference was not statistically significant. The incidence of prolonged hospitalization was 49.3%, 42.4%, and 41.3% for non–hospital shoppers, light hospital shoppers, and heavy hospital shoppers, respectively. Non–hospital shoppers had the highest incidence of prolonged hospitalization, but this difference was not statistically significant.

The results of the null models of hierarchical linear models demonstrated a significant variation in surgical outcomes among hospitals. The ICC of patients with the same hospital was 5.27% [0.114/(0.183 + 3.289); 

 = .003] for in-hospital complications, and the ICC of patients with the same hospital was 19.37% [0.790/(0.790 + 3.289); 

 <.001] for prolonged hospitalization. The variance in the null model was significant in level 2, suggesting that there were significant differences in the incidence of in-hospital complications and prolonged hospitalization among patients of different hospitals.This result indicates that it is suitable to use a generalized hierarchical linear model and further mean-as-outcomesmodels. In further models, patient-level predictors were included in level 1, resulting in this equation:

. Further, hospital predictors are included in level 2, resulting in equations of π

. Age, CCI, and hospital volume were continuous variables. Therefore, we entered them in the model by subtracting them from the grand mean.

 shows the results of mean-as-outcomes models of hierarchical generalized linear models. The initial incidence of in-hospital complications was 21.32% (0.271/[1 + 0.271]). The risk of in-hospital complications was significantly higher for men  than for women (odds ratio [OR] = 1.333; 

 = .003). The risk also increased significantly in older patients (OR = 1.023; 

 <.001) and those with higher CCI scores (OR =1.076; 

 = .004), and the risk of in-hospital complications was significantly lower for those who received the preoperative screening colonoscopy (OR = 0.648; 

 = .006). After adjusting for other characteristics,  heavy doctor shoppers were at a higher risk of in-hospital complications than non—doctor shoppers (OR = 1.675; 

=.037). Conversely, heavy hospital shoppers had a lower risk of in-hospital complications than non—hospital shoppers (OR = 0.272; 

 = .007). The average incidence of prolonged hospitalization was estimated as 41.28% (0.703/[1 +0.703]) among hospitals. The risk of prolonged hospitalization also increased significantly in older patients (OR = 1.028;

 <.001) and those with higher CCI scores (OR =1.059; 

= .001). However, the frequencies of hospital shopping and doctor shopping were not significant predictors of prolonged hospitalization after adjusting for other characteristics.

We found that the factors that influenced the incidence of in-hospital complications for colorectal resection patients were sex, age, CCI score, the screening colonoscopy, doctor shopping frequency, and hospital shopping frequency. After controlling for patient and hospital characteristics, heavy doctor shoppers were at a higher risk of in-hospital complications than non—doctor shoppers. Conversely, heavy hospital shoppers had a lower risk of in-hospital complications than non–hospital shoppers. In addition, prolonged hospitalization was significantly influenced by age, CCI score, and whether a patient had the preoperative screening colonoscopy. Nevertheless, after controlling for patient and hospital characteristics, the frequencies of doctor shopping and hospital shopping were not significantly associated with prolonged hospitalization.

With regard to the risk of in-hospital complications, we found that heavy doctor shoppers were at a higher risk than non—doctor shoppers, which suggests that the quality of care was not enhanced for patients who were heavy doctor shoppers. Previous studies have demonstrated that a patient’s choice of a physician is influenced by the professional knowledge, skills, and attitude of the physician. Patients can obtain more accurate therapeutic information after seeking a second opinion, which leads to a higher level of satisfaction with their treatment.

Our finding is not in accordance with the results of several previous studies, although no previous study has discussed the effect of doctor shopping on treatment outcomes in detail. This finding was not surprising, as it confirmed what previous researchers have discovered about the generally poor health status of frequent doctor shoppers.

 A possible reason for heavy doctor shoppers being at higher risk for in-hospital complications than non—doctor shoppers is that a longer delay from diagnosis to surgery negatively affects surgical outcomes. Some studies have indicated that longer delays for cancer surgery potentially influence prognosis.

Therefore, we inferred that the behaviors of heavy doctor shoppers might increase the delay from diagnosis to surgery, which may result in worse surgical outcomes.

Another important finding in this study was that hospital shopping was significantly correlated with in-hospital complications. Heavy hospital shoppers had a lower risk of in-hospital complications than non—hospital shoppers. Previous studies have shown that hospital quality is 1 of the reasons why patients select a hospital for surgery. Patients choose a high-quality hospital based on a physician’s recommendation, hospital reputation, or quality reports.

 In the present study, the effect of heavy  hospital shopping on the incidence of in-hospital complications was not in favor of heavy doctor shopping. There may be more variation in the quality of hospitals than in the quality of physicians. Developing medical technologies have made medical care more complex, and complicated patients need professional teamwork for successful treatment. Other than physician skills, the quality of care is also correlated with hospital equipment and the abilities of medical professionals.

 Cancer care requires the coordination of physicians, nurses, medical laboratory scientists, radiologists, and nutrition support staff, and several studies have reported that surgical outcomes are associated with professional staffing.

 In addition, quality-of-care problems that lead to surgical complications include problems with technical care during a surgical procedure, problems with medications administered, failure to provide preventive care, and failure to monitor a patient’s condition or medications.

 Preventing complications was associated not only with the skills of physicians and support staff, but also with the processes of care. The   superior technique of medical professionals and superior management of a hospital can prevent and detect complications at an early stage. By receiving treatment in capable hospitals, patients can obtain better surgical care and higher-quality therapy.

