Healthy behaviours and productive activities among Thai older adults: A repeated cross-sectional analysis

Healthy behaviours and productive activities among Thai older adults: A repeated cross-sectional analysis

Social Science & Medicine 213 (2018) 12–19 Contents lists available at ScienceDirect Social Science & Medicine journal homepage: www.elsevier.com/lo...

313KB Sizes 0 Downloads 26 Views

Social Science & Medicine 213 (2018) 12–19

Contents lists available at ScienceDirect

Social Science & Medicine journal homepage: www.elsevier.com/locate/socscimed

Healthy behaviours and productive activities among Thai older adults: A repeated cross-sectional analysis

T

Nopphol Witvorapong Centre for Health Economics, Faculty of Economics, Chulalongkorn University, Pathumwan, Bangkok, 10330, Thailand

A R T I C LE I N FO

A B S T R A C T

Keywords: Thailand Social participation Labour participation Lifestyle Elderly Multivariate probit

Based on a nationally representative repeated cross-sectional sample of older individuals from the 2007 and 2011 Surveys of Older Persons in Thailand (n = 50,138, with the participation rate of 95.79%), this study investigates the extent to which healthy behaviours are interrelated with productive activities in old age. Healthy behaviours were represented by alcohol abstinence, tobacco abstinence, physical exercise, and consumption of a nutritious diet, encompassing all major lifestyle choices that could lower mortality risks among the general population. Productive activities were represented by social participation and labour participation, consistent with the paradigms of Active and Productive Ageing promoted by the World Health Organization. A multivariate probit model, whereby all six behaviours were jointly estimated, was explored. Conditional on pairwise combinations of the two productive activities, the probabilities of contemporaneously undertaking all four healthy behaviours were calculated. The results illustrate that the relationships among productive activities and health behaviours are inextricable and complex. Considering each of the four healthy behaviours separately, social participation and labour participation are associated with lower probabilities of abstaining from alcohol and tobacco but higher probabilities of exercising and keeping a nutritious diet among older adults. Considering all four healthy behaviours together, the productive activities are associated with a significant increase in the probability that a Thai older adult would simultaneously abstain from alcohol, abstain from tobacco, exercise, and eat healthily, compared to if the same individual undertakes neither social participation nor labour participation. This study calls for a consistent set of multiple-behaviour interventions to promote healthy and productive ageing.

1. Introduction

productive activities serve as a potential catalyst for health-related behavioural changes in old age. They refer to a process whereby older individuals are encouraged to remain active, serving as “a resource to their families, communities, and economies” (WHO, 2002). Productive activities can be market-based or non-market, for which labour participation and social participation represent an important example respectively. The literature has persistently linked labour participation and social participation with healthy behaviours and good health outcomes among older adults (Behncke, 2012; Godard, 2016; Robroek et al., 2013; Ronconi et al., 2012; Sirven and Debrand, 2008, 2012). The policy implication is clear: that the promotion of productive activities benefits health in old age. The relationships among health behaviours, social participation, and labour participation in old age can nevertheless be more deeply investigated. The term ‘health behaviours’ is a broad umbrella for different activities. Within the same individual, it is possible to observe several healthy behaviours, unhealthy behaviours, or a combination of both (Dias, 2010; Poortinga, 2007). It is also possible to observe

Defined as lifestyle choices that directly or indirectly impact health (Peel et al., 2005), health behaviours constitute a critical risk factor for non-communicable diseases (NCDs), which are now the leading cause of deaths globally (IHME, 2016). Unhealthy behaviours—notably, smoking, drinking, lack of exercise, and poor diet—have been associated with poor health and higher risks of mortality across settings (Balia and Jones, 2008; Dias, 2010; Holmes and Joseph, 2011; IHME, 2016). The NCD burden falls disproportionately on older adults, who are the most likely age group to be affected (IHME, 2016). The fact that health behaviours are modifiable and, consequently, that adverse impacts of NCDs could be partially avoided at a later stage of life provides a basis for the United Nation's Research Agenda on Ageing for the 21st Century which calls for more research on health behaviours in old age (Peel et al., 2005). Consistent with the World Health Organization (WHO)'s paradigms of Active and Productive Ageing (Walker and Aspalter, 2015),

E-mail address: [email protected]. https://doi.org/10.1016/j.socscimed.2018.07.031 Received 7 January 2018; Received in revised form 12 June 2018; Accepted 18 July 2018 Available online 20 July 2018 0277-9536/ © 2018 Elsevier Ltd. All rights reserved.

Social Science & Medicine 213 (2018) 12–19

N. Witvorapong

2008; 2012). Since respondents from the two surveys were independently sampled (NSO, 2008; 2012), observations between the years cannot be matched, and the combination of the two waves does not yield panel data. Instead, the two surveys offer repeated crosssectional data that include information on personal and household characteristics, lifestyle choices, as well as labour and social participation. The unit of analysis is older individuals. The 2007, 2011 surveys respectively comprise 20,841 and 31,501 individuals, who were older than 60. In addition to the age cut-off, sample selection is based on the fact that the use of a multivariate probit model requires complete data on all variables. The process excludes 5 observations in 2007 and 2199 observations in 2011 with missing data on labour participation and whether the older individual co-resided with adult children, which is an explanatory variable, respectively. The final, repeated cross-sectional sample consists of 50,138 observations, with 20,836 and 29,302 observations from 2007 to 2011 respectively. This is equivalent to 95.79% of the original sample prior to the sample-selection process. According to the World Development Indicators database, the size of the Thai older population was 5,426,606 in 2007 and 6,205,430 in 2011, therefore the final sample addresses 0.38% and 0.47% of the older population in 2007 and 2011 respectively. The high participation rate (95.79%) implies that sample selection bias, if any, is likely to be minimal. In light of the absence of panel data, the choice of a repeated crosssection over a single cross-section is deliberate. Using repeated crosssectional data increases the number of observations and degrees of freedom in the analysis (Witvorapong, 2015). This is especially important, since this study employs an empirical strategy that requires a large data set, jointly estimating several outcomes. Repeated crosssectional data can also be used to (partially) capture time-varying components of outcomes of interest, allowing for the inclusion of a time trend or time fixed effects in the regression (Øvrum, 2011; Witvorapong, 2015). Nevertheless, even though repeated cross-sectional data are preferred, this study takes full advantage of the availability of multiple years of comparably collected data and investigates each year of data separately. The comparison of repeated cross-sectional versus crosssectional samples can be used to assess consistency of the results in this study (Williams et al., 2014).

