|Home | About | Journals | Submit | Contact Us | Français|
The Chronic Care Model (CCM) is a framework aimed at improving chronic illness care. We used empirical data from previously published literature to address two research questions: 1) do interventions that incorporate one or more elements of the CCM result in improved outcomes of interest for specific chronic illnesses; and 2) are any elements essential for improved outcomes?
Articles were identified from narrative literature reviews and quantitative meta-analyses, each of which covered multiple bibliographic databases from inception to March 2003. We supplemented this identification strategy by conducting our own search of the MEDLINE database (1998-2003) and consulting experts. We included randomized and non-randomized controlled trials of interventions that contained one or more elements of the CCM for asthma, congestive heart failure (CHF), depression, and diabetes. From these studies we extracted data on specified clinical outcomes, quality of life, and processes of care. We then used random effects modeling to compute pooled standardized effect sizes and risk ratios.
Of 1,345 abstracts screened, 112 studies contributed data to the meta-analysis: asthma, 27 studies; CHF, 21 studies; depression, 33 studies; and diabetes, 31 studies. In the aggregate, interventions that contained one or more CCM elements had beneficial effects on clinical outcomes and processes of care, and these effects were consistent across all conditions studied. The effects on quality of life were mixed, with only the CHF and depression studies showing benefit. Publication bias was noted for the CHF studies and a subset of the asthma studies.
Interventions that contain one or more elements of the CCM improve clinical outcomes and processes of care -- and to a lesser extent, quality of life -- for patients with chronic illnesses.
Approximately 90 million persons in the U.S. live with one or more chronic illnesses.(1) The management of health care delivery for persons with chronic illnesses has advanced substantially in recent decades, yet considerable deficiencies in the quality of chronic illness care remain.(2-6) Critics argue that systems of care designed to deal with acute episodes do not serve the needs of patients with chronic illness well.(7) Chronic illnesses rank among the nation’s costliest conditions,(8) and five chronic illnesses -- asthma, diabetes, heart disease, hypertension, and mood disorders -- account for nearly one-half of U.S. health care expenditures(9). Managed care organizations are increasingly becoming the main source of health care services for persons with chronic illnesses such as diabetes (10) and may be particularly well-suited to adopting reforms to optimize care for such patients.
The Chronic Care Model (CCM) is a primary care-based framework aimed at improving the care of patients with chronic illnesses.(11-14) The model integrates a number of elements into a plausible package designed to foster more productive interactions between prepared, pro-active teams and well-informed, motivated patients. Details on the concept, implementation, and evidence base of the CCM are available at http://www.improvingchroniccare.org.
Although individual components of the CCM have been rigorously studied, accounts of beneficial effects of implementing the CCM as a whole on processes, outcomes, and/or costs come largely from self-reported uncontrolled studies.(11;14;15) We therefore sought in this meta-analysis to classify previously published studies according to CCM component(s) implemented, to address two related research questions: 1) to what extent do interventions that incorporate one or more elements of the CCM result in improved outcomes of interest for specific chronic illnesses; and 2) are some elements of the CCM more effective than others? Although the CCM is intended to be generic, applicable across all types of chronic illnesses,(11;18) we focused on evaluating the effects of diverse interventions on clinical outcomes, quality of life, and processes of care for four chronic illnesses that would be of particular interest to managed care organizations: asthma, congestive heart failure, depression, and diabetes.
The CCM identifies six elements deemed to be essential for providing high-quality care to patients with chronic illnesses: Delivery System Design, Self-Management Support, Decision Support, Clinical Information Systems, Community Resources, and Health System Organization.(11;13;14) Because of the limited information available in published descriptions of interventions, and because the CCM elements are broadly defined, we categorized interventions according to the CCM elements they incorporated (Box 1).(19) Any given intervention could contain more than one element, up to a maximum of six.
One author (A.C.T.) was responsible for all aspects of study selection and data abstraction. First, we used MEDLINE and the Cochrane Library to identify 20 recently published systematic reviews and meta-analyses of the four chronic illnesses of interest: asthma,(20-24) congestive heart failure,(25-30) non-insulin dependent diabetes mellitus,(31-37) and depression.(38;39) These 20 reviews, in addition to three other reviews that were not condition-specific,(40-42) formed the primary substrate for identifying studies for potential inclusion in our study sample. Although these previously published reviews covered multiple databases -- including CINAHL, the Cochrane Library, the Dissertation Index, Embase, Healthstar, HELMIS, MEDLINE, PsycINFO, PsychLit, and Sigle -- from inception to March 2003 (depending on the date the review was published), we also performed a MEDLINE search for more recent studies (January 1998-June 2003) that may have been missed. Finally, additional studies were identified by consulting experts and searching independently maintained bibliographies (e.g., the Chronic Care Bibliography(43)).
