3.1 Data Source and the Selection of Sample
The data used in this study are obtained from a longitudinal claims data set that is created through a two-stage process. First, this data set randomly selects a sample of one million individuals from the registry of NHI beneficiaries in 2005.4 Second, this sample is then merged with insurance claim files that record all medical utilization claims for all individuals in the sample from 1996 to 2007. Hereafter, we refer to this data set as the NHI sampling claims data. The NHI sampling claims data are publicly available through the National Health Research Institute. This data set contains detailed records on the utilization of personal health care services, including outpatient visits, hospital admissions, and prescription drugs. The advantage of this data set is that all the medical utilization data can be linked together for the same patient. Thus, this data set allows us to trace the same individual's inpatient expenditures between 1996 and 2007. In this data set, each inpatient service claim documents the date of admission and discharge, age and sex of the patient, diagnosis, treatment, hospital characteristics, and, most importantly, the claimed expenditures.5
To identify the end-of-life patients who died during our study period, we utilize the NHI enrollment records that accompany the claims data.6 At any given point in time, an individual is coded as one of the three statuses: “enrolled”, “suspended”, or “withdrawn”. If the individual dies, his/her status will be coded as “withdrawn”. However, death is not the only reason for being coded as “withdrawn”. An individual can also be coded in this way if he/she is (i) incarcerated for longer than two months; (ii) missing for six months or longer; or (iii) ineligible for the NHI due to emigration. Unfortunately, the exact reason for being coded as “withdrawn” is not provided in the data.
However, we argue that the “withdrawn” cases due to nondeath reasons mostly occur to younger cohorts under the age of 50. To see this, we calculate the sample age-specific “death rates” in 2005 by treating all the “withdrawn” cases within each 10-year age group as dead and dividing them by their corresponding group size. We then compare them with the population age-specific death rates, which are obtained from the Statistical Yearbook of Interior 2005 published by the Ministry of Interior.7 As shown in Figure 1, the two death rates are almost identical, except for ages 10–49. In particular, the sample “death rate” is higher than the population rate over ages 10–49, implying that some of the “withdrawn” cases over this age range were actually alive in 2005. Therefore, in order to reduce the misidentification of those deceased, we restrict our analysis sample to those who were aged 50–99 in 2005 and turned into “withdrawn” during 2005–2007. Based on this definition, our analysis sample includes 11 863 individuals, whom we define as end-of-life patients and who presumably died during 2005–2007.
In Figure 2, we plot the mean NHI hospital expenditure and death rate by age. As shown, most deaths and hospital expenditures occurred for those aged 50 and older, indicating that the concern for a potential sample selection bias resulting from restricting the age to 50–99 in our study should be minimal.
We then backwardly trace their end-of-life hospital expenditures for 24 months, which are further grouped into eight quarters. More specifically, we define the day on which an individual's insurance status turns into “withdrawn” as the day of death and then trace his/her hospital expenditures backwardly.8 In the end, we have 48 971 observations of hospital stays during the last eight quarters before death for the 11 863 end-of-life patients.
3.2 Descriptive Analysis
Table 1 provides summary statistics for our sample. Among the 11 863 end-of-life patients, 60% are male and 40% are female. Their mean age is 74 in 2005. Of the 48 971 observations for hospital stays, the average expenditure is 41 357 New Taiwan Dollars (NTD), and the average length of stay is 9.58 days. 67% of stays are in private hospitals, and only 33% are in public hospitals. The market shares are consistent with the nationally aggregate statistics, suggesting that our sample is a good representation for the whole population.