In contrast, doctor shopping and hospital shopping conferred no significant effect on prolonged hospitalization, although they did significantly affect the risk of in-hospital complications. A possible explanation for this finding may be that prolonged hospitalization is both a medical indicator and a managerial indicator. In the present study, prolonged hospitalization, a measure for quality of care, conflicted somewhat with the pressures to decrease health utilization.

 Therefore, we suggest that other outcome measures (eg, recurrence, readmission rate) be considered to better examine quality of care.

There are several limitations in this study. First, we did not have information on cancer staging, which is an important factor that impacts surgical outcomes and is also an influential factor for patients seeking a second opinion. Nevertheless, the use of  screening colonoscopy, which can effectively identify cancers in early stages, was controlled for in these analyses. Second, we used administrative data to determine the incidence of in-hospital complications. However, the accuracy of administrative data remains in doubt, as highlighted in previous studies.

In the present study, the complication rate might be underestimated due to coding inaccuracies. Furthermore, all complications do not result from bad care. Although patient characteristics were controlled for in the present study, the use of complications as an intermediate outcome indicator is still limited. Conversely, using postoperative unplanned procedures as a complication marker may facilitate future outcome studies.

This is the first study, to our knowledge, to test the effects of seeking a second opinion on surgical outcomes in cancer patients. We found that colorectal cancer patients can obtain a higher quality of surgical care by selecting a hospital for their operation. However, heavy doctor shopping may be unfavorable to the quality of surgical care. We inferred that hospital shopping may be beneficial for patient care. Therefore, we suggest that healthcare authorities increase patient access to information about the quality of cancer care in various hospitals. This information may help patients shorten the process of finding quality healthcare providers; at the same time, it may help reduce the waste of medical resources that result from the long process of seeking care.

 Long-term Care Insurance Preparatory Task Force (H-RC), Department of Health, Executive Yuen, Taiwan; Institute of HealthPolicy and Management (H-RC, M-CY, K-PC), National Taiwan University, Taipei, Taiwan.

 The authors (H-RC, M-CY, K-PC) report no relationship or financial interest with any entity that would pose a conflict of interest with the subject matter of this article.

Concept and design (H-RC, M-CY, K-PC); acquisition of data (H-RC); analysis and interpretation of data (H-RC, K-PC);drafting of the manuscript (H-RC, K-PC); critical revision of the manuscript for important intellectual content (K-PC); statistical analysis (H-RC, K-PC); provision of study materials or patients (H-RC); obtaining funding (H-RC);administrative, technical, or logistic support (M-CY,); and supervision (MCY, K-PC).

: Kuo-Piao Chung, PhD, No 17, Xu-Zhou Road, Taipei, Taiwan 100, Taiwan. E-mail: kpchung@ntu.edu.tw.1. Harris KM. How do patients choose physicians? evidence from a national survey of enrollees in employment-related health plans. Health Serv Res. 2003;38(2):711-732.

2. Sato T, Takeichi M, Hara T, Koizumi S. Second opinion behaviour among Japanese primary care patients. Br J Gen Pract. 1999;49(444): 546-550.

3. Marquis MS, Davies AR, Ware JE Jr. Patient satisfaction and change in medical-care provider: a longitudinal study. Med Care. 1983;21(8): 821-829.

4. Hagihara A, Tarumi K, Odamaki M, Nobutomo K. A signal detection approach to patient-doctor communication and doctor-shopping behaviour among Japanese patients. J Eval Clin Pract. 2005;11(6): 556-567.

5. Wang MJ, Lin SP. Study on doctor shopping behavior: insight from patients with upper respiratory tract infection in Taiwan. Health Policy. 2010;94(1):61-67.

6. Rosenberg SN, Gorman SA, Snitzer S, Herbst EV, Lynne D. Patient reactions and physician-patient communication in a mandatory surgical second-opinion program. Med Care. 1989;27(5):466-477.

Articles

MANAGERIAL

Colorectal resection patients in Taiwan with heavy hospital shopping behaviors got better surgical care than those who did not shop or hospitals.

Can Cancer Patients Seeking a Second Opinion Get Better Care?

To estimate the direct medical costs associated with type 2 diabetes, its complications, and its comorbidities among US managed carepatients.

Data were from patient surveys, chart reviews, and health insurance claims for 7109 people with type 2 diabetes from 8 health plans participating in the Translating Research Into Action for Diabetes (TRIAD) study between 1999 and 2002.

 A generalized linear regression model was developed to estimate the association of patients’ demographic characteristics, tobaccouse status, treatments, related complications, andcomorbidities with medical costs.

The mean annualized direct medical cost was $2465 for a white man with type 2 diabetes who had been diagnosed fewer than 15 yearsearlier, was treated with oral medication or dietalone, and had no complications or comorbidities. We found annualized medical costs to be 10% to 50% higher for women and for patients whose diabetes had been diagnosed 15 or moreyears earlier, who used tobacco, who were being treated with insulin, or who had several other complications. Coronary heart disease, congestive heart failure, hemiplegia, and amputation were each associated with 70% to 150% higher costs. Costs were approximately 300% higher for end-stage renal disease treated with dialysis and approximately 500% higher for end-stage renal disease with kidney transplantation.

 Most medical costs incurred by patients with type 2 diabetes are related to complications and comorbidities. Our cost estimatescan help when determining the most cost-effectiveinterventions to prevent complications and comorbidities.

Because currently available simulation models often do not differentiate costs by patient characteristics, our study has several advantages over previous ones.

		Data on cost and patient characteristics were from both patient surveys and chart reviews.

		Variations in cost associated with different patient characteristics and treatments were assessed in a large and demographically and geographically diverse sample of adults with diabetes.

		Economic researchers can use these estimates in simulation models to assess the potential cost-effectiveness of interventions intended to prevent or delay type 2 diabetes and its complications.