substitutability or complementarity among health behaviours (Kaestner et al., 2014). Social participation and labour participation are also interrelated. The two activities compete with each other, given the time constraint (Hank and Stuck, 2008). They may also complement each other, with social participation and labour participation in old age being considered complementary forms of social interactions (Saffer and Lamiraud, 2012). The complex interrelationships indicate that health behaviours, social participation, and labour participation are endogenously determined (Balia and Jones, 2008; Giordano et al., 2012; Sirven and Debrand, 2008). In addition to observable time and budget constraints, older adults differ in terms of life circumstances and preferences, which are unobserved. The presence of common unobserved factors implies that the decisions to adopt a certain lifestyle, participate in social events, and engage in paid employment are all jointly made, and that empirical investigation of these behaviours entails more than simple reduced-form regressions. This study investigates the complex interrelationships among health behaviours, social participation, and labour participation in old age. More specifically, it tests the following hypotheses on a sample of older adults:

• Social participation is associated with health behaviours; • Labour participation is associated with health behaviours; • Social participation and labour participation are associated with each other; • Health behaviours are associated with each other. The literature suggests that the first two hypotheses can be signed; social participation and labour participation have proved to motivate healthy behaviours. The evidence is mixed for the last two hypotheses. While it is expected that social participation and labour participation, on the one hand, and health behaviours, on the other, are correlated among themselves, directions of these relationships are not well established. This study is a repeated cross-sectional analysis, using the nationally representative 2007 and 2011 Surveys of Older Persons, Thailand. Distinctly sampled, the two surveys do not form panel data (NSO, 2008, 2012). The final sample consists of 50,138 Thai older adults in the postretirement ages (i.e. older than 60 years), and is analysed under a multivariate probit model, where all outcomes of interest are jointly estimated. The 2007, 2011 surveys are also separately investigated; the results are compared with those from the repeated cross-sectional sample to provide robustness checks. The study offers important contributions. While existing studies are largely based on developed countries (Holmes and Joseph, 2011), this study utilises data from a developing country in Asia: Thailand. The country represents a valid case study, as it is phasing into an aged society, with 13% of the population being considered old (Witvorapong, 2015), and is experiencing an epidemiological shift towards NCDs, similar to the global trend (IHME, 2016). This study represents one of the few studies to address four main behavioural risk factors, namely poor diet, smoking, alcohol consumption, and physical inactivity (Poortinga, 2007); existing studies often focus on one risk factor (Godard, 2016; Lindstrom et al., 2001). Finally, it methodologically improves upon the existing literature by jointly modelling the outcomes of interest and explicitly accounting for unobserved heterogeneity (Balia and Jones, 2008; Dias, 2010; Sirven and Debrand, 2008, 2012).

2.2. Empirical model There are six binary outcomes of interest: alcohol abstinence (Ai ) , tobacco abstinence (Ti ) , physical exercise (Ei ) , healthy diet consumption (Di ) , social participation (Si ) , and labour participation (Wi ) . A multivariate probit model is used, assuming the following latent variable equations:

2. Methods

Ai∗ = x i′ βa + εai, where Ai = 1(Ai∗ > 0)

(1)

Ti ∗ = x i′ βt + εti, where Ti = 1(Ti ∗ > 0)

(2)

Ei∗

(3)

= x i′ βe + εei, where Ei =

1(Ei∗

> 0)

Di∗ = x i′ βd + εdi, where Di = 1(Di∗ > 0)

(4)

Si∗ = x i′ βs + εsi, where Si = 1(Si∗ > 0)

(5)

Wi ∗

(6)

= x i′ βw + εwi, where Wi =

1(Wi ∗

> 0)

x i represents a vector of explanatory variables that are the same across all six equations and are composed of characteristics of individual i as well as regional fixed effects. Exploiting the nature of repeated crosssectional data, x i also includes, where applicable, time fixed effects. εji represents the error term for outcome j that pertains to individual i . βj is a vector of coefficients to be estimated. The multivariate probit model assumes that error terms are jointly

2.1. Study design and data This study employs a repeated cross-sectional study design. Data used for estimation are from the nationally representative 2007 and 2011 Surveys of Older Persons in Thailand, collected under the same sampling frame and method by the National Statistical Office (NSO, 13

Social Science & Medicine 213 (2018) 12–19

N. Witvorapong

Table 1 Descriptive Statistics of Repeated Cross-sectional Sample, 2007 Sample, and 2011 Sample of Thai Older Adults (60 + years). Variables

Panel A: Dependent variables Social participation, S (=1) Labour participation, W (=1) Alcohol abstinence, A (=1) Tobacco abstinence, T (=1) Physical exercise, E (=1) Healthy diet consumption, D (=1) Clustering of Healthy Behaviours No Healthy Behaviour A=0∩T=0∩E=0∩D=0 One Healthy Behaviour A=1∩T=0∩E=0∩D=0 A=0∩T=1∩E=0∩D=0 A=0∩T=0∩E=1∩D=0 A=0∩T=0∩E=0∩D=1 Two Healthy Behaviours A=1∩T=1∩E=0∩D=0 A=1∩T=0∩E=1∩D=0 A=1∩T=0∩E=0∩D=1 A=0∩T=1∩E=1∩D=0 A=0∩T=1∩E=0∩D=1 A=0∩T=0∩E=1∩D=1 Three Healthy Behaviours A=1∩T=1∩E=1∩D=0 A=1∩T=1∩E=0∩D=1 A=1∩T=0∩E=1∩D=1 A=0∩T=1∩E=1∩D=1 Four Healthy Behaviours A=1∩T=1∩E=1∩D=1

Mean (SD) Repeated Cross-Section (I)

2007 Sample (II)

2011 Sample (III)

Mean-Difference T-test Statistics (IV)

0.743 0.353 0.843 0.795 0.388 0.676

0.757 0.329 0.817 0.677 0.414 0.710

0.732 0.370 0.861 0.880 0.370 0.652

−6.311*** 9.503*** 13.435*** 57.285*** −10.019*** −13.427***

(0.437) (0.478) (0.364) (0.403) (0.487) (0.468)

(0.429) (0.470) (0.387) (0.468) (0.493) (0.454)

(0.443) (0.483) (0.346) (0.325) (0.482) (0.476)

0.024 (0.151)

0.023 (0.150)

0.024 (0.153)

0.701

0.032 0.014 0.006 0.021

(0.176) (0.119) (0.075) (0.143)

0.060 0.014 0.006 0.028

(0.238) (0.119) (0.078) (0.166)

0.012 0.014 0.005 0.016

(0.109) (0.119) (0.073) (0.124)

−30.572*** −0.104 −1.205 −9.920***

0.211 0.005 0.041 0.004 0.018 0.037

(0.408) (0.074) (0.199) (0.062) (0.132) (0.190)

0.154 0.009 0.079 0.004 0.019 0.049

(0.361) (0.094) (0.271) (0.061) (0.136) (0.216)

