Cheng S, Chen C, Hou Y. A Longitudinal Examination of Continuity of Care and Avoidable HospitalizationEvidence From a Universal Coverage Health Care System. Arch Intern Med. 2010;170(18):1671–1677. doi:10.1001/archinternmed.2010.340
Few studies have examined the effect of continuity of care on avoidable hospitalization, and the results have been inconclusive. This study aimed to examine the effects of continuity of care on avoidable hospitalization and hospital admission for any condition in a health care system with a high level of access to care.
We used a longitudinal design to examine claims data that captured health care utilization between January 1, 2000, and December 31, 2006, under a universal coverage health insurance program in Taiwan. In total, 30 830 randomly selected subjects with 3 or more physician visits per year between 2000 and 2006 were analyzed in 3 age groups. The main outcome was avoidable hospitalization and hospital admission for any condition. A random intercept logistic regression model was used to control for age, sex, low-income status, health status, time effect, and random subject effect.
Higher continuity of care was significantly associated with lower likelihood of avoidable hospitalization in all 3 age groups. Similar associations were found for hospital admission for any condition in the 3 age groups.
Better continuity of care is associated with fewer avoidable hospitalizations and fewer hospital admissions for any condition in a health care system with easy access to care. Therefore, improvement of continuity of care is an appropriate path to follow in a universal coverage health care system.
Primary care forms the foundation of health care delivery systems and has an important role in providing integrated and accessible health care services to the public.1 However, the health care systems in many Asian and central and eastern European countries largely focus on specialist and hospital care. An important feature of primary care is the continuity of care1 (ie, the relationship between patients and physicians that goes beyond any specific episode of illness or disease2).
Previous studies have found that greater continuity of care may lead to better patient outcomes. This has been measured in the context of fewer emergency department visits,3- 7 more preventive care,8- 12 decreased hospital admissions,5,13- 16 better chronic disease control,17- 19 and less intensive care unit use.20
A classification of avoidable hospitalizations was developed by Billings and colleagues21,22 in the late 1980s to evaluate problems associated with inadequate access to primary care. It was hypothesized that timely ambulatory care could prevent unnecessary hospitalization by preventing the onset of disease, controlling an acute episodic illness, or managing a chronic illness.21 Various studies23- 27 have focused on the relationship between access to primary care and avoidable hospitalization. However, few studies have examined the relationship between continuity of care and avoidable hospitalization, and the results from these studies have been inconsistent.
Gill and Mainous14 analyzed data from Medicaid beneficiaries younger than 64 years and found that patients with higher continuity of care tended to have a lower likelihood of hospitalization rates for chronic avoidable hospitalization but not for acute avoidable hospitalization. Menec et al28 reported that a higher continuity of care score was associated with a lower risk of avoidable hospitalization among patients older than 67 years in Manitoba, Canada. Clearly, a more comprehensive analysis across multiple age groups is needed to reach a conclusion concerning the association between continuity of care and avoidable hospitalization.
Most previous studies4,6- 8,14,15 examining the relationship between continuity of care and hospital admission used a cross-sectional study design. One study13 was a double-blind randomized controlled trial, and 2 studies3,5 used a longitudinal design, which has 2 advantages over the cross-sectional design. First, the longitudinal design can determine the change over time of the outcome measurements and the factors that influence the changes, while the cross-sectional design can only assess the relationship between outcome measurements and related factors.29,30 Second, the longitudinal design may capture unobserved time-invariant characteristics (eg, health care–seeking behavior) of patients. In addition, findings from cross-sectional investigations may be confounded by unobserved characteristics that exist between patients who have higher or lower continuity of care.30,31 The present study used a longitudinal design with a representative sample from a universal coverage health care system.
Taiwan implemented a compulsory National Health Insurance (NHI) program in 1995 to ensure good access to care. Approximately 99% of Taiwan's 23 million residents were enrolled in the NHI program. Those who were not enrolled were living abroad or were incarcerated. Under the NHI system in 2006, patients sought care based on their preferences from about 508 NHI-contracted Western medicine hospitals or from approximately 9152 NHI-contracted Western medicine clinics nationwide.32 As in many Asian countries, there is no formal family physician arrangement in Taiwan. Family medicine is recognized as a specialty among 23 medical specialty groups listed by the health authority. Patients may visit a physician according to the symptoms they have and see a specialist (eg, internist, pediatrician, or dermatologist) in a community clinic or in a hospital without a referral. The lack of referral may be viewed as a form of patient self-referral.