The worldwide prevalence of type 2 diabetes and the demand for and costs of medical care for treating it have increased over the past decade.

Simulation models have been developedto estimate the long-term health and economic consequences of diabetes and to help policy makers identify the most cost-effective interventions for preventing and controlling diabetes.

 The cost data needed for diabetes cost-effectiveness models should be accurate and broadly applicable, come from large samples of patients with type 2 diabetes, and account for variations in costs due to differences in treatments, demographic characteristics, complications, and comorbidities among patients.

Many previous cost studies have estimated the economic cost of diabetes for a country or region as an aggregate, rather than based on individual-level variation.

 Although others have used data from a single health plan, the resulting estimates might not be widely comparable to those from other settings

 Most studies that estimated the costs of type 2 diabetes at the individual level did not consider the variation of the costs among patients with different characteristics.

 Other studies have used self-reported healthcare costs,

 which might have been inaccurate due to incomplete recall or potential bias from the influence of social desirability. Still others have obtained components of diabetes costs from multiple data sources, each of which might have had its own biases, thus limiting the comparability of the estimates.

The purpose of this study was to use a large and demographically and geographically diverse sample of adults with diabetes in the United States to provide cost data for diabetes simulation models. We described  the relationship between direct medical costs and individual patient demographic characteristics, tobacco use status, diabetes treatments, complications, and comorbidities in persons with type 2 diabetes.

We analyzed patient surveys, medical records, and administrative data for 7109 patients with type 2 diabetes who participated in the Translating Research Into Action for Diabetes (TRIAD) study between 1999 and 2002.13 TRIAD was designed to assess how the organization and structure of managed care health plans influence the processes and outcomes of diabetes care. The study involved 10 health plans and 68 provider groups serving approximately 180,000 persons with diabetes across the United States. TRIAD participants—who had to be 18 years or older, community dwelling, English or Spanish speaking, continuously enrolled in the same health plan for at least 18 months, not pregnant, and with more than 1 claim for health services&mdash;were sampled from the 10 health plans. We classified patients as having type 2 diabetes if the onset was before 30 years of age without current insulin treatment or if the onset was after 30 years of age with or without current insulin treatment. A total of 11,927 people from the initial sample met these criteria.

Of the 8820 patients whose medical records were abstracted at baseline, 8364 had Charlson Comorbidity Index scores. (The Charlson Comorbidity Index predicts the 10-year mortality for a patient who may have a range of comorbid conditions.) Of the 10  participating health plans, 2 (1255 patients) were excluded due to unavailability of some elements of the health plan administrative data, leaving 7109 patients for our analyses.

TRIAD collected baseline and follow-up data from health plans, provider groups, and diabetes patients. For this analysis, we used patient survey data gathered from the third quarter of 2000 to the first quarter of 2002, and chart review data and administrative data for the 18 months prior to each respondent’s survey date.

Direct medical costs from the 18 months prior to each patient’s baseline survey were determined from health plan administrative data; costs included inpatient, outpatient, emergency department treatment, pharmacy, and other expenses such as outpatient radiology and laboratory tests. We attempted to minimize price variations due to different labor, nonlabor, and other costs for the same services provided in different health plans. Inpatient costs were calculated based on the patient’s final diagnosis using the group weight rate from the Centers for Medicare & Medicaid Services (CMS) for fiscal year (FY) 2002 multiplied by the FY 2000 multiplier ($4328). Outpatient costs represented the estimated costs for procedures based on standardized reimbursement rates developed from the FY 2002 Medicare fee schedules by procedure code. Outpatientpharmaceutical costs were based on average wholesale prices per unit. We used the actual cost in dollars as described above without adjusting it to a single year’s cost index because the way health plans reported the data did not allow  us to do so. However, we adjusted for the survey interview year in the regression model to attenuate the inflation effect. We also reported the costs in 2010 dollars using the Consumer Price Index for medical services to reflect inflation in treating diabetes over the past decade.Patient copayments and other out-of-pocket costs were not considered in our analyses. Thus, our analyses represent direct medical costs from the perspective of a large health system.

Patient characteristics (eg, age, sex, race and ethnicity, education, income, tobacco use) were determined from patient surveys. Diabetes-related variables, including time since diagnosis of diabetes and methods of treatment,  were also determined from patient surveys. Diabetic complications and comorbidities—including retinopathy, nephropathy, neuropathy, hypertension, dyslipidemia, cerebrovascular disease, cardiovascular disease,  and peripheral vascular disease&mdash;were determined from both patient surveys and medical record reviews. A  comorbid condition was considered present if the 18-month medical record or chart review showed that the patient had the condition or if, in the patient survey, the patient recounted being told that in the past 18 months he or she had had the condition.  Additional comorbid conditions that are components of the Charlson Comorbidity Index were determined from medical record reviews.

We divided 18-month cost figures by 1.5 to standardize them as annual cost amounts. Cost distributions were right skewed; for regression analysis, to account for the skewed distribution we developed a generalized linear model (GLM) with log-link function to estimate the association between costs and patient demographic characteristics, tobacco use status, diabetes treatments, complications, and comorbidities. Health plan indicators and the year of the interview were also included in the calculations to control for health plan fixed effects and cost inflation, respectively. As the GLM with log-link function regression model required original cost data to be log-transformed, coefficients and 95% confidence intervals from the regression model needed to be back-transformed to the ordinal scale using an exponential function to get the cost estimates. We called these back-transformed coefficients cost multipliers in this study.

