2.3. Variable definitions

The Agency for Healthcare Research and Quality provided a definition for prevention quality indicators,^[31] identifying hospital admission for asthma using a code pertaining to the main diagnosis (ICD-9-CM code, 493.xx). Therefore, in this study, avoidable hospitalization was defined as an event that occurred during the outcome period. The follow-up time was defined as the number of days from the date of the end of the COC period to the occurrence of the avoidable hospitalization for asthma. However, if no avoidable hospitalization for asthma occurred, the patient was censored at the end of the outcome period.

The COCI score, used as the independent variable, was measured during the COC period. The score ranges from 0 to 1 (values close to 1 represent a greater COC) and measures the dispersion of contact between patient and physician.^[32] The COCI has been widely adopted in studies based on health care claim databases,^{[4,9,14,15,19,25]} because it is less sensitive to the number of physician visits and is suitable for application to a large amount of outpatient visit data.^[25]

The general formula is 

where N is the total number of physician visits, n_i is the number of visits to the i_th physician, and k is the total number of physicians. In this study, the total number of physician visits (N) and the number of visits to a given physician (n_i) included ambulatory claims for asthma as the major diagnosis. The patients were categorized into three groups based on the first and third quartile value of COCI as follows: low, medium, and high.^[25,29]

Confounding factors were identified in 3 mutually exclusive periods. First, variables measured on the index date were sex,^[9,33,34] age,^[9,33,34] and insurance premium (<20,000 NTD, 20,000–40,000 NTD, and ≥40,000 NTD), which was used as a proxy indicator of income status.^[35] Second, the variables measured in the year prior to the index date included chronic obstructive pulmonary disease (COPD) (ICD-9-CM codes 491, 492, or 496),^[36] pulmonary-related diseases (ICD-9-CM codes 490, 494, or 495),^[36] diabetes mellitus (DM) (ICD-9-CM code 250),^[36] the Charlson comorbidity index (CCI), and the number of asthma-related ED visits. The CCI and number of asthma-related ED visits were used as proxy indicators of health status^[9] and disease severity.^[25] The CCI score contains 17 categories of comorbid conditions defined by ICD-9-CM codes, and it is calculated according to enhanced ICD-9-CM coding algorithms.^[37] Third, because patients’ health status during the COC period also may impact the outcome, the number of asthma-related ambulatory visits was used as a proxy for patients’ health status.^[9,19]

In this study, descriptive statistical analysis was used to present the distribution of patient characteristics. In addition, chi-squared tests and one-way analysis of variance were used to analyze associations between patient characteristics and COC.

The Cox regression model assumes that the ratio of the hazards of two subjects is the same at all times; in this study, the scaled Schoenfeld residual was used to test whether this assumption was valid.^[38] With the valid proportional hazard assumption (P = 0.7921), the Cox regression model was applied to examine the association between COC levels and the risk of avoidable hospitalization for asthma among elderly patients. Multivariate analysis was used to calculate adjusted hazard ratios (aHRs) by adjusting for sex, age, insurance premium, COPD, pulmonary-related diseases, DM, CCI, number of asthma-related ED visits, and number of asthma-related ambulatory visits. The variance inflation factor (VIF) is used to detect the presence of multicollinearity; a value greater than 10 indicates the severity of multicollinearity in the regression model. In our model, no multicollinearity was represented by a VIF of less than 5 in each variable. The aHR stratified by each variable was then calculated to investigate the effect of COC levels on the extent to which avoidable hospitalization for asthma.

Two-sided criteria with P values of less than 0.05 were considered to be statistically significant in this study. All statistical analyses and data management were conducted using SAS software version 9.4 (SAS Institute, Cary, NC).