Facilitated by the ease of accessibility, the mean number of annual physician visits in Taiwan, about 15 per person in 2005, is one of the highest in the world. Accordingly, patients in Taiwan are often criticized for their “physician shopping” behaviors.33 Features of the health care system in Taiwan may hinder communication or trust between patients and physicians and deteriorate the continuity of care. By using longitudinal analysis, this study aimed to determine the effect of continuity of care on avoidable hospitalization under a health care system with easy access to care.
The present study was conducted using a longitudinal study design. The main data source was the Longitudinal Health Insurance Database, which covered the enrollee population in 2005; this data set was a specific subset of the NHI Research Database (NHIRD) in Taiwan. The 2005 Longitudinal Health Insurance Database consisted of 1 million subjects chosen in 25 subsets of randomly selected samples from the entire NHI enrollee profile. Random sampling was conducted for every subset of 40 000 subjects enrolled at the end of 2005 in the NHI.34
The first 2 subsets of the 2005 Longitudinal Health Insurance Database were used to form a data set with 80 000 subjects. For each subject, we were able to examine all health care utilization information between January 1, 2000, and December 31, 2006, under the NHI program, including physician visits, emergency department visits, and hospital admissions. Exclusion criteria were applied to select appropriate study subjects. First, subjects were excluded if they had no medical claim records in any of the years between 2000 and 2006. Second, because a meaningful continuity of care index cannot exist for few visits and because a reliable continuity of care index is more robust with more visits, we excluded subjects with fewer than 3 physician visits in any of the years between 2000 and 2006. This exclusion criterion was adopted in previous studies.4,6,14 Third, to increase the comparability of the continuity of care index, we excluded the following services: outpatient surgery, dental care, traditional Chinese medicine, and some specific services such as long-term care or home health care services. The remaining 30 830 subjects were included in the study. Data from 2000 were considered baseline information, and each subject contributed data for the subsequent 6 years.
According to Jee and Cabana,35 continuity of care measures are classified into 5 types (ie, duration, density, dispersion, sequence, and subjective measures). When using claims data sets for analyses, researchers tend to use continuity of care indexes such as the usual provider of care index for measuring the density of visiting a physician frequently,36 the continuity of care index (COCI) for measuring the dispersion of visits,37 and the sequential continuity index for sequentially measuring the various physicians visited.38 Compared with the usual provider of care index and the sequential continuity index, the COCI is less sensitive to the number of physician visits.39 Considering the physician shopping and many physician visits in Taiwan, we chose the COCI as our primary predictor variable.
The COCI is composed of the number of different physicians seen and the number of visits to each physician.37 The general formula is as follows:
where N indicates the total number of physician visits; nj, the number of visits to physician j; j, a given physician; and M, the number of providers.
The COCI ranges between 0 and 1, with a higher value corresponding to greater continuity of care. Because the COCI values have no inherent clinical meaning, we categorized them into 3 equal tertiles (low [0.00-0.16], medium [0.17-0.33], and high [0.34-1.00]) according to the distribution of scores across the entire study population.
The main outcomes were whether a subject had an avoidable hospitalization or had a hospital admission for any condition in each of the years from 2001 to 2006; these outcome measures were coded as dichotomous variables. In defining a hospital admission for any condition, we considered the possibility that some hospital admissions could not be the result of a high level or a low level of continuity of outpatient care; therefore, we excluded admissions with diagnoses of an injury or poisoning (International Classification of Diseases, Ninth Revision, Clinical Modification [ICD-9-CM ] codes 850-995), as well as all supplementary classifications (V-codes), such as chemotherapy.
We identified an avoidable hospitalization based on the definition by the Institute of Medicine.40 An avoidable hospitalization was defined by a lesser need for hospitalization if patients received good primary care. The diagnoses for avoidable hospitalizations in our study included asthma, angina, convulsions, cellulitis, dehydration, hypertension, bacterial pneumonia, congestive heart failure, hypoglycemia, gastroenteritis, congenital syphilis, diabetes mellitus, immunization-related and preventable conditions, grand mal status and other epileptic convulsions, chronic obstructive pulmonary disease, skin grafts with cellulitis, kidney or urinary tract infection, and severe ear, nose, and throat infections.
