Study design
This study aimed to examine the changes in medical utilization after the expansion of the national disabled support service in Korea using a difference-in-difference (DID) approach. As the national disabled support service expanded the implementation of services for people with severe disabilities, starting on January 1, 2013, the pre-and post-intervention periods were from January 1, 2010, to December 31, 2012, and January 1, 2013, to May 31, 2015, respectively. A prerequisite for DID estimation is that the trends in the test and control groups should be parallel, which can be partially verified using graphical evidence16. Figure 1 and Supplementary Table 1 shows that the treatment (severely disabled) and control (mildly disabled) groups satisfied the basis for a common trend.
Unadjusted trend in hospital visits due to hypertension.
Data
This study used sample cohort data from the National Health Insurance Service (NHIS). The data contained socioeconomic qualification variables (including death and disability), the status of medical resource utilization, and the clinical status of approximately one million people. The data consists of a stratified sample representing 2% of the entire national population, stratified by gender, age, and region, and is derived from health insurance claims data. The classification of severe and mild disabilities was available as a sociodemographic variable in the dataset. The study population comprised individuals with disabilities and hypertension (ICD-code: I10–I15) between 20 and 64 years old (PAS program service target). The total number of observations were 17,126, recorded from January 1, 2010, to May 31, 2015. The final study sample consisted of 38,499 individuals, including 15,103 in the treatment group and 23,396 in the control group.
Variables
In this study, the main outcome variable was defined as the number of hospital visits due to hypertension during the entire observation period, including both pre- and post-intervention. This outcome was selected to assess changes in hospital utilization associated with hypertension before and after the intervention. The independent variables included treatment status, gender, age, health insurance type, disability type, the Charlson Comorbidity Index (CCI), and a time variable.
The treatment variable distinguishes between the treatment (severely disabled) and control (mildly disabled) groups allowing for the evaluation of intervention effects. Gender and age were included to capture demographic characteristics, with age treated as a continuous variable. Health insurance type was categorized into employer-sponsored, local subscribers, and Medical Aid recipients. Disability type was classified into physical disabilities, brain lesions, visual impairment, hearing impairment, and other disabilities. The Charlson Comorbidity Index (CCI) was incorporated as a measure of comorbidity to adjust for the severity of concurrent health conditions. Lastly, a time variable was included to distinguish the periods before and after the intervention, which is essential for the Difference-in-Differences (DID) analysis used to estimate the causal effect of the intervention.
Statistical analysis
In this research, the DID methodology is employed to estimate the causal effect of policy interventions by analyzing changes in outcomes across different groups. The DID approach controls for unobserved time-invariant confounders by comparing the difference in outcomes before and after the intervention for the two groups. By modeling the pre-intervention outcome, an appropriate counterfactual is established, representing what would have occurred in the absence of the policy intervention. To estimate the average treatment effect, given the outcome variable is count-based, a negative binomial distribution model is utilized. The negative binomial model includes an additional parameter to handle this overdispersion, making it more suitable for modeling count data compared to a Poisson regression. This ensures a more accurate estimation of the effect size while accounting for the inherent variability in the data. The model incorporates key covariates such as age, gender, severity of disability, and comorbidities to control for potential confounding factors. Interaction terms between the treatment group and the post-intervention period are included to estimate the DID effect, representing the average treatment effect. All analyses were conducted using the SAS software, version 9.4 (Cary, NC, USA).
link