For this model, the base case was determined to be the 1-year direct medical costs for a white man diagnosed with diabetes for fewer than 15 years, treated with diet or oral agents, and with no complications or comorbidities. Because the costs for  treating such a person differed among health plans and we did not know which health plan represented the “true” cost, it was not appropriate to use a cost estimate from any of the health plans as the base-case cost. We decided to use the mean of the estimated base-case costs among all of the health plans. To do this, we included all of the indicators for each health plan in the model and omitted the intercept to get the mean base-case cost in each of the 8 health plans, then computed the mean of the estimated mean costs in the health plans. That provided a modeled mean cost to use as the uniform direct medical cost for a  base-case patient.

Because the model had a log-link function, it was a multiplicative model. To determine the relative increase in direct medical costs for a given patient with characteristics other than those of a base-case patient, we multiplied the direct medical costs for a base-case patient by the product of the cost multipliers calculated for each demographic characteristic, tobacco use status, diabetes treatment, complication, or comorbidity that applied to that patient.

All of the independent variables were coded as dichotomous or discrete variables. In the regression analysis, missing values for independent variables were imputed 5 times using a multiple imputation method.

We did not impute missing dependent variables. Because the purpose of our model was cost prediction, we used stepwise regression and only variables with regression coefficients significant at the 

.05 level were kept in the model. For the same group of covariates at different levels, we collapsed the groups that were not statistically significant into 1 level. When several independent variables were highly correlated with each other (correlation coefficient 

.25), only 1 was included in the model. For example, income and education are highly correlated, so income was deleted from the model. We did not consider interaction effects for our analyses. We used SAS version 9.1.3 (SAS Institute, Cary, North Carolina) and STATA version 10.1 (StataCorp, College Station, Texas) to perform the analyses.

 shows the unadjusted means and standard deviations of the aggregate costs for treatments, complications, and comorbidities associated with the demographic characteristics of the 7109 patients in the study. Hispanic, non-Hispanic black, and Asian people made up almost half of the sample. More than half of the patients were 60 years or older. Women made up 54% of the sample. Unadjusted costs differed depending on patient demographic and socioeconomic characteristics.

 shows the unadjusted means and standard deviations of the aggregate costs for treatments, complications, and comorbidities associated with diabetes. One-third of the patients were treated with insulin. The most common comorbidities or complications were hypertension, dyslipidemia, retinopathy, neuropathy, and cardiovascular disease. Other comorbidities, assessed as components of the Charlson Comorbidity Index, were present in 4% to 13% of patients. Unadjusted costs differed depending on time since diagnosis of diabetes, treatment type, complications, and comorbidities. Unadjusted costs also varied substantially across health plans, ranging from $1900 to $2700.

Regression coefficients and cost multipliers associated with tobacco use status and each patient’s demographic characteristics, diabetes treatment, complications, and comorbidities are shown in 

. Using the model, the mean annualized direct medical cost for a white man diagnosed with type 2 diabetes for fewer than 15 years, treated with oral antidiabetic medication or diet alone, and with no complications or comorbidities was calculated as $2465. Cost multipliers for other characteristics, treatments, complications, and comorbidities were calculated based on differences in costs associated with that characteristic compared with those of the base-case patient. Being Hispanic, non-Hispanic black, Asian, or of other race/ethnicity was associated with a 20% to 30% lower direct medical cost compared with being a non-Hispanic white. Being awoman, having diagnosed diabetes for 15 or more years, using tobacco, and having proliferative diabetic retinopathy, diabetic nephropathy, neuropathy, treated hypertension, dyslipidemia, transient ischemic attack, cerebral vascular accident, angina, or peripheral vascular disease were each associated with direct medical costs 10% to 50% higher than those of the base-case patient. Coronary heart disease, congestive heart failure, hemiplegia, and amputation were each associated with 70% to 150% higher direct medical costs. End-stage renal disease (ESRD) treated with dialysis or kidney transplantation was associated with approximately 300% and 500% higher costs, respectively.

For patients with more than 1 characteristic distinguishing them from the base case, total direct medical costs would likely be many times higher than those for the base-case patient. For example, for a woman who had type 2 diabetes for 15 or more years, with an amputation, hypertension, stroke, and congestive heart failure, direct medical costs would be $24,897, 10.1 times those for the base-case patient. This was calculated as $2465 (the basecase cost) X 10.1 (the product of the multipliers for being female [1.10], diabetes duration 

15 years [1.20], amputation [2.38], hypertension [1.13], stroke [1.34], and congestive heart failure [2.13].

We estimated the association between patient demographic characteristics, tobacco use status, diabetes treatments,  complications,and comorbidities and the direct medical costs of  type 2 diabetes across 8 managed care organizations in the United States from 1999 to 2002. Costs varied among patients with different characteristics. Being female, having a longer time since diagnosis of diabetes, and having complications and comorbidities were associated with higher costs. The most costly complication was ESRD treated with dialysis or by kidney transplantation. Direct medical costs for patients with ESRD and dialysis or transplantation were approximately 3 to 5 times higher than those for patients with no or early-stage renal complications (microalbuminuria).

Relative to direct medical costs for whites, those for Hispanics, non-Hispanic blacks, and Asians were lower. Since the cost was standardized, the cost difference among race/ethnic groups reflected differences in  utilization. However, for the current analysis, we were unable to determine whether these differences were due to different treatments received by different racial and ethnic groups or whether they reflected different needs for healthcare among racial  and ethnic minority groups relative to non-Hispanic whites.

Models used for assessing the cost-effectiveness of diabetes interventions often aim to simulate disease progression for a diverse diabetes population and specific diabetic subpopulations, and assign incremental costs to specific diabetes treatment regimens or to the development of diabetes-related complications and comorbidities. Currently available models often do not  differentiate among the costs incurred by patients with different characteristics. Our study has several relative strengths in providing the cost data required for such computer simulation models. First, the information on costs came from administrative  records, and data on the prevalence of comorbidities and complications were obtained from both patient surveys and chart  reviews, thus potentially reducing both recall and social desirability biases compared with studies that used only surveys to gather information.

