Conceptual model
This study is based on the Andersen and Newman Framework of Health Service Utilisation. The Andersen model is used in analysing the factors that affect health service utilisation (including hospitalisation, doctor visits, dental care, etc.). The initial Andersen model includes predisposing characteristics, enabling resources and needs [18]. Amongst them, predisposing characteristics refer to the demographic characteristics, social structure and health beliefs of the population who tend to use medical services before the occurrence of a disease. Enabling resources refer to the individual's ability to obtain medical services and the accessibility of medical service resources. Need refers to the direct factors of health service utilisation, including perceived illness measures and evaluated illness measures [19,20]. The revised Andersen model believes that individual medical behaviour is the result of the interaction of contextual characteristics, individual characteristics and medical results. Since its creation, the Anderson model has been widely used in researches related to health expenditures and health service utilisation and is therefore an authoritative research model for medical and health services [20,21]. Thus, this study selected variables based on this model. In addition, informal care provided by family members was incorporated into the regression model as an enabling component by referring to the practices of previous literatures [22].
Data and Sample
The data used in this study were derived from the China Longitudinal Study of Health and Retirement (CHARLS). CHARLS is a national survey aimed at providing comprehensive and high-quality data, such as demographic background, family characteristics, health behaviours and conditions and retirement information, of residents aged 45 years and older. A multi-stage stratified proportional probability sampling design was used to randomly select households from 450 villages and residential communities in 150 counties and regions in 28 provinces [23]. The wave 3 study conducted in 2015 and the wave 4 study in 2018 were used for analysis in this paper. A balanced panel data with 5704 participants was formed according to the following criteria: 1) 65 years and older and 2) information about outpatient, hospitalisation and family care were provided in both waves. The details of the sampling process are shown in Fig. 1. The respondents were categorised as urban and rural residents according to the family residential area defined by the National Bureau of Statistics of China.
Dependent variables
This paper analysed two types of health care utilisation: outpatient and inpatient services. The respondents were asked about how many times they had been visited by medical institutions in the last month and how many times they had received inpatient care during the past year. Outpatient and inpatient visits were used to evaluate the utilisation of health services in the elderly in this study. Besides, the dichotomous variable, namely, whether or not the participants use outpatient or inpatient services, was also used to examine health service utilisation.
Key variables
Interviewees were asked who helped them with their daily life activities in CHARLS. The 10 options for this question are: a) Spouse; b) Father, Mother, Father-in-law or Mother-in-law; c) Children, Children's spouses, Grandson or Granddaughter; d) Sibling, Brother-in-law, Sister-in-law, Sibling of spouse, Children of sibling, Brother-in-law of spouse, Sister-in-law of spouse, Children of brother-in-law or Children of sister-in-law; e) Other relatives; f) Paid helper (such as nanny); g) Volunteer; h) Employee(s) of facility; i) Community; j) Others. The participants who selected one or more of options a–e were considered to have accepted informal care from family members [13,14]. In addition, the interviewees were asked about the number of days in the past month that family members had assisted them.
In this research, informal care was examined by two aspects:
a) Whether the respondents have received informal care from family members. The answers were coded as ‘Yes’ and ‘No.’
b) The number of days of informal care that respondents received in a month. The answers were coded as ‘none,’ ‘less than 15 days,’ ‘between 15 and 29 days’ and ‘30 days or more.’
The Control variables
Control variables were selected in accordance with predisposing characteristic, enabling resources and needs according to the Andersen model [24–27]. The following individual-level characteristics were considered as control variables (Table 1).
Table 1: Definition/codes of the control variables.
|
Variables
|
Codes/definition
|
Predisposing Characteristics
|
Gender
|
0=Male; 1=Female
|
Age
|
Continuous variable
|
Marital statues
|
0=Single; 1=Partnered
|
Education
|
1 = Illiterate; 2 = Primary school and lower; 3 = Junior middle school; 4 = Senior middle school and higher
|
Enabling Resources
|
Medical insurance
|
0=None; 2=Yes
|
Pension
|
Whether received any pension or not: 1=No; 2=Yes
|
Financial support from children
|
Financial support received from children: 0=None; 1=0-2000yuan; 2=2000-5000yuan; 3=5000-10000yuan; 4=10000yuan or more
|
Smoke
|
0=No; 2=Yes
|
Drink
|
0=No; 2=Yes
|
Need
|
Chronic diseases
|
0=None; 1=Yes
|
Number of ADL limitations
|
Range from 0 to 6
|
Number of IADL limitations
|
Range from 0 to 6
|
Self-rated health
|
1=very good; 2=good; 3=fair; 4=poor; 5=very poor
|
|
Year
|
2015; 2018
|
Note: ADL: activities of daily living; IADL: instrument activities of daily living.
Statistical analysis
The dependent variables of this study are outpatient visits and inpatient visits, which are count variables; therefore, the analysis methods considered were Poisson regression and negative binomial regression [28]. However, for a good-fitting model, Poisson regression requires the mean and variance of the dependent variable to be equal. Substantial departures from this measure may indicate a problem with model specification and also suggest that the estimated standard errors may be downwardly biased. In comparison, negative binomial regression is more suitable when overdispersion occurs [29]. This study examined whether the two dependent variables have overdispersal through the likelihood-ratio test of alpha value. The results show that the alpha value is significantly not 0 (95% confidence interval [CI]: 4.75–6.03; 95% CI: 1.41–2.09), which means negative binomial regression should be used in this study instead of Poisson regression [30,31]. Subsequently, this study conducted Hausman test (both P<0.001) and LR test (both P<0.001) using inpatient and outpatient visits as the outcome variables, respectively. The results showed that the fixed effect is more applicable. Therefore, fixed-effect negative binomial regression for panel data was conducted to estimate the impact of informal care on the health service utilisation of the elderly. [32]. Cluster-robust standard errors were used to control heteroscedasticity [33]. The specification of the model was as follows:
ln(λit) = β1careit+β2Pit+β3Eit+β4Nit+μit+ε,
where λit represents the outpatient or inpatient visits of individual i in period t; careit represents the time of informal care obtained by individual i in period t; and Pit, Eit, and Nit represent the control variables related to predisposing characteristics, enabling resources and needs in the Anderson Model, respectively. μit is the individual fixed effect, and ε is the error term. Incidence rate ratio (IRR) was calculated to facilitate the interpretation of the results. [34].
Model uncertainty is ubiquitous in social science. Thus, substitution dependent variables and substitution regression model were used to test the robustness [35]. The dichotomous variable, namely, whether outpatient and inpatient services were used, was regarded as the outcome variable. A fixed-effect binary choice model for panel data was used for sensitivity test.