Information concerning patient diagnoses was obtained from NHI claims. Using ICD-9-CM codes, each condition was identified by the principal diagnosis except for dehydration, which was defined by principal and secondary diagnoses. For several avoidable hospitalization conditions, specific criteria for exclusion included age, diagnosis, and procedure.40
Many previous studies4- 6,9,15 used continuity of care measures for 1 year to predict the health care outcomes in the subsequent year. However, Gill et al4 reported that continuity of care was associated with fewer emergency department visits in the same year but not in the subsequent year. We considered that patients may have had different illnesses and would have visited different physicians during these 2 years or that they may have changed their health care–seeking behavior. Therefore, we examined the effects of continuity of care on avoidable hospitalization and hospital admission for any condition in the same year.
The covariates concerning patient characteristics included in the model were age, sex, low-income status, health status, and time dummy variables. The following 3 proxy indicators were used to represent the patients' health status: the number of physician visits in the previous year, the likelihood of hospitalizations in the previous year, and the Charlson index score. The Charlson index score contains 17 categories of comorbid conditions defined by ICD-9-CM codes.41 To ensure the accuracy of a patient's condition, only diagnoses appearing at least 3 times in a patient's annual claims records were included in the calculation for the Charlson index score. In addition, the study included time dummy variables, with 2001 as the reference group, to control for characteristics that may change over time, such as the severe acute respiratory syndrome outbreak in 2003.
To enhance the comparability of our findings with results of previous studies, our analyses were performed based on 3 age groups (≤18, 19-64, and ≥65 years). We used a binary distribution for the dependent variable in the model with a logit link.30 Some unobserved patient characteristics such as health care–seeking behavior may simultaneously affect the continuity of care and health care outcomes; therefore, we fit random intercept logistic regression models using a longitudinal technique that could consider the unobserved time-invariant characteristics for patients. Covariates in this model included time-variant variables (age, number of physician visits in the previous year, the likelihood of hospitalizations in the previous year, and Charlson index score), time-invariant variables (sex and low-income status), and the 5-year dummy variables for time effect. For the random subject effect, we calculated the intraclass correlation coefficient, which represented the proportion of variance accounted for by the subject level.30 Our analyses were performed using commercially available statistical software (SAS, version 9.1.3; SAS Institute, Cary, North Carolina; and STATA 9; StataCorp LP, College Station, Texas).
The characteristics of the study subjects in 2001 are given in Table 1, categorized in age groups of 18 years or younger, 19 to 64 years, and 65 years or older. There were fewer male (36.1%) than female (63.9%) subjects in the middle age group. The mean numbers of physician visits in the previous year among the younger, middle, and older age groups were 19, 17, and 27 visits, respectively. The proportions of subjects hospitalized in the previous year were 7.8%, 9.0%, and 16.6% among the younger, middle, and older age groups, respectively. We also noted that 95.8%, 81.3%, and 48.8% of subjects among the younger, middle, and older age groups, respectively, had a Charlson comorbidity index score of 0. This indicated that subjects in the older group were more likely to have comorbid conditions than those in the younger and middle age groups.
The variables of interest are given in Table 2. The COCI values tended to decrease from 2001 to 2006 in the younger group; however, the COCI values remained stable in the middle and older age groups. Different trends were observed among the 3 age groups for avoidable hospitalization and for hospital admission for any condition. For the younger group, the rates of avoidable hospitalization decreased over time (from 4.2% to 1.4%), as did hospital admission for any condition over time (from 6.8% to 3.3%). For the middle age group, the hospitalization rates remained stable during these years. For the older group, the rates of avoidable hospitalization increased (from 4.6% to 8.7%), as did hospital admission for any condition (from 16.9% to 24.2%).
Significant dose-response trends for the effects of continuity of care on avoidable hospitalizations were observed for the 3 age groups (Table 3). The likelihood of avoidable hospitalization for subjects with a high or medium COCI was lower than that for subjects with a low COCI among all 3 age groups. With regard to the random subject effect, the intraclass correlation coefficient ranged from 27.04% to 32.78%, indicating that a large proportion of the variance existed at the subject level.
The COCI was associated with hospital admission for any condition in all 3 age groups (Table 4). The likelihood of hospital admission for subjects with a high or medium COCI was lower than that for subjects with a low COCI in all 3 age groups. Regarding the random subject effect, the intraclass correlation coefficient ranged from 20.16% to 26.59%, indicating that a sizable proportion of the variance was attributable to the subjects.