 Second, cost data related to complications and comorbidities were collected for the same study participants, thus avoiding overlap or discrepancies that can be introduced when using cost components from different data sets. For example, in 1 prior study, different sources of information (eg, literature reviews, acute care discharge databases, government reports, fee schedules) were used to assess different costs, which could introduce inconsistencies to aggregate cost estimates.

 Third, our study assessed variation in costs associated with different patient characteristics and treatments, which is an advance over prior research that aggregated costs across large patient groups with different complications and comorbidities

 or did not assess direct medical costs specific to patients with different demographic characteristics or diabetes treatments.

 Using data specific to so many variables (ie, demographics, tobacco use status, treatments,  complications, comorbidities) allowed this study toidentify differences in component medical costs better than studies in which these costs were aggregated. Therefore, we likely arrived at a more realistic overall direct medical cost for a diverse diabetic population. Finally, our data were more representative of the wide array of US health systems. They were from health plans across the United States and accounted for the high degree of regional variation in both disease severity and practice patterns. This is an advance over past studies that used data from single health plans or from more homogeneous populations.

Despite these strengths, our study is limited in that TRIAD participants were all enrolled in managed care organizations, so their healthcare utilization and costs might not represent those of uninsured patients or those enrolled in fee-for-service health plans. Second, our study sample might inadvertently have included some people with type 1 diabetes, although this number would likely be small. Third, we may have underestimated costs because we reported only direct medical costs to the health plan and did not include patient copayments, deductibles, out-of-pocket costs, direct nonmedical costs, or indirect costs. Fourth, our calculations did not consider the possibility that interactions between multiple complications and comorbidities might increase or decrease cost estimates compared with simply using the multipliers for each individual characteristic. To obtain anidea of the direction of the bias, we chose a subgroup with multiple comorbidities and found that the predicted cost was greater than the actual mean cost; however, the 95% confidence intervals of the 2 costs were wide. Thus, the biasis likely to be  omparable to the error of measurement in  magnitude. Fifth, we were not able to convert all the cost data into a single-year cost unit because the way in which health plans reported the data did not allow us to do so. In the regression model, we tried controlling for the year in which the patient was interviewed. This allowed us to average costs from different years and smooth some of the cost inflation in the later years. We found that the coefficients were not significantly different from zero, so we subsequently dropped the year indicators from the model. In addition, in constructing the cost variable, we had already considered to some extent the inflation factor: costs were constructed using healthcare utilization multiplied by the standard feeor payment schedule used by CMS in a single year. Sixth, the cost data were almost 10 years old. We addressed this issue by reporting adjusted cost to 2010 dollars using the Consumer Price Index on medical care in addition to the original cost. However, using multipliers instead of absolute differences in cost, our estimated multipliers were relatively independent of the year of data. Finally, due to the small numbers of patients in some complication and comorbidity subgroups and the multicollinearity of a number of the factors, cost estimates for some complications and comorbiditieswere not included in the  model.

We used rigorous modeling methods to estimate component component costs specific to several demographic characteristics, tobacco use status, diabetes treatments, complications, and comorbidities for patients with type 2 diabetes in managed care organizations in the United States from 1999 to 2002. We found that the large proportion of overall direct medical costs for type 2 diabetes is attributable to diabetes complications and comorbidities, and especially to ESRD with dialysis or kidney transplantation treatment. Economic researchers can use these estimates in simulation models to assess the potential cost-effectiveness of interventions intended to prevent or delay type 2 diabetes and its complications

 From Division of Diabetes Translation (RL, PZ), Centers for Disease Control and Prevention, Atlanta, GA; Department ofInternal Medicine (DB, WHH), University of Michigan, Ann Arbor, MI; School of Public Health (MBB), University of Michigan, Ann Arbor, MI; Department of Health Services (SLE), UCLA, Los Angeles, CA; FeinbergSchool of Medicine (RTA), Northwestern University, Chicago, IL; Department of Family Medicine and Community Health (JCC), UMDNJ-Robert Wood Johnson Medical School, New Brunswick, NJ.

The authors (RL, DB, PZ, SLE, RTA, WHH) report no relationship or financial interest with any entity that would pose a conflict of interest with the subject matter of this article.

Concept and design (RL, PZ, SLE, RTA, JCC, WHH); acquisition of data (DB, WHH); analysis and interpretation of data (RL, DB, MBB, SLE, RTA, WHH); drafting of the manuscript (RL, DB, MBB, RTA); critical revision of the manuscript for important intellectual content (RL, MBB, PZ, SLE, RTA, JCC, WHH); statistical analysis (RL); provision of study materials or patients (WHH); and obtaining funding (WHH).

This study was jointly funded by Program Announcement number 04005 from the Centers for Disease Control and Prevention (Division of Diabetes Translation) and the National Institute of Diabetes and Digestive and Kidney Diseases. This analysis was also supported by the Biostatistics and Economic Modeling Core of the Michigan Diabetes Research and Training Center, Grant P60DK020572 from the National Institute of Diabetes and Digestive and Kidney Diseases.

 Rui Li, PhD, Division of Diabetes Translation, Centers for Disease Control and Prevention, 4770 Buford Hwy, NE, MS K-10, Atlanta, GA 30341. E-mail: eok8@cdc.gov.1. King H, Aubert RE, Herman WH. Global burden of diabetes, 1995-2025: prevalence, numerical estimates, and projections. Diabetes Care.1998;21(9):1414-1431.