0.252 0.003 0.014 0.004 0.017 0.029

(0.434) (0.054) (0.119) (0.063) (0.130) (0168)

26.638*** −8.941*** −36.475*** 0.529 −1.295 −11.706***

0.028 0.251 0.038 0.034

(0.165) (0.433) (0.192) (0.181)

0.021 0.208 0.068 0.040

(0.142) (0.406) (0.252) (0.195)

0.034 0.281 0.017 0.030

(0.180) (0.449) (0.129) (0.170)

8.685*** 18.730*** −29.445*** −6.160***

0.236 (0.424)

0.218 (0.413)

Panel B: Explanatory variables Female (= 1) 0.527 (0.499) 0.477 (0.499) Age (years) 71.046 (7.649) 71.262 (7.507) Non-Buddhist (= 1) 0.052 (0.221) 0.049 (0.216) Marital status (Excluded group = single) Married (= 1) 0.578 (0.494) 0.527 (0.499) Widowed (=1) 0.382 (0.486) 0.461 (0.499) Divorced (= 1) 0.012 (0.108) 0.011 (0.106) Separated (= 1) 0.014 (0.119) 0.018 (0.132) Education (Excluded group = no education) Primary education (= 1) 0.720 (0.449) 0.672 (0.469) Lower secondary education (= 1) 0.038 (0.192) 0.036 (0.185) Upper secondary education (= 1) 0.033 (0.178) 0.039 (0.193) Higher than upper secondary education (= 1) 0.063 (0.243) 0.083 (0.276) Savings (Excluded group = no savings) 1 – < 50,000 THB (= 1) 0.090 (0.287) 0.144 (0.351) 50,000 - < 200,000 THB (= 1) 0.152 (0.359) 0.211 (0.408) 200,000 - < 700,000 THB (= 1) 0.162 (0.368) 0.238 (0.426) 700,000 THB or more (= 1) 0.100 (0.299) 0.114 (0.318) Home ownership (= 1) 0.792 (0.406) 0.775 (0.418) Functional health (Excluded group = no limitations) With ADL & IADL limitations (=1) 0.087 (0.282) 0.093 (0.291) With IADL limitations only (=1) 0.141 (0.348) 0.146 (0.353) Total number of children 3.841 (2.149) 3.995 (2.334) Living arrangements (Excluded group = living alone or with spouse, having no children or grandchildren) Co-residing with adult children (=1) 0.285 (0.452) 0.275 (0.447) Living in a skipped-generation household (=1) 0.090 (0.286) 0.088 (0.283) Living in a three-generation household (=1) 0.310 (0.463) 0.328 (0.469) Living alone, children in same province (=1) 0.267 (0.442) 0.245 (0.430) Living alone, children in different province (=1) 0.115 (0.319) 0.100 (0.300) Number of observations 50,138 20,836

0.248 (0.432)

7.858***

0.563 (0.496) 70.893 (7.744) 0.053 (0.025)

19.120*** −5.3218*** 2.061**

0.610 0.326 0.012 0.012

(0.477) (0.469) (0.110) (0.109)

38.910*** −30.984*** 0.692 - 5.211***

0.754 0.040 0.029 0.049

(0.431) (0.196) (0.167) (0.215)

20.238*** 2.517** −6.062*** −15.751***

0.052 0.110 0.108 0.089 0.804

(0.223) (0.313) (0.310) (0.285) (0.397)

−35.526*** - 31.327*** - 39.619*** - 9.354*** 7.921***

0.083 (0.275) 0.137 (0.344) 3.731 (1.999)

- 4.157*** - 2.658*** −13.565***

0.292 (0.455) 0.091 (0.288) 0.298 (0.457) 0.283 (0.450) 0.125 (0.331) 29,302

4.161*** 1.319 −7.032*** 9.494*** 8.598*** d.f. = 50,136

Notes: ***p < 0.01, **p < 0.05, *p < 0.1; d.f. = Degrees of freedom. Statistical significance in the last column based on two-group mean comparison t-tests.

14

Social Science & Medicine 213 (2018) 12–19

N. Witvorapong

accounts for temporal changes in the outcomes of interest.

and normally distributed. That is,

⎡ ⎢ εai ⎢⎡ 0 ⎤ ⎤ ⎡ ε ti ⎢⎢ 0 ⎥ ⎥ ⎢ ⎢⎢ 0 ⎥ ⎢ εei ⎥ ⎢ εdi ⎥ ∼ N ⎢ ⎢ 0 ⎥, ⎢⎢ ⎥ ⎢ εsi ⎥ ⎢⎢ 0 ⎥ ⎢ε ⎥ ⎢⎣ 0 ⎦ ⎣ wi ⎦ ⎢ ⎢ ⎣

⎡ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎣

1

ρat

ρta

1

ρea ρda ρsa ρwa

ρet ρdt ρst ρwt

ρae ρad ρas ρaw ⎤ ⎤ ⎥⎥ ρte ρtd ρts ρtw ⎥ ⎥ ⎥⎥ ρed ρes ρew ⎥ ⎥ 1 ⎥ ⎥ = N (0, Ω) 1 ρds ρdw ⎥ ⎥ ρde ⎥⎥ ρse ρsd 1 ρsw ⎥ ⎥ ρwe ρwd ρ 1 ⎥ ⎥ ws ⎦⎥ ⎦

3.1. Descriptive statistics Table 1 reports descriptive statistics. Columns (I)-(III) are based on the 2007–2011 repeated cross-section (n = 50,138), the 2007 crosssection (n = 20,836), and the 2011 cross-section (n = 29,302) respectively. Column (IV) displays two-group mean-comparison t-test statistics, evaluating whether the mean values of a given variable differ statistically between 2007 and 2011. Panel A of Table 1 shows descriptive statistics of the dependent variables. Their operational definitions are as follows:

N is the standard normal distribution, with 0 representing the expected values and Ω the (notably symmetric) variance-covariance matrix of the error terms. For individual i , Ω contains 15 unique values of ρ. Each ρ represents the conditional correlation for each pair of outcomes. It is assumed that each ρ is statistically different from zero, which indicates that the error terms for individual i are not independently distributed and that the outcomes are jointly determined. The multivariate probit model ensures that unobserved characteristics common across all equations for individual i are accounted for, thereby reducing estimation bias. Given the correlated error structure and the joint cumulative probability density function of the sixth order that is being required, the model is estimated using a simulated maximum likelihood approach (Roodman, 2011). All statistical analyses are performed using STATA Version 12. To investigate the relationships between health behaviours, on the one hand, and social participation and labour participation, on the other, alternative estimation techniques could be used. An obvious option is to specify the model such that social and labour participation become explanatory variables, reducing the number of dependent variables to the four healthy behaviours, i.e.