In identifying abstracts and articles for potential inclusion in our sample, the authors and their affiliations were not masked. We screened all identified abstracts for potentially relevant studies that assessed the effects of interventions containing one or more of the CCM elements. Randomized and non-randomized controlled studies were eligible for inclusion. Uncontrolled studies were excluded, as were studies written in a language other than English. Because we sought to include only the most recent decade of published evidence in our report, we excluded studies published prior to 1993.
These studies employed a wide range of outcome variables. We identified a clinical outcome, a quality of life outcome, and a key process of care for each condition (Table 1). To make the analysis more tractable and to maximize our ability to pool data, we opted to collect data on the variables that were reported by the greatest number of studies (e.g., for the diabetes studies we selected HbA1c as the clinical outcome rather than cholesterol level and HbA1c monitoring as the key process of care rather than funduscopic testing). We also accepted a variety of similar measures for some domains. For example, we accepted any of the most commonly used measures for the clinical outcome among the depression studies, including the Hopkins Symptom Checklist-20 (9 studies), the Hamilton Depression Rating scale (6 studies), and the Center for Epidemiological Studies-Depression scale (6 studies). The most commonly used measures of disease-specific quality of life were the St. George’s Respiratory Questionnaire (4 studies), the Juniper Asthma Quality of Life Questionnaire (4 studies), and the Minnesota Living with Heart Failure Questionnaire (3 studies).
Among the studies selected with data on the clinical outcomes, some reported data as count variables or continuous variables while others reported data as dichotomous variables. We extracted data in both forms where possible. Too few studies on asthma and congestive heart failure reported the clinical outcome data as count variables (number of ED visits or readmissions during the study period), so for those two condition-specific analyses we only used the data that were reported as dichotomous variables (number with any ED visit during study period, number readmitted during study period).
Using a standardized data collection form, we abstracted data on study characteristics, including CCM elements incorporated in the interventions and effects on clinical outcomes, quality of life, and processes of care. If the results of a study were reported through multiple articles, data were abstracted from all articles and attributed to the primary citation. If a study reported comparisons at multiple follow-up times, we abstracted data on outcomes closest to 12 months of follow-up. For each study that reported data as continuous variables, the mean and standard deviation values were extracted (if available). For studies that did not report a standard deviation value at follow-up, the standard deviation was assumed to be equal to one-quarter of the theoretical range for that measure. This imputation approach is based on the assumption that the underlying distributions were approximately normal, and it is a very conservative assumption in that the multiplication factor for truly normal distributions should be closer to one-sixth. For the clinical outcomes, a higher score represented a worse outcome, whereas for quality of life and processes of care, a higher score represented a better outcome; some studies required recoding of quality of life measures for consistency. For example, if a study used the St. George’s Respiratory Questionnaire as a measure of quality of life (0-100, with higher scores indicating poorer quality of life), we recoded the variables so that higher scores represented better quality of life. We excluded studies that did not provide sufficient statistics for data abstraction if our efforts to contact the study’s authors for additional data were unsuccessful.
We conducted all analyses using the Stata statistical software package (version 8.0, Stata Corporation, College Station, Tex.).(44-49) For each study, effect sizes were calculated for the intervention group relative to the comparison group at follow-up. If outcomes were measured on a continuous scale, we computed Hedges g, the standardized effect size, with an adjustment to correct for small sample bias.(50) For each intervention-control comparison of a dichotomous outcome, we calculated the risk ratio and its standard deviation.(51) Effect sizes were estimated for each type of outcome variable using random effects models (that incorporate both between- and within-study variance(52)), first pooled across conditions and then stratified by condition. We a priori deemed pooled estimates based on fewer than five studies to be unreliable for statistical hypothesis testing, as noted in the results. To check for publication bias (which may result from the non-publication of small negative studies) we visually assessed funnel plots for asymmetry and employed the regression asymmetry test.(53)
We used a multivariate approach to independently assess the effect of each CCM element on the estimated pooled effect size, after adjusting for the presence of the other elements if the study’s intervention contained more than one. To do this, we fit random-effects meta-regression models(54) for each of the four types of outcomes. The only covariates included in these regressions were a constant term and six indicator variables equal to unity if the intervention included that particular CCM element, zero otherwise.(55) Some of the CCM elements were implemented in too few studies for a pooled estimate to be computed, so we labeled those situations as not estimable (NE) in the results. All statistical hypothesis tests were carried out at the two-sided 0.05 level of confidence.