2. Wild S, Roglic G, Green A, Sicree R, King H. Global prevalence of diabetes: estimates for the year 2000 and projections for 2030. DiabetesCare. 2004;27(5):1047-1053.

3. Hogan P, Dall T, Nikolov P; American Diabetes Association. Economic costs of diabetes in the US in 2002. Diabetes Care. 2003;26(3): 917-932.

4. American Diabetes Association. Economic costs of diabetes in the U.S. in 2007 [published correction appears in Diabetes Care. 2008;31(6): 1271]. Diabetes Care. 2008;31(3):596-615.

5. American Diabetes Association Consensus Panel. Guidelines for computer modeling of diabetes and its complications. Diabetes Care. 2004;27(9):2262-2265.

6. Brown JB, Nichols GA, Glauber HS, Bakst AW. Type 2 diabetes: incremental medical care costs during the first 8 years after diagnosis. Diabetes Care. 1999;22(7):1116-1124.

METHODS

A large proportion of medical costs for type 2 diabetes are attributable to complications and comorbidities, especially end stage renal disease with dialysis or kidney transplantation.

Medical Costs Associated With Type 2 Diabetes Complications and Comorbidities

 It is unknown to what extent the gap between need and care for depression among patients with diabetes differs across racial/ethnic groups. We compared, by race/ethnicity, the likelihood of clinical recognition of depression (diagnosis or treatment) of patients who reported depressive symptoms in a well-characterized community-based population with diabetes.

We used a survey follow-up study of 20,188 patients with diabetes from Kaiser Permanente Northern California. Analyses were limited to 910 patients who scored 10 or higher on the Patient Health Questionnaire (PHQ-8) which was included in the survey and who had no clinical recognition of depression in the 12 months prior to survey. Clinical recognition of depression was defined by a depression diagnosis, referral to mental health services, or antidepressant medication prescription.

Among the 910 patients reporting moderate to severe depressive symptoms on the survey and who had no clinical recognition in the prior year, 12%, 8%, 8%, 14%, and 15% of African American, Asian, Filipino, Latino, and white patients, respectively, were clinically recognized for depression in the subsequent 12 months. After adjusting for sociodemographics, limited English proficiency, and depressive symptom severity, racial/ethnic minorities were less likely to be clinically recognized for depression compared with whites (relative risk: Filipino: 0.30, African American: 0.62).

 More work is needed to understand the modifiable patient and provider factors that influence clinical recognition of depression among diabetes patients from different racial/ethnic groups, and the potential impact of low rates of clinical recognition on quality of care.

 A total of 88% of patients with no clinical recognition of depression in the prior year and reporting moderate to severe depressive symptoms on a research survey were not clinically recognized within the subsequent 12 months; Filipinos and African Americans were less likely to be clinically recognized with depression than whites. Depression is one of the most common and costly mental disorders among primary care patients in the United States.

 Depression is even more common among people with diabetes, 

 among whom it is associated with less adequate self-care (eg, poorer diet, physical inactivity, medication nonadherence, and poorer glycemic control) and poorer quality of life.

Comorbid depression and diabetes has also been linked to increased risk of mortality and cognitive decline.

 Findings from the TRIAD study involving 8790 adults with diabetes enrolled in 10 managed care health plans in 7 states indicate that depressed patients achieve poorer diabetes control compared with patients who were not depressed.

 Depression is treated less aggressively in patients with multiple comorbidities

 and among patients with diabetes, depression is often undertreated, particularly in racial/ethnic minority groups.

Compared with whites, US racial/ethnic minority groups have a greater prevalence of diabetes,

 have greater concerns about medication use,

 are more likely to have poorly controlled diabetes,

 and experience greater incidence of some of the major complications related to diabetes.

These groups may also receive less adequate treatment for comorbid depression. Depression is underdiagnosed in general

 and may be even more so in some racial/ethnic minority groups.

 While findings from psychiatric epidemiologic studies indicate that rates of depression diagnosis for racial/ethnic minorities are generally lower than those for whites, members of racial/ethnic minority groups diagnosed with clinical depression report a greater burden of depressive symptoms and rate their depression as more severe and disabling than for whites.

 Higher rates of depressive symptoms among racial/ethnic minorities despite lower rates of diagnoses may indicate disparities in the rate of clinical recognition of depression.

Randomized controlled trials investigating the efficacy of different depression treatment modalities suggest that reductions in the recurrence of depression may lead to improvements in diabetes-related outcomes such as hemoglobin A1C,

 and a recent trial of patient-centered management of depression and chronic disease showed significantly improved control of medical disease and depression, suggesting the utility of depression treatment when tailoring care.

Despite the well-documented clinical importance of depression for patients with diabetes and epidemiologic evidence of racial/ethnic disparities in diabetes prevalence, diabetes control, and outcomes, there is relatively little research regarding racial/ethnic disparities in the recognition and care of depression among multi-ethnic populations with diabetes in usual care settings. We evaluated whether there were significant racial/ethnic differences in the likelihood that patients with diabetes who self-reported significant depressive symptoms would be diagnosed or treated for depression during a 12 month follow-up. The current study is based on a sample drawn from a single integrated healthcare delivery system, which may reduce confounding by access to care.

Study participants were drawn from the Diabetes Study of Northern California (DISTANCE), a follow-up study among members of the Kaiser Permanente Northern California Diabetes Registry conducted in 2005-2006. Kaiser Permanente, an integrated, non-profit, group-practice healthcare delivery organization, provides comprehensive medical services to over 3 million members in Northern California, more than 25% of the region’s population. Care is provided by more than 7000 providers at 19 hospitals and 152 medical offices. Kaiser Permanente members are predominantly employed or retired individuals and closely approximate the general population of the region ethnically and socioeconomically except for the extreme tails of the income distribution.