• Social participation is a binary variable, taking the value of 1 if the

• •

Hi∗ = x i′ βh + γSi + ηWi + εhi, where Hi = 1(Hi∗ > 0) and H ∈ {A , T , E , D } However, such specification is problematic in cross-sectional settings. The model is vulnerable to simultaneity bias, as health behaviours, social participation, and labour participation contemporaneously influence each other, compromising causality conclusions. To redress simultaneity bias requires valid instrumental variables (IVs) (Rocco et al., 2014; Ronconi et al., 2012), yet this strategy raises two further problems: (1) that valid IVs cannot always be found (which is the case here), and (2) that, even if statistically justified IVs can be found, they might not be sufficiently strong to provide explanatory power needed to eliminate bias. The choice of a multivariate probit model is consistent with the existing literature (Dias, 2010; Hank and Stuck, 2008; Sirven and Debrand, 2012). The joint estimation is useful. Instead of identifying the relationships of interest in a direct manner, this study assumes that the above six outcomes are jointly determined by the same attributes. This assumption rules out the need to identify valid and strong IVs and tackle endogeneity bias directly. Here, the linkages among health behaviours, social participation, and labour participation are stipulated through the covariances: ρ ’s. If the ρ ’s are statistically different from zero, then the assumption of jointly determined outcomes holds and the signs of the ρ ’s provide information as to how the outcomes are associated with each other.

older adult reported having participated in community-based activities in the past 12 months or holding an active membership in an association. In the surveys, examples given for community-based activities include a Thai equivalent of town hall meetings and other forms of social events. Examples of community-based associations are village-scout associations and village cooperatives, which are available uniformly throughout the country. Labour participation is binary, taking the value of 1 if the older adult reported having been employed/worked/self-employed in the past 12 months. The other dependent variables are self-reported healthy behaviours: alcohol abstinence, tobacco abstinence, physical exercise, and consumption of a nutritious diet. They represent the opposite of NCD risk factors, taking the value of 1 if the older adult reported having engaged in the respective behaviour in the past 6 months and 0 otherwise. Alcohol abstinence refers to current abstention from alcoholic beverages. Tobacco abstinence refers to current abstention from cigarettes, betel quid, and other forms of tobacco. Physical exercise refers to whether the individual reported having exercised on a “regular” basis. Nutritious diet refers to the daily consumption of vegetables, fruits, and at least 8 glasses of water on a “regular” basis. Since the term “regular” is not defined in the surveys, the cutoff for the exercise and diet variables, differentiating individuals who undertook the activities regularly vis-à-vis those who did not, is subjective.

The dichotomisation of the dependent variables is necessary. The surveys do not contain information on the extent to which individuals invested their time into the above activities, therefore it is not possible to form a continuous or a meaningful multi-category measure of any of the outcomes. The top half of Panel A suggests that data from the repeated crosssection, the 2007 sample, and the 2011 sample are largely consistent. Focusing on the repeated cross-section, the majority of the sample participated in social events (74.3%) and approximately one-third (35.3%) still worked even after the retirement age of 60. With the exception of physical exercise (38.8%), more than two-thirds of the sample adopted a healthy behaviour, where alcohol abstinence (84.3%) was the most common form. The 2007, 2011 samples can be similarly described, except that, in 2011, tobacco abstinence (88.0%), as opposed to alcohol abstinence, was the most frequently observed healthy behaviour. The bottom half of Panel A suggests that multiple healthy behaviours clustered within individuals. Based on the repeated cross-section, only 2.4% of the sample performed none of the healthy behaviours, living a completely unhealthy lifestyle in old age. The rest had at least one healthy behaviour. Undertaking all four healthy behaviours was the second most frequently observed combination (23.6%), after a combination of alcohol abstinence, tobacco abstinence, a healthy diet, and no/irregular physical exercise (25.1%). The repeated cross-section largely mirrors the 2007, 2011 surveys. The proportions of health-behaviour clustering are consistent across samples, with the exception being

3. Results This section is composed of descriptive and regression analyses. To assess consistency of the main results, the empirical analyses below are performed on three samples: the 2007–2011 repeated cross-sectional sample, as well as the 2007 and the 2011 cross-sectional samples. The discussion mainly revolves around the repeated cross-sectional sample, which is preferred, given its larger sample size and the fact that it 15

Social Science & Medicine 213 (2018) 12–19

N. Witvorapong

Table 2 Conditional Correlation Coefficients of Error Terms from Joint Estimation of Productive Activities and All Healthy Behaviours (Number of Equations = 6) Based on Repeated Cross-sectional Sample, 2007 Sample, and 2011 Sample of Thai Older Adults (60 + years). Pair of error terms

ρ _12: Alcohol abstinence and tobacco abstinence ρ _13: Alcohol abstinence and exercise ρ _14: Alcohol abstinence and healthy diet ρ _15: Alcohol abstinence and social participation ρ _16: Alcohol abstinence and labour participation ρ _23: Tobacco abstinence and exercise ρ _24: Tobacco abstinence and healthy diet ρ _25: Tobacco abstinence and social participation ρ_ 26: Tobacco abstinence and labour participation ρ _34: Exercise and healthy diet ρ _35: Exercise and social participation ρ _36: Exercise and labour participation ρ _45: Healthy diet and social participation ρ _46: Healthy diet and labour participation ρ _56: Social participation and labour participation Wald test statistics (Chi-squared) Number of observations

Conditional correlations Repeated cross-section (I)

2007 Sample (II)

2011 Sample (III)

0.700*** (0.012) −0.0885*** (0.010) 0.0151 (0.010) −0.0848*** (0.011) −0.165*** (0.010) −0.0357*** (0.009) 0.0226** (0.009) −0.0822*** (0.010) −0.112*** (0.010) 0.669*** (0.009) 0.135*** (0.009) 0.322*** (0.008) 0.0724*** (0.008) 0.0832*** (0.008) 0.0951*** (0.009) 5695.22*** 50,138

0.543*** (0.016) −0.092*** (0.014) 0.002 (0.015) −0.114*** (0.017) −0.166*** (0.015) −0.009 (0.012) 0.018 (0.012) −0.121*** (0.014) −0.086*** (0.013) 0.674*** (0.014) 0.105*** (0.013) 0.288*** (0.013) 0.089*** (0.013) 0.073*** (0.014) 0.082*** (0.015) 3915.58*** 20,836

0.910** (0.018) −0.085*** (0.019) 0.028** (0.013) −0.063*** (0.015) −0.169*** (0.016) −0.052*** (0.017) 0.040*** (0.013) −0.053*** (0.015) −0.127*** (0.015) 0.665*** (0.021) 0.154*** (0.012) 0.349** (0.172) 0.061*** (0.011) 0.093*** (0.010) 0.107*** (0.012) 7549.50*** 29,302

