Does Life Satisfaction Determine Subjective Health?

Abstract

The majority of previous studies on life satisfaction and health status were conducted in the advanced developed countries, while less attention has been focused on transitional countries, especially those in Central Asia, the Caucasus, and the Balkans. This study is among a very few studies that focused on the regions which faced on the prolonged economic and political crisis during the transition. Drawing on comparable data from 28 transitional countries in Eastern and Central Europe, the Caucasus, the Central Asia, and Turkey, we quantify the effect of self-reported life satisfaction on the self-reported health status of the population. To rule out reverse causality and to reduce estimation biases, we employed simultaneous equation models with instrumental variables. Two models used standard simultaneous equation regression (2SLS) and bivariate ordered probit regression (bioprobit) for categorical ordered variables. Our main finding is that, regardless of the model used, higher levels of life satisfaction determine higher health status. The mechanisms regarding the effects of life satisfaction on health are discussed. Future researchers are encouraged to include life satisfaction in their analyses of health status. From a methodological standpoint, we demonstrate that a strong endogeneity exists between life satisfaction and health status, regardless of the models used. Ignoring endogeneity and estimating a single stage regression model with life satisfaction and health status will likely lead to biased results.

Appendix 1

Estimation of bivariate ordered probit (bioprobit) regression model. We commence equations relating the latent health (H ^*) and life satisfaction (LS^*) status to individual characteristics of the respondents x (Sajaia 2006):

$$ {\mathrm{LS}}_i^{*}={x}_{1i}^{\hbox{'}}{\beta}_1+{\varepsilon}_{1i} $$

(1)

$$ {H}_i^{*}={\gamma}_i{\mathrm{LS}}_i^{*}+{x}_{2i}^{\hbox{'}}{\beta}_2+{\varepsilon}_{2i} $$

(2)

where x _1i and x _2i denote vector of observable characteristics, while β ₁ and β ₂ denote a vector of unknown parameters. Parameter gamma (γ _i ) is an unknown scalar that indicates the effect of LS ^* _i on H ^* _i for individual i. Two error terms, ε_1i and ε_2i , are normally distributed N (0, ∑) and the conditions of endogeneity such that E(x _1i ε_1i ) = 0 and E(x _2i ε _2i ) = 0. To observe two categorical variables LS and H such that

$$ {\mathrm{LS}}_i\left\{\begin{array}{l}1\kern0.48em \mathrm{if}\;{\mathrm{LS}}_i^{*}\le {g}_{11}\\ {}\dots \\ {}l\kern0.24em \mathrm{if}\;{g}_{1l-1}<{\mathrm{LS}}_i^{*}\le {g}_{1l}\\ {}\dots \\ {}L\kern0.24em \mathrm{if}\;{g}_{1L-1}<{\mathrm{LS}}_i^{*}\end{array}\right\} $$

(3)

and

$$ {H}_i\left\{\begin{array}{l}1\kern0.48em \mathrm{if}\;{H}_i^{*}\le {g}_{21}\\ {}\dots \\ {}m\kern0.24em \mathrm{if}\;{g}_{2m-1}<{H}_i^{*}\le {g}_{2m}\\ {}\dots \\ {}M\kern0.24em \mathrm{if}\;{g}_{2M-1}<{H}_i^{*}\end{array}\right\} $$

(4)

where the unknown cutoffs meet the following condition: g ₁₁ < g ₁₂ < · · · < g _1l … < g _1L − 1 and g ₂₁ < g ₂₂ < · · · < g _2m … < g _{2M − 1.} We define g ₁₀ = g ₂₀ = −∞ and g _1L  =  ₂  = ∞ for the sake of handling the boundary cases together.

The probability of observing LS _i  = l and H _i  = m is:

$$ \begin{array}{cc}\hfill \Pr \left({\mathrm{LS}}_i=l,{H}_i=m\right)\hfill & \hfill \begin{array}{l}={\varPhi}_2\left({g}_{1l}-{x}_{1i}^{\hbox{'}}{\beta}_1,\left({g}_{2m}-\gamma {x}_{1i}^{\hbox{'}}{\beta}_1-\gamma {x}_{2i}^{\hbox{'}}{\beta}_2\right)\psi, \overset{\sim }{\rho}\right)\\ {}-{\varPhi}_2\left({g}_{1l-1}-{x}_{1i}^{\hbox{'}}{\beta}_1,\left({g}_{2m}-\gamma {x}_{1i}^{\hbox{'}}{\beta}_1-{x}_{2i}^{\hbox{'}}{\beta}_2\right)\psi, \overset{\sim }{\rho}\right)\\ {}-{\varPhi}_2\left({g}_{1l}-{x}_{1i}^{\hbox{'}}{\beta}_1,\left({g}_{2m-1}-\gamma {x}_{1i}^{\hbox{'}}{\beta}_1-{x}_{2i}^{\hbox{'}}{\beta}_2\right)\psi, \overset{\sim }{\rho}\right)\\ {}+{\varPhi}_2\left({g}_{1l-1}-{x}_{1i}^{\hbox{'}}{\beta}_1,\left({g}_{2m-1}-\gamma {x}_{1i}^{\hbox{'}}{\beta}_1-{x}_{2i}^{\hbox{'}}{\beta}_2\right)\psi, \overset{\sim }{\rho}\right)\kern1em \end{array}\hfill \end{array} $$

(5)

where Φ ₂ is the bivariate standard normal cumulative distribution function and Ψ and $\overset{\sim }{\rho }$ are, respectively, defined as follows: $\psi =\frac{1}{\sqrt{1+2\gamma \rho +{\gamma}^2}}$ and $\overset{\sim }{\rho }=\psi \left(\gamma +\rho \right)$.

Provided that observations are independent, the logarithmic likelihood for the entire sample of size N is:

$$ \ln \Im ={\displaystyle \sum_{i=1}^N{\displaystyle \sum_{l=1}^L{\displaystyle \sum_{m=1}^MI\Big({\mathrm{LS}}_i}}}=l,{H}_i=m\Big) \ln \Pr \left({\mathrm{LS}}_i=l,{H}_i=m\right) $$

(6)

Similar to 2SLS, a set of instruments could be introduced in Eq. 3, and the system of Eqs. 1 and 2 could be computed based on a full-information maximum-likelihood estimation.