DISTANCE was designed to assess the association of patient, provider, and health system factors with health outcomes among patients with diabetes from 5 racial/ethnic groups. A randomly selected, ethnically stratified sample of members with diabetes receiving care from Kaiser Permanente Northern California (“Kaiser Permanente”) was invited in 2005 to 2006 to complete the survey. Of these, 20,188 patients (62% response rate among eligible members) completed surveys: 3420 African Americans (16.9%), 2312 Asians (excluding Filipinos) (11.4%), 4602 whites (22.8%), 2404 Filipinos (11.9% [we examined Filipinos separately from the broader Asian racial/ethnic group because Filipinos had a significantly different sociodemographic profile compared with the Asian group. While the DISTANCE sample did include other Pacific Islander groups, the numbers were substantially lower, so they were combined into 1 category which also included Native Americans, Eskimo, multiracial, and other/unknown]), 3717 Latinos (18.4%), 2222 multiracial (11.0%), and 1511 other (7.5%). Demographic, clinical, behavioral, and census data were available for all invited participants. No response bias was detected when comparing, among responders versus non-responders, the associations of poor glycemic control (A1C >9%) with race (P = 0.55); subsequent assessments of selection bias also failed to detect response biases.

The DISTANCE survey took an average of 45 to 60 minutes to complete and included 4 modes of administration: 1) a computer-assisted telephone interview (CATI) administered by a third party, 2) a password-enabled, Internet-accessible survey (“web survey”) maintained on a secure server at the Kaiser Division of Research, 3) a self-administered, written survey, or 4) a short version of the written survey (the short written version was abridged and contained 40 questions). Offering the survey by oral interview in multiple languages was intended to mitigate the language and/or literacy barriers. The written and web surveys were in English only, but the CATI was available in English, Spanish, Cantonese, Mandarin, and Tagalog, using certified translations of an English script. The content of each survey mode was identical except for slight adjustments in wording as needed. Details on study recruitment procedures and sample characteristics have been published previously.

 The DISTANCE study was approved by the Institutional Review Boards of the Kaiser Foundation Research Institute and the University of California, San Francisco.

The survey included the 8-item Patient Health Questionnaire (PHQ-8), a validated screener for depression, which has been found to function similarly across different racial and ethnic groups.

 The PHQ-8 yields a valid proxy of depression diagnosis compared with other, longer clinician-administered diagnostic instruments and is widely used in clinical practices, including Kaiser Permanente.

 The PHQ-8 was used to identify current depressive symptoms. The self-report questionnaire consists of 8 out of the 9 

Diagnostic and Statistical Manual of Mental Disorders, 4th Edition

, depressive disorder criteria: anhedonia, depressed mood, trouble sleeping, loss of energy, changes in appetite, trouble concentrating, and psychomotor retardation or agitation experienced over the last 2 weeks. In line with current practice, patients who scored >10 were classified as meeting criteria for depression care.

 We also examined severity of depressive symptoms using validated cut points for depression severity: moderate (PHQ score 10-14), moderatesevere (PHQ score 15-19), and severe (PHQ score >19).

Clinically recognized depression (CRD) was our outcome of interest. Similar to previous studies, CRD was determined by the presence of a diagnosis, referral, or treatment of depression in the patient’s medical records.

 More explicitly, CRD was defined by the presence of any of the following within 12 months following respondents’ participation in the DISTANCE survey: 1) the diagnosis of depression in patient’s history/medical chart (

International Classification of Diseases, Ninth Edition

 codes 296.2X, 296.3X, 296.5X, 296.8X, 296.9X, 300.4X, 309.0X, 309.1X, 309.2X, 311, 648.4X, V790, or 307.44), 2) physician referral to mental health services for depression treatment captured electronically within the Kaiser Permanente system, and/or 3) prescription written for first-line anti-depressant medications (citalopram, fluoxetine, fluvoxamine, paroxetine, sertraline, or escitalopram). We included diagnoses and treatments regardless of the clinical setting from which they were generated (eg, primary care visits or mental health visits).

There were 10,543 DISTANCE respondents who selfidentified as African American/black, Asian, Filipino, Latino, or white (

). Of these, 1370 (13%) met criteria for moderate to severe depression according to the PHQ-8 (score >10)41 on the DISTANCE survey. We excluded the 33.1% (460/1370) of these who had been clinically recognized by a Kaiser Permanente provider in the 12 months prior to DISTANCE survey participation and limited our analyses to the remaining 910 patients with self-reported depressive symptoms and no diagnosis, referral, or treatment for depression in the prior year. Within our study cohort, 24% (217/910) had an earlier history (ie, more than 12 months prior to the survey) of CRD within Kaiser medical records.

Covariates included sociodemographic factors, including respondents’ age and gender, depression symptom severity, limited English proficiency, number of medical visits, and patients’ level of medical illness severity using the Deyo version of the Charlson Comorbidity Index.

 Age at survey completion was included as a continuous variable. Limited English proficiency was a binary variable defined as respondents’ selfreported difficulty in reading or speaking English.

All data were analyzed using SAS version 9.13.

 We evaluated whether there were racial/ethnic differences in the likelihood of CRD during 12 months of follow-up after the survey among patients who self-reported significant depressive symptoms in the DISTANCE survey. We used modified Poisson regression models to obtain relative risk estimates to examine the association between race/ethnicity and likelihood of CRD.

 Results from previous studies suggest that modified Poisson regression is a superior method to log binomial regression for estimating relative risks when examining prospective data with binary outcomes.