Notes: ***p < 0.01, **p < 0.05, *p < 0.1. Standard errors in parentheses.

item, (2) walk up the stairs, (3) take a means of public transportation, and (4) count money. Descriptive statistics of these characteristics are consistent among the three samples. Notable exceptions include the fact that there were more male than female (47.7%) respondents in 2007 and that respondents in 2007 had higher savings than those in 2011. The number of children and living arrangements represent family characteristics. Considering the repeated cross-section, the average number of living children of respondents, regardless of their marital status, was 3.841. Excluding single/never-married respondents, the figure would have become 3.893. The construction of living arrangements pertains also to all respondents. Approximately 28.5% of the repeated cross-sectional sample lived with adult children, 9.0% with grandchildren (known as a skipped-generation household) and 31.0% with both adult children and grandchildren, i.e. a three-generation household. The rest of the sample either lived alone with children or grandchildren nearby (26.7%) or alone with children or grandchildren further away (11.5%). The excluded group was older adults (regardless of their marital status), who lived alone or with their spouses and had no living children or grandchildren. Descriptive statistics of the family characteristics are similar across samples. Table 1 suggests that the repeated cross-section is appropriate. It reflects the 2007, 2011 samples well, with descriptive statistics being largely consistent among all three samples. Additionally, Column (IV) shows statistically significant differences between the 2007 and 2011 data. This indicates that, albeit sampled under the same sampling frame, the two years differ markedly, and that using either of the two cross-sections alone could lead to temporally inconsistent results. If the two surveys were statistically similar, then it would have been possible to use only one cross-section to represent both, and the repeated crosssection would have been less justifiable.

that, in 2007, performing all four healthy behaviours was the most frequently observed combination (21.8%). Explanatory variables in the model are composed of personal characteristics, socioeconomic status, functional limitations, number of living children, and living arrangements. The specification corresponds closely to the literature. Existing studies have demonstrated that, in addition to social and labour participation, statistically significant predictors of health behaviours include gender, age, education, marital status, socio-economic status, living arrangements, and health status (Dias, 2010; Godard, 2016; Kaestner et al., 2014; Sirven and Debrand, 2008). The same set of predictors also explains the two productive activities (Behncke, 2012; Rocco et al., 2014; Ronconi et al., 2012; Saffer and Lamiraud, 2012), and is extended in certain instances to include additional area characteristics (e.g. presence of public transportation and GDP per capita in the area of residence) (Rocco et al., 2014; Ronconi et al., 2012). As evidenced in Table 1, this study addresses most of the abovementioned predictors in the estimation. The exception is area characteristics. However, time and regional fixed effects (i.e., an indicator variable that separates the sample into 2007 versus 2011 observations, and a group of indicator variables that separate the sample into regions of residence respectively) are added to the specification. They are intended to capture—at least, partially—temporal and regional heterogeneity across individuals in the data set, and minimise risks of residual confounding/omitted variable bias that may arise from lack of explicit area characteristics. Descriptive statistics of explanatory variables are provided in Panel B of Table 1. Following the afore-described sample selection process (which addresses the estimation requirements to have complete data on all explanatory variables), there are no (internally) missing observations for any of the variables. The repeated cross-section contains more female than male respondents (52.7%). Given the range of 61–107 years, the average age of respondents was 71.046 years. Most respondents were married (57.8%), attained primary education (72.0%), had no savings (49.6%, the excluded category), and were functionally independent (77.2%, the excluded category). The variable representing Activities of Daily Living (ADL) limitations was constructed from five questions in the surveys, referring to the older individual's ability to independently (1) eat, (2) get dressed, (3) perform personal hygiene tasks, (4) squat, and (5) walk. The variable representing Instrumental Activities of Daily Living (IADL) limitations was similarly constructed, referring to the older individual's ability to independently (1) lift a heavy

3.2. Estimation results This section presents results from multivariate probit regressions, where the jointly estimated outcomes are social participation, labour participation, and the four healthy behaviours. Table 2 reports conditional correlations ( ρ ’s) of all pairs of outcomes. Column (I) represents the repeated cross-sectional sample. Columns (II) and (III) represent the 2007 and the 2011 samples respectively. Results of the three samples are derived from the same estimation process, using a simulated maximum likelihood method that draws each of the error terms 100 times. Coefficient estimates from the repeated cross-section are relegated to 16

Social Science & Medicine 213 (2018) 12–19

N. Witvorapong

Columns (I)-(III) are based on the repeated cross-sectional sample, the 2007 sample, and the 2011 sample respectively. Results from the three columns are similarly signed, but the magnitude varies. The repeated cross-sectional sample offers estimates that fall in between the 2007 and 2011 samples. Focusing on the repeated cross-section henceforth, Table 3 suggests that the probabilities of undertaking all four healthy behaviours increase, when older adults either participate in social events or work or both. The first row unveils that, conditional on participating in social events and working, the probability of engaging in all four healthy behaviours increases by 14.84% compared to when social and labour participation are not undertaken. The association between labour participation and the healthy behaviours is stronger than that between social participation and the healthy behaviours. Compared to when older adults do not work nor participate in social events, the probability of undertaking all four healthy behaviours rises by 11.55% when they work yet do not participate in social events. The probability increase is only 3.92% when older adults participate in social events but do not work. In summary, regardless of the samples, Tables 2 and 3 offer important conclusions. They include the fact that (1) the productive activities are positively associated with the consumption of a nutritious diet and exercise; (2) the productive activities are negatively associated with alcohol and tobacco abstinence; (3) the consumption of nutritious diets and physical exercise seem to be complementary; (4) alcohol abstinence and tobacco abstinence seem to be complementary; (5) the consumption of healthy diets and physical exercise, on the one hand, and alcohol and tobacco abstinence, on the other, appear to be substitutable, and (6) the productive activities, on the whole, promote a healthy lifestyle among older adults.

