Does social isolation affect physical and mental health? A test of the social causation hypothesis using dynamic panel models with fixed effects

Abstract

Objectives

The widely acknowledged negative association between social isolation and physical and mental health is commonly interpreted in terms of social causation and has served as an important frame of reference for many interventions. However, evidence of social causation is likely biased because most studies cannot differentiate between social causation and health selection. The public attention given to this field of research indicates a need for analytical strategies that improve the understanding of the underlying link between social isolation and physical and mental health.

Methods

Using data from the German Socio-Economic Panel (GSOEP) study (2004 to 2012) of 6740 men and 7189 women aged 50 and above, we estimated dynamic panel models with fixed effects that allow us to probe the social causation hypothesis while accounting for direct selection (reverse causality) and indirect selection (unobserved heterogeneity). All analyses were conducted for women and men separately.

Results

We found that social isolation adversely affected mental health among older men and women to a degree that suggests practical relevance. However, we could not find a similar effect on physical health. A considerable part of the association between social isolation and both health outcomes was attributable to indirect selection, whereas direct selection led to underestimating the relevance of social isolation for mental health.

Conclusion

The results provide more convincing evidence that social isolation has adverse effects on mental health among older people. We conclude that effective interventions targeting social isolation might indeed be suitable for improving mental health among older people.

Keywords:

• Social isolation
• health
• structural equation modeling
• fixed-effects regression
• panel data

Acknowledgements

The authors would like to thank Henning Best, Volker Ludwig, Ingmar Rapp, Patrik Dahl, and Sonja Linder for helpful comments regarding earlier versions of the manuscript.

Disclosure statement

No potential conflict of interest was reported by the authors.

Ethical approval

The GSOEP data are collected in accordance with the data protection laws of the Federal Republic of Germany. All participants provided free and informed consent to participate in the survey. Because this study involved only a secondary analysis of anonymized GSOEP data, no further ethical approval was required.

Data availability statement

The data applied in this study are available from DIW Berlin (German Institute for Economic Research). Restrictions apply to the availability of these data, which were used under license for this study. The data are available at https://www.diw.de/en/diw_02.c.222829.en/access_and_ordering.html with the permission of DIW Berlin. Signing a contract for data distribution with DIW Berlin is a precondition for working with SOEP data. After this contract is signed, the data will be available upon request.

Notes

1 The two-year lags between previous and current health and the one-year lags between social isolation and health in our model are implemented to account for the lags in the GSOEP data by design.

2 There is a critical discussion in the literature about whether and when it is appropriate to include autoregressive effects. If misspecified, they can lead to biases in the estimates for all other variables in the model. For example, the random-effects model (M1) could conflate spurious and true state dependence by not considering unobserved heterogeneity, which could lead to an upwardly biased autoregressive effect and, thus, downwardly biased estimates for social isolation and other variables in the model. We decided to proceed cautiously and not include autoregressive effects unless theoretically justified. We adjusted the models step-by-step according to the three theoretical explanations to understand how the specific changes in the model affect the estimates.

3 Standard cross-lagged panel models are often used to estimate reciprocal associations between variables but are potentially biased by unobserved heterogeneity (indirect selection). In contrast, fixed effects models are well suited to control for unobserved heterogeneity but are potentially biased by reverse causality (direct selection). The ML-SEM approach combines the advantages of cross-lagged panel models and fixed-effects regressions and allows us to analyze the effect of isolation on health while considering both direct and indirect selection. However, using this approach, we were unable to directly estimate the reverse effect of health on the risk of social isolation because social isolation was a categorical variable in this study, and simulation studies have not yet shown that the ML-SEM approach to dynamic panel models with fixed effects can be generalized to categorical outcomes. However, the reverse effect of health on the risk of social isolation was still considered in M3 and M4 because these models allow for a correlation between the idiosyncratic error at each time point and social isolation in all subsequent waves. Thus, we considered reverse causality without obtaining a direct estimate of the reverse effect.

4 The means and regression coefficients reported in our study must be multiplied by 10 to obtain values on the original scale ranging from 0 to 100.

Additional information

Funding

This work was supported by the German Research Foundation (Deutsche Forschungsgemeinschaft) under Grant EC433/3-1.