 Estimated relative risk (RR) and 95% confidence intervals were generated using a robust error variance procedure known as sandwich estimation,

 which was implemented using the SAS PROC GENMOD procedure with the REPEATED statement.

 We first created unadjusted Poisson regression models to test for racial/ethnic differences in CRD. Given the race-stratified random sampling design, all models were weighted for sampling fractions (expansion weighting). Weighting for sampling fractions compensates for the unequal probability potential respondents in the sample population had in being selected into the DISTANCE study.

 Although the clinical recognition of depression should not depend on whether someone is of a particular race/ethnicity, sex, age, or clinical state, we specified additional models adjusted for age, sex, depression symptom severity, number of visits, and limited English proficiency. Finally, we examined whether there was effect modification by sex by specifying interaction terms between race/ethnicity and sex.

There were some significant differences in sociodemographic characteristics of the racial/ethnic groups. African Americans were, on average, older and Latinos were younger compared with the other racial/ethnic groups (see 

), while African American, Latino, and Filipino subgroups had relatively higher proportions of women versus men compared with Asians and whites. Education also varied: Latinos were the least likely to have completed high school while Filipinos, Asians, and whites were the most likely to report having a college degree. There were no significant racial/ethnic differences in the distribution of depression symptom severity based on the self-reported screener (PHQ-8). There were significant differences in the number of medical visits across racial/ethnic groups. On average, African Americans had the highest number of visits during the follow-up period and Asians had the fewest number of visits. This utilization pattern tracked the significant differences in comorbidity burden, with the highest levels in African Americans, lowest in Asians, and intermediate and similar in Filipinos, Latinos, and whites.

Our first finding concerns the rate of CRD during the 12 months after self-reporting moderate to severe depressive symptoms (PHQ-8 >10) on our survey, among those whose depression was not currently recognized in the medical system (ie, no CRD in the prior 12 months). In this sample, 12% (110/910) overall (12%, 8%, 8%, 14%, and 15% of African American, Asian, Filipino, Latino, and white patients, respectively), were clinically recognized (ie, diagnosed, referred, or treated for depression) within 12 months of each individual’s survey date. The second finding concerns the relative differential in rates of clinical recognition across racial/ethnic groups (Table 2). In the unadjusted model, patients from each of the 4 racial/ethnic minority groups were less likely than whites to be clinically recognized for depression within 12 months, although the absolute differences in rates were small and were statistically significant only for Filipinos (RR, 0.33; 95% confidence interval [CI], 0.17-0.65). Across all models, Filipinos remained significantly less likely to have CRD compared with whites. While Asians were less likely to have CRD compared with whites, these differences were not significant in any model. For Latinos and African Americans, differences were evident in some models but not others. Latinos were less likely than whites to have CRD when adjusted for age and sex in Model 2 (RR, 0.58; CI, 0.36-0.95) and for age, sex, depression severity, and limitations on English proficiency in Model 3 (RR, 0.53; CI, 0.29 to 0.97), but this was not significant either in the initial unadjusted model or when further adjusted for medical visits in Model 4 (RR, 0.57; CI, 0.31-1.04) and Charlson index in Model 5 (RR, 0.57; CI, 0.31-1.05). Differences in CRD between African Americans and whites were not significant in Models 1 to 3, but became significant once analyses were further adjusted for medical visits in Model 4 (RR, 0.58; CI, 0.35-0.97) and remained significant in Model 5 after adjusting for the Charlson index (RR, 0.58; CI, 0.35-0.97).

Finally, we tested whether the race/ethnic differences in CRD were modified by sex but we found no statistically significant interactions between race/ethnicity and gender in ourmodels of CRD, indicating that the patterns of clinical recognition variation across racial/ethnic groups were similar for men and women (analysis not shown).

The first finding from this study was that among patients with diabetes who had significant depressive symptoms and who had not been clinically recognized with depression in the prior year, few (less than 15%) were clinically recognized any time during the following year. This finding is consistent with the prior evidence regarding the underrecognition of depression in populations with diabetes from primary care settings.53 In the current study design, we excluded the 34% (460/1370) of subjects with PHQ-8 scores consistent with depression who had been clinically recognized in the previous 12 months; this exclusion may have resulted in a study sample with a lower likelihood of subsequent depression recognition. Rates of recognition may also have increased in recent years within the Kaiser system with the subsequent introduction of regular screening and prioritization of care for depressive symptoms in routine diabetes care within the Kaiser system.

Low rates of clinically recognized depression may be due to provider or patient factors; both may be uncomfortable about discussing symptoms of depression and patients may want to avoid the stigma sometimes associated with a diagnosis of depression. There may also be cultural differences in communication style and in what is seen as appropriate to share in a medical encounter.

Depression often goes unrecognized in primary care among symptomatic diabetes patients, especially in some minorities.

May 2013

Can Cancer Patients Seeking a Second Opinion Get Better Care?

Medical Costs Associated With Type 2 Diabetes Complications and Comorbidities

Differences in the Clinical Recognition of Depression in Diabetes Patients: The Diabetes Study of Northern California (DISTANCE)

Patient Experience Over Time in Patient-Centered Medical Homes

Medicare Prescription Drug Plans as Perceived by Public Health Providers

Variability in Resource Use: Diagnosing Colorectal Cancer

A Predictive Model of Hospitalization Risk Among Disabled Medicaid Enrollees

Emergency Department Use Associated With Primary Care Office Management

Pragmatic Clinical Trials: US Payers' Views on Their Value

Effects of Integrated Delivery System on Cost and Quality

Well-Child Care Visits and Risk of Ambulatory Care-Sensitive Hospitalizations