Table A.1 in Appendix 1 to save space. Table 2 shows that, individually, the conditional correlations (ρ ’s) for most error-term pairs are statistically significant at the 1–5% level. The Wald tests suggest they are also collectively significant. The results hold for all three samples, indicating that the decisions to undertake the six activities are jointly determined, correlated through attributes that are not captured in the data. These unobserved attributes may be rooted in the individual, e.g. preferences to be active, or may stem from ‘opportunity structures’ (Hank and Stuck, 2008), e.g. the functionality of an infrastructure that is amenable to active and healthy ageing in the community. The implication is that omission of any of the dependent variables would likely cause estimation bias, with unobserved heterogeneity not adequately accounted for. A closer look at the conditional correlations reveals that the interrelationships among productive activities and healthy behaviours are complicated and non-linear. The correlation coefficient for social participation and labour participation is positive, indicating the existence of unobserved factors that influence the two activities in the same direction. A more casual interpretation is that, if an older individual socialises, there is a good chance that he/she will also be in paid employment, and vice versa. Relative to the productive activities, a positive sign is also found with regard to physical exercise and healthy diet consumption. However, a negative sign is observed with regard to alcohol and tobacco abstinence. In other words, Table 2 suggests that older individuals who engage in either or both of the productive activities are likely to exercise and keep a healthy diet, but are also likely to smoke and drink, all else being equal. These results are again consistent across samples. Table 2 also highlights the complexities with which healthy behaviours cluster within individuals. The four healthy behaviours indirectly display some degree of complementarity and substitutability. The conditional correlations are positive for the following pairs: (1) alcohol and tobacco abstinence, (2) tobacco abstinence and healthy diet consumption, and (3) exercise and healthy diet consumption. They are negative for (1) alcohol abstinence and exercise, and (2) tobacco abstinence and exercise. The results indicate that the fact that an older individual exercises (thereby seemingly having a healthy lifestyle) does not necessarily mean that he/she will abstain from smoking and drinking. This is consistent with the idea of moral licensing, whereby performing a “moral” (healthy) behaviour can license and compensate for an “immoral” (unhealthy) behaviour (Cascio and Plant, 2015). To further contextualise the results, conditional probabilities of undertaking the four healthy behaviours, given different combinations of social participation and labour participation, are calculated. Each conditional probability is defined as:

4. Discussion and conclusions This study investigates the interrelationships among social participation, labour participation, and four healthy behaviours: alcohol abstinence, tobacco abstinence, physical exercise, and consumption of a nutritious diet. Based on a nationally representative repeated crosssectional sample of older individuals from the 2007 and 2011 Surveys of Older Persons, Thailand (n = 50,138), a multivariate probit model was used to jointly estimate the six outcomes of interest. Conditional on pairwise combinations of social participation and labour participation, the probabilities of undertaking all four healthy behaviours were calculated. The results illustrate that the relationships among social participation, labour participation, and healthy behaviours are inextricable and complex. Considering each healthy behaviour separately, social participation and labour participation are associated with lower probabilities of abstaining from alcohol and tobacco (which are not beneficial for health) but higher probabilities of exercising and keeping a nutritious diet for older adults (which are beneficial for health). Considering all healthy behaviours together, social participation and labour participation are associated with a significant increase in the probability that an older adult would simultaneously abstain from alcohol, abstain from tobacco, exercise, and eat healthily, compared to if the same individual undertakes neither social participation nor labour participation. The conclusions remain robust, regardless of whether the repeated cross-section or the cross-sectional samples are considered. This study has some limitations. First, despite the use of a repeated cross-section, which already represents an improvement over a single cross-section (Williams et al., 2014; Witvorapong, 2015), the paper is not based on panel data, as they are not available in the Thai context. The missing time dimension affects the causality conclusions (Giordano et al., 2012; Øvrum, 2011). Given the cross-sectional nature of the data, an alternative method to the non-recursive multivariate probit model employed here is to tackle endogeneity of the outcomes directly and rely on identification strategies to establish causality (Balia and Jones, 2008; Ronconi et al., 2012). However, that would involve having

Pr (Ai = 1 ∩ Ti = 1 ∩ Ei = 1 ∩ Di = 1 ∩ Si = j ∩ Wi = k ) Pr (Ai = 1 ∩ Ti = 1 ∩ Ei = 1 ∩ Di = 1 ∩ Si = j ∩ Wi = k ) = Pr (Si = j ∩ Wi = k ) The numerator is the predicted joint probability that individual i undertakes all four healthy behaviours, has a j outcome for social participation and a k outcome for labour participation. The denominator is the predicted joint probability that individual i has a j and a k outcome for social and labour participation respectively. The calculation of joint probabilities is based on linear predictions from coefficients of the multivariate probit regression for the numerator, and a bivariate probit regression for the denominator. Coefficient estimates used to calculate the conditional probabilities under the repeated cross-section are provided in Table A.1 in Appendix 1 and the last two columns of Table A.2 in Appendix 2 respectively. Under the Cholesky factorisation of the variance-covariance matrix of the error terms from the regressions, a simulation method with 100 pseudorandom draws from the standard uniform density is used (Cappellari and Jenkins, 2006). Table 3 shows within-individual differences of the conditional probabilities, tested with t-tests with the equal variance assumption. 17

Social Science & Medicine 213 (2018) 12–19

N. Witvorapong

Table 3 Differences in Conditional Probabilities from the Multivariate Probit Model Based on Repeated Cross-sectional Sample, 2007 Sample, and 2011 Sample of Thai Older Adults (60 + years). Pair of health behaviour clusters

Difference in conditional probabilities

Pr(A = 1 ∩ T = 1 ∩ E = 1 ∩ D = 1 (S = 1 ∩ W = 1)) - Pr(A = 1 ∩ T = 1 ∩ E = 1 ∩ D = 1 (S = 0 ∩ W = 0)) Pr(A = 1 ∩ T = 1 ∩ E = 1 ∩ D = 1 (S = 1 ∩ W = 1)) - Pr(A = 1 ∩ T = 1 ∩ E = 1 ∩ D = 1 (S = 1 ∩ W = 0)) Pr(A = 1 ∩ T = 1 ∩ E = 1 ∩ D = 1 (S = 1 ∩ W = 1)) - Pr(A = 1 ∩ T = 1 ∩ E = 1 ∩ D = 1 (S = 0 ∩ W = 1)) Pr(A = 1 ∩ T = 1 ∩ E = 1 ∩ D = 1 (S = 1 ∩ W = 0)) - Pr(A = 1 ∩ T = 1 ∩ E = 1 ∩ D = 1 (S = 0 ∩ W = 0)) Pr(A = 1 ∩ T = 1 ∩ E = 1 ∩ D = 1 (S = 1 ∩ W = 0)) - Pr(A = 1 ∩ T = 1 ∩ E = 1 ∩ D = 1 (S = 0 ∩ W = 1)) Pr(A = 1 ∩ T = 1 ∩ E = 1 ∩ D = 1 (S = 0 ∩ W = 1)) - Pr(A = 1 ∩ T = 1 ∩ E = 1 ∩ D = 1 (S = 0 ∩ W = 0)) Number of observations

Repeated cross-section (I)

2007 Sample (II)

2011 Sample (III)

0.1484*** (0.0003) 0.1091*** (0.0002) 0.0329*** (0.0001) 0.0392*** (0.0001) −0.0762*** (0.0002) 0.1155*** (0.0002) 50,138

0.083*** (0.0004) 0.070*** (0.0003) 0.001*** (0.0001) 0.013*** (0.0001) −0.071*** (0.0002) 0.085*** (0.0003) 20,836

