Data
The study used data from five waves of the Chinese Longitudinal Healthy Longevity Survey (CLHLS) in 2002, 2005, 2008/2009 (hereafter referred to as 2008), 2011/2012 (hereafter as 2011), and 2014. The first two waves of the CLHLS (1998 and 2000) were not used because several key variables were not included and the sample only included adults aged 80 and older at baseline. The CLHLS was conducted in a randomly selected 50% of counties/cities in 22 Han-dominated provinces [26]. The total population of the 22 provinces accounts for approximately 87% of the total population in China in 2010. The CLHLS was designed to oversample very old adults. Thus, sample weights were constructed based on the age-sex-residence-specific distributions of the populations in the sampled provinces and were released with the dataset [27]. The CLHLS has an overall response rate of more than 90% across waves [5, 26].
In 2002, the CLHLS recruited nearly 16,000 older adults. In addition to tracking these older adults, the CLHLS recruited more than 7300 new sample members in 2005, nearly 9000 new sample members in 2008, and nearly 1300 newly old adults in 2011. Consistent with prior research [5, 27], we pooled the five waves of CLHLS data to obtain robust estimates. Specifically, we examined how those who were interviewed in 2005, 2008, and 2011 waves were at risk of dying at 2008, 2011, and 2014 waves. Excluding those who were lost to follow-up, the final analytic sample included 27,794 adults aged 65 and older who were interviewed in 2002–2011.
Access to healthcare
The National Academies of Sciences, Engineering, and Medicine of the United States defines healthcare (or healthcare) as wide range of services that include preventative care, chronic disease management, emergency services, mental health services, dental care, and other community services that promote health over the lifespan [28]. In the CLHLS, healthcare was defined in the same context and included several questions on the respondents’ access to such healthcare services. In this study, we used the overall question of whether respondents reported having adequate access to healthcare (medical care) services when needed (yes vs. no). On average, about 7.8% of women and 7.4% of men reported having inadequate access to healthcare over the 2002–2014 study period—with some variation based on rural and urban residence (9.0% vs. 4.7%, respectively). To reduce missing data on this question, the CLHLS included responses from proxies (i.e., family members, etc.) to ascertain information on access to healthcare for participants who may be too sick to provide such information. Approximately 23% of sampled older adults used a proxy—with a weighted proportion of only 4.8% using a proxy (because most proxies were for adults aged 90 or older).
Covariates
The analyses adjusted for a wide range of covariates that have been shown to be associated with either access to healthcare or mortality at older ages [2, 5]. The analyses included a number of demographic, socioeconomic, family/social support, behavioral, and health-related factors that have been shown to be associated with either access to healthcare or mortality [2, 5]. Demographic factors included age (in years), sex, and residence (urban vs. rural). Socioeconomic factors included ethnicity (Han vs. non-Han), years of schooling (0, 1–6, or 7+ years), primary lifetime occupation (white collar vs. other occupations), economic independence (having a retirement wage/pension and/or having own earnings vs. no), and having a health insurance (yes vs. no). Having health insurance referred to whether the respondent was covered by any of the following insurance programs: in addition to abovementioned public free medical scheme, Urban Employer-sponsored Medical Scheme (UEMS), the Urban Resident Medical Scheme (URMS), and the New Cooperative Medical Scheme (NCMS), the CLHLS collected data on whether a respondent purchased severe diseases insurance program. Family/social support included current marital status (married vs. not married) and whether the respondent had close proximity to their children (either co-residing with a child or having a child living in the same village [yes vs. no]).
Health behaviors included whether the respondent was currently smoking (yes vs. no). Health status included (1) physical function, (2) chronic disease, and (3) cognitive impairment. Physical function was measured by disability of activities of daily living (ADLs). ADLs consisted of six activities with three responses: “able to do without help,” “need some help,” and “need full help.” The six items included: (a) bathing, (b) dressing, (c) indoor transferring, (d) toileting, (e) eating, and (f) continence [26]. The six items of ADLs were adopted from the Katz scale and used similar response categories as IADLs [23]. We categorized respondents as ADL disabled (coded as 1) if they reported needing any help in performing any of the six items (otherwise coded as 0). Chronic diseases were ascertained in the CLHLS from 22 major conditions, including hypertension, diabetes, cardiovascular disease, stroke, cancer, and so on. The conditions were self-reported and more than 90% were reported as being diagnosed by a physician. A dichotomous measure for any chronic disease (yes vs. no) was included in the analyses. Cognitive function was measured using the Mini-mental State Examination (MMSE) that includes six domains of cognition—i.e., orientation, reaction, calculation, short-term memory, naming, and language—with a total score of 30. The MMSE items in the CLHLS were adopted from the Folstein MMSE scale [26]. Respondents were categorized as cognitively impaired if his/her MMSE score was below 24 [26]. Given the low level of educational attainment among older adults in China, we assessed alternative criteria (e.g., score of 18) for those with no education to test the sensitivity of different cut-points for defining cognitive impairment. Results were very similar to those presented here and are available upon request. As documented in prior research [26], the Chinese version of MMSE used in the CLHLS was culturally translated from the international standard version of the MMSE questionnaire. The validity and the reliability of the MMSE measure were also carefully tested in pilot surveys and verified in each wave of the CLHLS [26]. With the exceptions of sex, ethnicity, education, and primary lifetime occupation, all measures included in the models were time-varying covariates.