The degree of between-study heterogeneity was assessed by the chi-squared test for heterogeneity based on Cochran’s Q.(56) We also calculated the heterogeneity statistic I2, which is independent of the number of studies and the effect size metric and can be interpreted as the proportion of total variation in the estimated treatment effect that is due to between-study heterogeneity rather than chance.(57) We fit random-effects meta-regression models(54) to determine whether study-level variables explained the heterogeneity in the treatment effects, including in these models the following study-level variables: indicators for CCM element, indicators for type of chronic illness, duration of follow-up, inpatient or outpatient setting, and Jadad quality score.
In post hoc meta-regression analyses, we included variables to determine whether there were synergies associated with the use of more CCM elements. The first specification included an ordinal variable for the number of CCM elements in the intervention. The second specification examined only the subset of studies that included Delivery System Design (the most frequently studied CCM component) in the intervention. Studies that incorporated Delivery System Design plus any additional CCM elements were compared to studies that incorporated Delivery System Design only. The third specification was similar but examined only the subset of studies that included Self-Management Support (the second most frequently studied CCM component).
We used the Jadad scale to assess the quality of studies in our sample.(58) Each study received from zero to five points (with higher scores representing higher methodological quality), depending on whether it was described as randomized or double blind, if the randomization sequence or blinding procedure was appropriate, and if it provided detailed information on withdrawals and dropouts. Because studies of low methodological quality have been found to overestimate treatment benefit,(59) we conducted a sensitivity analysis by re-estimating pooled effects and re-fitting meta-regression models using only the 36 studies with a Jadad quality score of three or greater.
Figure 1 illustrates the flow of literature from the original sources to final acceptance for our review. We identified 1,345 abstracts of studies published between January 1993 and June 2003. On the basis of screening the abstracts, we excluded 1,041 abstracts and requested 304 articles for detailed review. Of these, we excluded an additional 192 articles. The Evidence Table, available at http://www.***, provides a list of references for the 112 studies included, as well as further detail on each of these studies. Any given study could contribute data on one or more outcomes of interest, so we report sample sizes for the stratified analyses in the tables below.
Summary statistics for the sample of studies are provided in Table 2. Of the 112 studies in our analysis, 93 studies (83%) were described as randomized. Only 36 studies (32%) scored a three on the Jadad scale, and none scored higher than three. The primary limitation of these studies was the lack of double blinding, but double blinding is rarely possible in studies of organizational interventions. Almost half of the studies (46%) contained only one CCM element in the intervention. The most common CCM elements contained in the interventions were Self-Management Support (n=80), Delivery System Design (n=60), and Decision Support (n=38). The two most common elements, Self-Management Support and Delivery System Design, were frequently bundled together with at least one other element (64% [51/80] and 83% [50/60], respectively). Only eight studies included four CCM elements in the intervention. Most interventions were carried out in the outpatient setting, but four congestive heart failure studies and one depression study tested inpatient interventions.
We first examined the overall effectiveness of interventions with one or more CCM elements. Overall, the interventions led to statistically significant improvements in each of the four outcomes of interest. In particular, for the 52 studies that reported clinical outcomes as continuous variables (depression and diabetes only), there was a statistically significant pooled effect size of −0.23 in favor of the intervention (95% confidence interval [CI], −0.31 to −0.15; P<0.001 for whether the reduction was less than zero; see Figure 2). The 46 studies that reported clinical outcomes as dichotomous variables yielded a statistically significant pooled relative risk of 0.84 in favor of the intervention (95% CI, 0.78-0.90; P<0.001 for whether the relative risk was less than one; see Figure 3). The overall effect on quality of life outcomes, based on data from 24 studies, was statistically significant and favored the intervention (0.11; 95% CI, 0.02-0.21; P=0.023; see Figure 4). Finally, 32 studies contributed data on process outcomes of interest for a pooled relative risk of 1.19 (95% CI, 1.10-1.28; P<0.001; see Figure 5).
The overall effects were pooled across conditions. To enhance the clinical interpretation of our findings, we report the condition-specific pooled effect sizes in Table 3. Across conditions, the interventions had consistent effects on clinical outcomes and processes. With respect to quality of life, however, the asthma studies showed equivocal results, despite an adequate number of studies.