0.203** (0.0004) 0.148*** (0.0004) 0.057*** (0.0001) 0.055*** (0.0001) −0.091*** (0.0003) 0.145*** (0.0003) 29,302

Notes: ***p < 0.01, **p < 0.05, *p < 0.1. Standard errors in parentheses. Statistical significance based on paired difference t-tests.

may have certain benefits. It can reduce measurement error in subjectively reported outcomes and allow for an easier interpretation of the results. Lindstrom et al. (2001), for example, converted a continuous measure of physical activity into a binary variable, separating their sample into the categories of “low” versus “high” physical activity for ease of interpretation. A similar approach is observed in Poortinga (2007), who transformed multi-category measures of several health behaviours into binary outcomes based on health recommendations in the UK, thereby cataloguing the sample into two groups: those who met the recommendations and those who did not. Second, to infer the extent to which the binary definition of outcome variables biases estimates, results from this study and existing studies with different definitions of health behaviours are discussed. The comparison reveals that they are consistent for the most part, illustrating that the main conclusions are robust to the operationalisation of the outcome variables. Øvrum (2011) estimated the demand for physical activity and for fruit and vegetables, defining the two outcome variables in frequency (count data) terms. Focusing on socioeconomic characteristics, the author found that better socioeconomic status was associated with a healthier lifestyle. The conclusion is consistent with estimation results of this study (Table A.1 in the appendix). As another example, Kaestner et al. (2014) estimated the effects of statin use on several health behaviours, gauging the degree of substitutability of different disease-prevention activities. The authors explored different operational definitions for each health behaviour. For example, alcohol consumption was defined in two ways, as a binary variable of whether the individual was currently drinking and that of whether the individual was drinking more than 3 ounces of alcohol per week. On the other hand, obesity, as a proxy for unhealthy dieting, was represented in three ways: by the body mass index (BMI), a binary variable of whether the individual was overweight (BMI > 25), and a binary variable of whether the individual was obese (BMI > 30). Kaestner et al. (2014) found that, notwithstanding differences in the operational definitions of health behaviours, the results were largely consistent. Despite the limitations, the study has several contributions. First, using a developing country in Asia—an area that is underrepresented in the literature—as a case study, this study improves upon existing studies by jointly estimating the outcomes of interest, explicitly teasing out their complex interrelationships. In a similar fashion as Sirven and Debrand (2008), this study takes a step further than other joint estimation papers by calculating joint and conditional probabilities of the jointly estimated outcomes to make the results less abstract. Second, this study confirms findings in the existing literature. It lends empirical credence to the fact that, despite the time constraint that each individual faces, social participation and labour participation are not necessarily substitutable (Saffer and Lamiraud, 2012). More

credible exclusion restrictions, which are not possible given limitations of the data-collection instrument. Nevertheless, it should be noted that this study does not claim causality but instead sits well with other correlational studies (Dias, 2010; Ellaway and Mcintyre, 2007), and that the conclusions here, where productive activities are positively linked with healthier behaviours, are in line with studies using panel data (Behncke, 2012; Rocco et al., 2014; Sirven and Debrand, 2012). Second, consistent with the earlier limitation, the fact that the sample was not followed over time means that sample attrition cannot be modelled. The study may suffer what is known as a ‘survival bias’ (Peel et al., 2005), whereby older adults who had died or left the survey because of their unhealthy behaviours were not included. In other words, the current sample is composed only of older adults who were alive and available at the time of the survey, and their behaviours could be systematically different from older adults who had died or dropped out, in which case the results would at best be internally valid. On a related note, past behaviours of the sample were also not documented. Although past behaviours are typically not controlled for in cross-sectional studies (Poortinga, 2007; Sirven and Debrand, 2008), the missing information could be important. For example, it is possible that an older adult is observed to abstain from alcohol today because she is a lifetime abstainer, or because she is a ‘sick quitter’, having had to stop drinking because of the negative health consequences (Dawson et al., 2012). The inability to distinguish lifetime abstainers and sick quitters from the rest of the older population could overestimate the associations between the productive activities vis-à-vis alcohol (and tobacco) abstinence. Finally, all outcomes of interest in the model are binary, therefore they lack important insights that could be key to better understanding the results. As the paper stands, the presence of social and labour participation is associated with a lower likelihood of abstaining from alcohol and tobacco and a higher likelihood of eating healthily and exercising. Having information on the intensity (in addition to the presence) of all six activities could further unlock the complexities of their associations. For instance, it is possible that social participation and labour participation are associated nonlinearly with drinking, increasing the likelihood of moderate or ‘social’ drinking, yet decreasing the likelihood of heavy drinking. MacCallum et al. (2002) demonstrate that the dichotomisation of outcome variables leads to loss of information on individual differences and allows potentially nonlinear relationships to be more easily overlooked, thereby yielding imprecise results. Nevertheless, while the binary definition of the outcome variables in this study may lead to an oversimplification of the results, the problem is unlikely to be so severe as to reverse the main conclusions. Two reasons are provided. First, the dichotomisation of health behaviours 18

Social Science & Medicine 213 (2018) 12–19

N. Witvorapong

Acknowledgement

specifically, the results suggest that, among Thai older adults, there exist common unobserved factors that influence an older person to jointly hold a club membership and participate in the labour market. This study also shows that, in old age, social participation and labour participation are generally positively associated with a healthy lifestyle (Behncke, 2012; Godard, 2016; Sirven and Debrand, 2008, 2012). The conclusion holds, even after controlling for human capital determinants of health such as education and wealth, which—as pointed out by Veenstra (2000)—should already explain much of health-related behaviours. It is argued here that social participation and labour participation represent forms of human interactions and social capital that influence health behaviours through, inter alia, the transmission of health information among peers, the promotion of norms on healthy behaviours, and the provision of psychosocial support which helps reduce social isolation that is beneficial to health of older adults (Sirven and Debrand, 2012; Veenstra, 2000). Finally, the most important contribution of this study lies perhaps with the fact that it addresses social participation, labour participation, and four—as opposed to one— healthy behaviours jointly. Breaking down an older adult's lifestyle into four distinct components is informative, because the associations between social participation and labour participation, on the one hand, and healthy behaviours, on the other, are heterogeneous and because within-individual health-related behaviours can be inconsistent—i.e., a person may undertake both healthy and unhealthy behaviours at the same time. Considering only a subset of healthy behaviours may bring one to conclude that social participation and labour participation are necessarily conducive to a completely healthy lifestyle, when in fact social participation and labour participation may also be associated with unhealthy behaviours. Tables A.2 and A.3 in Appendix 2 provide results of alternative multivariate probit models, whereby a given healthy behaviour (as opposed to all four), social participation and labour participation were jointly modelled. The tables demonstrate that, had the research question been to consider only one healthy behaviour, the results could have been misleading. The findings have an important policy implication: that social participation, labour participation, and all healthy behaviours should be jointly promoted in order to most effectively induce behavioural changes among older adults. Policies aimed at encouraging productive activities and healthy lifestyle choices should be aligned such that they provide a coherent package of multiple-behaviour interventions (Poortinga, 2007), whereby costs of socialising, working, and being healthy in old age are simultaneously lowered. Given the positive associations among social participation, labour participation, exercising and keeping a healthy diet and the fact that all of these activities are time-consuming but not necessarily money-consuming (Godard, 2016), this study suggests that the policy direction be shifted towards facilitating older adults to make a more productive use of their ample time, highlighting for them activities that are not straining on their budgets yet beneficial for their wellbeing. Given the negative correlations between the productive activities and alcohol and tobacco abstinence (which may offset the generally positive health effects of the former), this study suggests that the government more firmly enforces alcoholand tobacco-control policies and targets older adults who may overconsume the above goods as a means of enhancing social welfare. Future studies should continue to further explore the complexities of the interrelationships among healthy behaviours, social participation, and labour participation and to investigate particularly their potential nonlinearity, which would require a more comprehensive dataset that follows the sample over time.