We conducted several sensitivity analyses to improve the robustness of the study. First, we performed analyses using various COCI cutoff points, including continuous scale, median, tertiles, quartiles, quintiles, and slot approach (0-33, 35-65, and 70-100). These analyses indicated that the results were robust to how the COCI was coded (data available in appendixes on request from the author). Second, we used the usual provider of care index and the sequential continuity index, in place of the COCI, to examine the association between continuity of care and avoidable hospitalization, as well as hospital admission for any condition. The results were similar to those obtained using the COCI (data available in appendixes on request from the author). Third, we examined whether the effect of the COCI on avoidable hospitalization and on hospital admission for any condition was independent of the number of physician visits in the previous year. We stratified the number of physician visits into 3 tertiles (groups with low, medium, and high numbers of visits) and found similar results within each group (data available in appendixes on request from the author).
Additional sensitivity analyses were conducted to examine the stability of the results. First, we performed analyses using the stratified Charlson index score (score of 0 vs ≥1) and found partly similar results. Second, we conducted analyses by excluding subjects who had fewer than 2 physician visits in any of the years between 2000 and 2006 and found similar results. Third, we performed analyses including subjects with excluded diagnoses, and the results were similar to previous findings. Supplementary tables of the results are available on request from the author.
The primary objective of this study was to examine the effect of continuity of care on avoidable hospitalization under a health care system with easy access to care. The results indicated that better continuity of care was associated with lower likelihood of avoidable hospitalization and hospital admission for any condition among all 3 age groups. Therefore, avoidable hospitalization could be reduced by increasing the continuity of care in a health care system with a high degree of access to care.
Many previous studies have examined the relationship between patients with a regular source of care and avoidable hospitalization; however, the findings have been inconclusive. One study42 reported that patients with a regular source of care experienced fewer hospitalizations, although other studies43,44 did not support this conclusion. Two studies14,28 examined the association between continuity of care and avoidable hospitalization using a cross-sectional design; however, the findings of these studies were also controversial. Gill and Mainous14 focused on a younger population and demonstrated only partial support for a positive association, while Menec et al28 focused on an older population and found a significantly positive association. In the present study, better continuity of care was associated with fewer avoidable hospitalizations among all 3 age groups.
Since 1995, the compulsory health insurance program has significantly increased accessibility to care and utilization of health care services for residents of Taiwan.45 A 2008 study46 reported that the NHI has narrowed, to some extent, the disparity of life expectancy among groups with varying degrees of health. The present study revealed that residents with better continuity of care were less likely to need hospital services under universal health care coverage. Improving continuity of care for patients should be an objective for health care reform in Taiwan.
Our study has several limitations. First, the study used NHI claims data, which do not contain information about self-paid physician visits. However, we assume that the number of these types of visits would be minimal. Second, this study did not include some measures of patient characteristics (such as education or household income) that may simultaneously affect continuity of care and outcome measures. Because our study used a longitudinal design, this concern might be mitigated; indeed, the study design accounted for unobserved time-invariant patient characteristics.30 Third, findings from this study could be generalizable to countries without a formal referral system but not to countries in which most citizens seek care first through a primary care physician, such as in the United States or in Canada.
Using comprehensive data from a health care system with universal coverage and with freedom to choose physicians, our study revealed that better continuity of care is associated with fewer avoidable hospitalizations and with fewer hospital admissions for any condition. Evidence from the longitudinal analysis indicated a positive association for every age group when measured using various COCIs. Although most previous studies were conducted in the United States or in Canada, our findings suggest that continuity of care is beneficial in a health care system without a referral arrangement. This study provides evidence for policy makers and concludes that improvement in continuity of care is an appropriate path to follow in a universal coverage health care system.
Correspondence: Shou-Hsia Cheng, PhD, Institute of Health Policy and Management, College of Public Health, National Taiwan University, Room 618, 17 Hsu-Chow Rd, Taipei, Taiwan 100 (firstname.lastname@example.org).
Accepted for Publication: March 5, 2010.
Author Contributions:Study concept and design: Cheng. Acquisition of data: Cheng. Analysis and interpretation of data: Cheng, Chen, and Hou. Drafting of the manuscript: Cheng. Critical revision of the manuscript for important intellectual content: Cheng, Chen, and Hou. Statistical analysis: Cheng and Chen. Obtained funding: Cheng. Administrative, technical, and material support: Chen and Hou. Study supervision: Cheng.
Financial Disclosure: None reported.
Funding/Support: This study was supported by grant NSC98-2410-H-002-054 from the National Science Council in Taiwan (Dr Cheng).
Additional Contributions: The National Health Research Institutes provided the data sets for our study.