The pooled effect size (−0.19) for the diabetes studies reporting continuous data on hemoglobin A1c indicates a lower hemoglobin A1c in the intervention group at follow-up as compared to the control group. Assuming a range of standard deviation values from 1.56 to 2.47 (the inter-quartile range for the studies in our sample), this effect size is equivalent to a reduction in hemoglobin A1c of −0.30% to −0.47%. The positive effect size reported for the congestive heart failure studies (0.28) indicates a higher quality of life scale score in the intervention group as compared to the control group and is equivalent to an increase of 5.6-6.7 points on the Chronic Heart Failure Questionnaire.
We assessed publication bias along two dimensions: outcome measure and condition. Based on funnel plot inspection and the regression asymmetry test, evidence of publication bias was apparent for the congestive heart failure studies (all outcomes) and the asthma studies with dichotomous data on the clinical outcome.
In the meta-regression analyses, we found that four elements of the CCM (Delivery System Design, Self-Management Support, Decision Support, and Clinical Information Systems) were associated with better outcomes and processes, after adjusting for the presence of other elements if the intervention contained more than one (Table 4). Our sample had too few studies that implemented the Community Resources and Health Care Organization elements to judge the relative effectiveness of those two elements. The statistically significant effects observed for Delivery System Design and Self-Management Support could be attributed to both the larger number of studies and larger estimated effects. Decision Support improved process significantly but not outcomes. We observed no statistically significant effects for Clinical Information Systems, perhaps due to the smaller number of studies. We also noted that no single element of the CCM was essential to improved outcomes (Appendix Table 5, available at http://www.***).
Evidence of substantial between-study heterogeneity was noted for the analyses of continuous clinical outcomes (I2 = 78%; Cochran’s χ2 test for homogeneity Q=230.4, df=51, P<0.001), dichotomous clinical outcomes (I2 = 67%; Q =135.19, df=45, P<0.001), quality of life (Q =92.81; df=23; P<0.001, I2 = 75%), and processes (I2 = 90%; Q=311.59, df=31, P<0.001). These I2 values indicate that more than two-thirds of the variation in the estimated treatment effects may be attributable to between-study heterogeneity. We fit random-effects meta-regression models in order to identify potential explanations for this variation, but the study-level variables we considered -- indicators for type of chronic illness, indicators for CCM element, duration of follow-up, inpatient or outpatient setting, and Jadad quality score -- did not appreciably reduce the unexplained variance.
In our random-effects meta-regression analyses, the number of CCM elements incorporated in the study intervention was not associated with better outcomes, with P-values ranging from 0.38 to 0.81. Only two CCM domains, Delivery System Design and Self Management Support, were represented in enough studies for a meaningful comparison of interventions consisting of that element alone vs. that element in conjunction with other elements. In our data, these comparisons were never statistically significant, with P-values ranging from 0.13 to 0.61.
We re-fit the meta-regression models to the data subset of studies (n=36) that scored a three on the Jadad scale.(59) The results of this sensitivity analysis yielded results that were qualitatively similar to the main analysis (Appendix Tables 6 and 7, available at http://www.***). The pooled effect sizes were generally larger (favoring the intervention), with larger confidence intervals.
Interventions that have incorporated one or more elements of the CCM have had beneficial effects on clinical outcomes and processes of care for patients, and the results were consistent across a variety of chronic illnesses. Our estimated pooled effect size estimates, although small-to-moderate-sized,(60) are also broadly consistent with those reported in prior meta-analyses.(24;30;37;42) Interventions directed at diabetes care, for example, led to a 0.30%-0.47% reduction in hemoglobin A1c. Managed care organizations may realize benefits from even smaller reductions in mean population values for continuous risk factors such as lipid levels and hemoglobin A1c. For example, the European Prospective Investigation of Cancer and Nutrition (EPIC-Norfolk) estimated that a population reduction of 0.2% in hemoglobin A1c could reduce the prevalence of men with high HbA1c levels (5%-6.9%) from 79% to 57% and reduce excess mortality by 10%.(61) We found that interventions directed at congestive heart failure led to a 5.6-6.7 point improvement in the Chronic Heart Failure Questionnaire, slightly less than the 7-9 point difference that is regarded as a minimal clinically important difference on that scale.(62)
In terms of quality of life, the evidence was mixed, with the asthma and diabetes studies showing no benefit. It has been well-established that condition-specific quality of life scales are more sensitive to changes in clinical status than are generic measures of quality of life. Most of the studies included in our meta-analyses used condition-specific quality of life scales (see the Evidence Table). For some conditions, one might reasonably expect disease-specific interventions to have a more direct effect on clinical outcomes and processes of care than on quality of life. Improvements in diabetes care, for example, are focused on preventing long-run microvascular complications beyond the end-point of the studies we examined,(63) with less focus on improving short-run quality of life. We speculate that our meta-analyses might have yielded different results, for example, had we used a quality of life measure that is more sensitive to the short-run benefits of improved glycemic control.(64;65)
The CCM elements most responsible for these benefits could not be determined from the data. Effects appeared to be somewhat stronger for Delivery System Design and Self-Management Support, although Decision Support had significant beneficial effects on processes. The other elements of the CCM may be critical infrastructure to providing high quality chronic care but are more difficult to test scientifically. For example, leadership support may be necessary to promote and sustain higher quality, but randomizing managed care organizations to receive changes in leadership support is clearly infeasible. It is no wonder that these elements have not had many scientific trials. The fact that such linkages are hard to study scientifically does not mean they are unimportant. Instead, they are supported by common sense and reports from successful organizations.