Mortality and life expectancy
All-cause mortality was defined as whether the respondent was dead or alive at the time of the 2014 survey (event measure). Exposure to mortality risk (duration measure) was ascertained by the number of days alive from the date of the baseline interview to the date of death (for deceased) or the date of the 2014 survey (for survivors). The date of death for deceased respondents was gathered from official death certificates whenever available; otherwise, the information was collected from next-of-kin and confirmed by the local residential committee. From 2002 to 2014, approximately 57.5% of sampled individuals died (it would be 68.3% if those lost to follow-up were excluded) and about 15.8% of respondents were lost to follow-up. The corresponding weighted numbers were 26.7% for death (it would be 31.8% if those lost to follow-up were excluded) and 16.4% for loss to follow. A number of studies have previously documented the high quality of mortality data in the CLHLS [26]. To ensure the reliability of our results, we compared the age-sex-specific death rates from the 2002–2014 CLHLS data to the rates obtained from the 2000 and 2010 Census [26, 29, 30] and the World Population Prospects (WPP) provided by the United Nations Population Division [31]. The high degree of concordance among the death rates from these sources is reported in Figures A1 and A2 in the Additional file 1.
The age-specific death rates were obtained by multiplying the number of time units in the survival analysis relative to a year to the hazard rate estimated from the following formula based on an exponential distribution of the hazard function:
$$ h(x)=\exp \left(\beta_0+\beta_1x_1+\beta_2x_2+\beta_3x_3+\beta_4x_4+\sum \limits_i=1^K\left(\gamma_iz_i\right)\right), $$
where β0 is the constant; exp.(β0) is also referred to as the baseline function. β1 is the coefficient associated with adequate access to healthcare (x1); β2 is the coefficient for age(x2); and β3 is the coefficient for sex (x3). β4 is the coefficient for urban-rural residence. The four variables x1, x2, x3,and x4 are all dichotomous variables. For a given sex (x3), the mean of x4 is used when the age-specific mortality rates do not distinguish urban-rural differences. In a similar vein, for urban or rural areas (x4), the mean of x3 is used when the age-specific mortality rates do not distinguish gender differences. γi is the coefficient of covariate zi. K is total number of covariates.
In the present study, we estimated the age-sex-specific and the age-residence-specific death rates according to access to healthcare by supplying the coefficients to the different categories of age, sex, urban-rural residence, and access to healthcare, respectively, along with the coefficients of the covariates in the models to their means. We then multiplied 365 to h(x) to estimate the annual death rate in that the analytical unit of time in our survival analysis is day.
Once the age-sex-specific or age-residence-specific death rates were estimated stratified by adequate and inadequate access to healthcare, life expectancy was calculated for adequate access and for inadequate access stratified by sex or by urban-rural residence using common/basic demographic methods for life tables [32]. The difference in life expectancy between adequate access and inadequate access to healthcare is the increase associated with adequate access or the contribution to life expectancy made by adequate access to healthcare while taking into account the covariates.
Analytical strategies
Three exponential parametric hazard models were used to examine the difference in life expectancy between adequate access and inadequate access to healthcare or the increase in life expectancy associated with adequate access and how the difference (or advantage) was altered by sex and by urban-rural residence when a wide range of confounding factors were taken into account. Model I included access to healthcare in addition to age, sex, and urban-rural residence. Model II added socioeconomic status to Model I. Model III included all study covariates. Supplementary analyses were also performed to examine whether each of the other three sets of confounding factors (i.e., family/social support, health behaviors, and health conditions) influenced the increase in life expectancy associated with adequate access to healthcare (see Additional file 1: Table A). To ensure the accuracy of model-based death rates in the hazard models, sample weights were always applied. Also, given the different patterns of healthcare utilization, mortality trajectories, and differences by sex and urban-rural residence [33,34,35,36,37,38,39], we included interactions between sex and access to healthcare and between urban-rural residence and access to healthcare.
Several sensitivity analyses were also conducted. First, preliminary analyses showed that the results from the parametric hazard models were nearly identical to the estimates from Cox proportional hazard models. Second, preliminary analyses also showed that the results were consistent when assuming that persons lost to follow-up had the same survival status (and length of survival in the survey interval) as those with known survival status if they had the same demographics, psychosocial characteristics, and health conditions. Third, we found no significant interactions between access to healthcare and survey year in our assessment of possible temporal changes in the associations during the 2002–2014 study period. Missing data among all covariates was generally low (less than 2%) and ultimately dropped from the final analyses. Alternative approaches were also assessed (e.g., multiple imputation, mean/mode imputation, etc.) and the results were nearly identical.
link