This work was supported by Chulalongkorn Economics Research Center (CERC), Faculty of Economics, Chulalongkorn University. The author would like to thank the anonymous reviewers for their useful and constructive comments. Appendix A. Supplementary data Supplementary data related to this article can be found at https:// doi.org/10.1016/j.socscimed.2018.07.031. References Balia, S., Jones, A.M., 2008. Mortality, lifestyle and socio-economic status. J. Health Econ. 27 (1), 1–26. Behncke, S., 2012. Does retirement trigger ill health? Health Econ. 21, 282–300. Cappellari, L., Jenkins, S.P., 2006. Calculation of multivariate normal probabilities by simulation, with applications to maximum simulated likelihood estimation. STATA J. 6 (2), 156–189. Cascio, J., Plant, E.A., 2015. Prospective moral licensing: does anticipating doing good later allow you to be bad now? J. Exp. Soc. Psychol. 56, 110–116. Dawson, D.A., Goldstein, R.B., Grant, B.F., 2012. Prospective correlates of drinking cessation: variation across the life-course. Addiction 108, 712–722. Dias, P.R., 2010. Modelling opportunity in health under partial observability of circumstances. Health Econ. 19, 252–264. Ellaway, A., Macintyre, S., 2007. Is social participation associated with cardiovascular disease risk factors? Soc. Sci. Med. 64, 1384–1391. Giordano, G.N., Bjork, J., Lindstrom, M., 2012. Social capital and self-rated health—a study of temporal (causal) relationships. Soc. Sci. Med. 75, 340–348. Godard, M., 2016. Gaining weight through retirement? Results from the SHARE study. J. Health Econ. 45, 27–46. Hank, K., Stuck, S., 2008. Volunteer work, informal help and care among the 50+ in Europe: further evidence for ‘linked’ productive activities at older ages. Soc. Sci. Res. 37, 1280–1291. Holmes, W.R., Joseph, J., 2011. Social participation and healthy ageing: a neglected, significant protective factor for chronic non-communicable conditions. Glob. Health 7 (43), 1–8. Institute for Health Metrics and Evaluation (IHME), 2016. Rethinking Development and Health: Findings from the Global Burden of Disease Study. IHME, Seattle, WA. Kaestner, R., Darden, M., Lakdawalla, D., 2014. Are investments in disease prevention complements? The case of statins and health behaviors. J. Health Econ. 36, 151–163. Lindstrom, M., Hanson, B.S., Ostergren, P.O., 2001. Socioeconomic differences in leisuretime physical activity: the role of social participation and social capital in shaping health related behavior. Soc. Sci. Med. 52, 441–451. MacCallum, R.C., Zhang, S., Preacher, K.J., Rucker, D.D., 2002. On the practice of dichotomization of quantitative variables. Psychol. Meth. 7 (1), 19–40. National Statistical Office (NSO), Thailand, 2008. Final Report: Survey of Older Persons 2007 (in Thai). National Statistical Office, Bangkok. National Statistical Office (NSO), Thailand, 2012. Final Report: Survey of Older Persons 2011 (in Thai). National Statistical Office, Bangkok. Øvrum, A., 2011. Socioeconomic status and lifestyle choices: evidence from latent class analysis. Health Econ. 20, 971–984. Peel, N.M., McClure, R.J., Bartlett, H.P., 2005. Behavioral determinants of healthy aging. Am. J. Prev. Med. 28 (3), 298–304. Poortinga, W., 2007. The prevalence and clustering of four major lifestyle risk factors in an English adult population. Prev. Med. 44, 124–128. Robroek, S.J.W., Schuring, M., Croezen, S., Stattin, M., Burdorf, A., 2013. Poor health, unhealthy behaviors, and unfavorable work characteristics influence pathways of exit from paid employment among older workers in Europe: a four year follow-up study. Scand. J. Work. Environ. Health 39 (2), 125–133. Rocco, L., Fumagalli, E., Suhrcke, M., 2014. From social capital to health – and back. Health Econ. 23, 586–605. Roodman, D., 2011. Fitting fully observed recursive mixed-process models with cmp. STATA J. 11 (2), 159–206. Ronconi, L., Brown, T.T., Scheffler, R.M., 2012. Social capital and self-rated health in Argentina. Health Econ. 21, 201–208. Saffer, H., Lamiraud, K., 2012. The effect of hours of work on social interaction. Rev. Econ. Househ. 10, 237–258. Sirven, N., Debrand, T., 2008. Social participation and healthy ageing: an international comparison using SHARE data. Soc. Sci. Med. 67, 2017–2026. Sirven, N., Debrand, T., 2012. Social capital and health of older Europeans: causal pathways and health inequalities. Soc. Sci. Med. 75, 1288–1295. Veenstra, G., 2000. Social capital, SES and health: an individual-level analysis. Soc. Sci. Med. 50, 619–629. Walker, A., Aspalter, C., 2015. Active Ageing in Asia. Routledge, Oxon and New York. Williams, A.J., Wyatt, K.M., Williams, C.A., Logan, S., Henley, W.E., 2014. A repeated cross-sectional study examining the school impact on child weight status. Prev. Med. 64, 103–107. Witvorapong, N., 2015. The relationship between upstream intergenerational transfers and wealth of older adults: evidence from Thailand. J. Popul. Res. 32, 215–242. World Health Organization (WHO), 2002. Active Ageing: a Policy Framework. A Contribution of the World Health Organization to the Second United Nations World Assembly on Ageing, Madrid, Spain, April 2002. WHO, Geneva.

Conflicts of interest We have no conflict of interest to declare.

19