The CCM has been promoted as a unified package. Evidence that interventions with multiple components do better than single components (66) has been interpreted as supporting synergistic effects in which the whole is bigger than the sum of the parts. Some components of the model, such as building an electronic patient registry, may facilitate other components and reduce their costs. We found that single interventions were quite successful. In post hoc analyses, we attempted to identify whether there may be some advantage to having more components, but that advantage was never statistically significant and does not appear to be more than additive.
One limitation of our work is that the studies in our sample only incorporated elements of the CCM and were not designed to test the entire CCM package.(11;13;14) The RAND/UC Berkeley Improving Chronic Illness Care Evaluation (ICICE) is nearing completion, and it is the first independent and controlled evaluation of the effects of implementing the CCM as a whole. Organizations signed up for the Institute of Healthcare Improvement’s Collaboratives to improve care for specific conditions (16) and worked together to learn about the CCM and about how to make organizational changes to improve quality of care. The design of the ICICE has been published,(17) and results from the evaluation are posted at http://www.rand.org/health/ICICE/ as they become available. Despite the large scale of the evaluation -- 24 organizations with both intervention and control sites, and 12 organizations with intervention sites only -- the number of participating organizations was too small to determine which components of the CCM were most critical to success. The organizations’ characteristics and what they did differed in many ways, many times more than the number that could be studied statistically.
A second limitation is that the use of meta-analytic methods necessarily forces what are likely complex, multivariate interventions into a narrow linear framework. In this meta-analysis we aggregated results across conditions and across interventions. We attempted to investigate the sources of variation between studies, but we were unable to explain much of it. We were also unable to assess interactions between CCM element and type of chronic illness. For example, a Clinical Information Systems intervention featuring physician reminders may be particularly effective for improving care for one type of chronic illness but not for other types, and a pooled analysis would not identify the interaction. A related limitation is that we were unable to assess the intensity of implementation in the study interventions.(19;67) Perhaps the interventions we studied were successful because doing trials requires energy and commitment to the intervention concept. This energy may be an important component to initial success that is hard to transfer. If there is significant variation in the intensity of implementation of these elements across studies, simplified comparisons by the presence or absence of these elements may mask important between-study differences. And finally, we focused our data collection on selected outcomes. We needed to do so in order to aggregate across studies, recognizing that interventions may have had different effects on other outcomes and processes of care. However, the outcomes we selected were reported in a large number of studies and likely reflect outcomes of interest to managed care organizations. A final limitation is that we used an unconventional search strategy by relying on prior meta-analyses as the primary substrate for identifying our sample of studies. Doing so may have introduced unpredictable biases, but we also systematically searched the MEDLINE database and the Chronic Care Bibliography(43) to identify more recently published studies.
Despite these limitations, our meta-analysis shows that interventions that contain one or more elements of the CCM can improve outcomes and processes for several chronic illnesses of interest to managed care organizations. How to transfer the gains from these efficacy studies into the chaotic real world of health care is a different but equally important issue.
FUNDING: This work was made possible by grants No. 034984 and No. 035678 from the Robert Wood Johnson Foundation. At the time the research was conducted, Dr. Tsai was a Graduate Student Summer Associate at RAND Health and a National Research Service Award Trainee supported by U.S. Agency for Healthcare Research and Quality Institutional Training Award (T32 HS 00059-06) and Case Western Reserve University.
An earlier version of this manuscript was presented at the AcademyHealth Annual Research Meeting in San Diego, Calif. on June 7, 2004.
Alexander C. Tsai, Department of Epidemiology and Biostatistics, Case Western Reserve University School of Medicine.
Sally C. Morton, RAND Health.
Carol M. Mangione, Division of General Internal Medicine and Health Services Research, Department of Medicine, David Geffen School of Medicine at UCLA.
Emmett B. Keeler, RAND Health.