Table 1. Summary statistics of the end-of-life patients
| ||Mean||Std. dev.||Min||Max|
|Age in 2005||74.14||11.04||50||99|
|Hospital characteristics|| || || || |
In Taiwan, the regulatory agency classifies hospitals into three accreditation levels: (i) medical centers; (ii) metropolitan hospitals; and (iii) community hospitals. The accreditation level is closely correlated with the size of the hospitals. In our sample, 34% of the hospital stays are in medical centers while metropolitan and community hospitals account for 40% and 26% of the hospital stays respectively. If we look at the distribution of the hospital stays over the last eight quarters, we find that the closer the individual is to death, the more hospital stays that occur. More specifically, only 6% of the stays occur in the seventh quarter prior to death, and the percentage increases all the way up to 40% in the last quarter.9
Before we move to our empirical test, it is helpful to briefly examine the dynamic patterns of the end-of-life hospital expenditures and related hospital characteristics. Figure 3 illustrates the average quarterly hospital expenditure (the solid line) for the last eight quarters prior to death.10 (Note that quarter “0” denotes the quarter that ended in death, quarter “–1” is the first quarter before the death quarter, and so forth.) Most of the time, the end-of-life hospital expenditure increases gradually and remains under 50 000 NTD until the terminal quarter when it suddenly soars more than threefold.11 One of the plausible explanations for this dramatic rise is the rapidly deteriorating health in the terminal quarter, because the average Charlson's comorbidity index (the dashed line; the higher the index, the more serious the morbidity), which measures the intensity of illness based on the ICD-9 code, parallels the hospital expenditure, and also jumps more than three times in the last quarter.12
In terms of hospital characteristics, we are mainly interested in the distribution of hospital expenditure across the ownership form and the accreditation level of the hospital. Figure 4 shows the quarterly hospital choices in the final stage of life based on these two characteristics. We define the quarterly choice in terms of the percentage of days spent in a certain type of hospital in each quarter.13 The left panel shows that metropolitan and community hospitals account for 30% of the hospital choices at the beginning of our observation period (the eighth quarter prior to death), while medical centers account for 40%. Then, the share of community hospitals decreases over time, while the share of higher-level hospitals increases. This suggests that patients tend to move to larger hospitals (in terms of the accreditation level) as they approach death. By contrast, the choice between public and private hospitals does not seem to respond to health deterioration. As shown in the right panel, private hospitals constantly account for roughly two-thirds of the total hospital days over the last eight quarters, while the market share of hospital days in public hospitals remains fairly stable in the range of one-third.
Figure 4. The distribution of hospital days in the final stage of life across the accreditation level and ownership form
Download figure to PowerPoint
3.3 Empirical Specification
Our benchmark empirical model is specified as follows:
where yijq is hospital expenditure or days of hospital stay for individual i's jth hospital admission in the qth quarter relative to death; Private is a dummy variable denoting a private hospital; Metropolitan is a dummy variable denoting a metropolitan hospital; Medicalcenter is a dummy variable denoting a medical center; X is a vector of control variables, including the patient's age, sex, Charlson's comorbidity index, and a full set of regional dummies and quarter dummies, which indicate the discharge quarter of each hospital admission; B is a vector of coefficients corresponding to the variables in X; and εijq is an error term.
Equation (1) is estimated by both the ordinary least squares (OLS) method and quantile regression method. The former method estimates the effect on the mean of the outcomes, while the latter estimates the effect on different quantiles of the outcomes. In our data, an individual may have multiple hospital admissions even within the same quarter. We thus use robust standard errors clustered at the individual level for statistical inference. The standard errors in the quantile regressions are calculated by using bootstrapping method. In addition to the benchmark model, we also use an alternative specification that adds to Equation (1) the interaction terms between the ownership and accreditation level. This alternative specification allows us to test if the private–public differences vary across accreditation levels.
We recognize the fact that the patient's choice of hospital is highly correlated with the severity and complexity of his/her medical condition. For example, magnetic resonance imaging (MRI) is often used to detect structural anomalies of the body for patients with severe medical conditions, such as a brain tumor. However, MRI machines are only available at large hospitals, such as metropolitan hospitals and medical centers in Taiwan, and MRI scans are rather expensive. To control for the severity and complexity of medical conditions, we calculate Charlson's comorbidity index for each admission and add it as a control variable in our regressions.
It is possible that the patient's choice of hospital is also associated with proximity to a certain type of hospital. For example, if sicker people tend to live closer to larger hospitals, failing to control for such proximity could bias our estimates. Although our regressions include regional dummies, which divide Taiwan into six regions, they may be too broad to fully control for the proximity to hospitals. Unfortunately, we do not have more detailed information about patients' residence.