Social Capital, Economic Hardship, and Health: A Test of the Buffering Hypothesis in Transition and Nontransition Countries

ABSTRACT

Social capital is positively associated with a number of health outcomes, and it is theorized that social capital serves as a “buffer” during economic hard times, reducing the negative quality of life impacts of economic hardship. Using self-rated health as the outcome variable, we test whether social capital modifies the effect of economic hardship in a sample of 35 transition and non-transition countries and multilevel ordinal logistic regression models that interact social capital with economic hardship variables. Overall, we find consistent evidence that social capital improves health but our analyses do not suggest that social capital conditions the effect of economic hardship. Hence, social capital does not appear to act as “buffer” during trying economic times. We suggest that more research is needed to truly understand how social capital improves health.

Notes

¹To some degree this distinction parallels the “public” and “private” dimensions identified by Putnam (2001).

²Only 55 respondents provided “don’t know” answers to this question. We suspect that this small amount of missing data does not bias our models to any large degree.

³This measure of GNI per capita was adjusted using the Atlas method, which minimizes the effect of currency exchange rates.

⁴Of course, countries differ by far more than just their level of economic development or economic growth. We accessed data from the World Health Organization (2016) for country-level health care expenditures. We found the percentage of gross domestic product per capita spent on health care correlated very weakly with self-rated health (r = .01), while health care spending per capita correlated mildly with self-rated health (r = .12). These modest correlations suggest that the exclusion of these variables does not induce omitted variable bias in our models. As an additional complication, national health care spending data are not available for Kosovo. Further, GNI per capita correlated very strongly with health care spending per capita (r = .98) and moderately with health care spending as a percentage of GDP (r = .62), indicating that GNI per capita is a reasonable proxy for health care spending.

⁵We do not estimate random coefficient models primarily because the relatively small number of countries (35) can lead to highly biased coefficients (Mueleman and Billet 2009). Stegmueller (2013) showed that to estimate unbiased random coefficients a large number of Level-2 units is ideal. We also include few country—level predictors because of similar concerns about bias. Bryan and Jenkins (2016) performed a series of Monte Carlo simulations on country-level predictors and found that parameter estimates are highly unreliable when the number of Level-2 units is small; the authors suggest that researchers include relatively few contextual predictors.

⁶We also estimated additional models treating the categorical predictors as categorical (i.e., including a dummy variable for each category and omitting the lowest categorical). Substantively, these models produced nearly identical results as our simpler model specification.

⁷Multicollinearity diagnostics can be found in Appendix D. These indicate that multicollinearity is not a problem among our predictors. In the past the standard practice to mitigate multicollinearity among main effect and interaction terms was to mean-center variables. However, a new consensus has emerged across a number of disciplines that the multicollinearity generated by interaction terms is “nonessential” and that mean-centering, while not especially harmful, is not necessary in the light of multicollinearity generated in moderation models (Arceneaux and Huber 2007; Dalal and Zickar 2012; Echambadi and Hess 2007; Gatignon and Vosgerau 2005; Franzese and Kam 2009; Kromrey and Foster-Johnson 1998; Shieh 2011; Tate 1984). Briefly, these articles demonstrate that mean centering does not alter hypothesis tests, does not affect whether modification can be detected, and in general is not necessary. For instance, Franzese and Kam (2009) noted that centering “alters nothing important statistically and nothing at all substantively” (4). In line with this research, we estimated unreported models with mean-centered variables and found the results substantively similar to those reported here.

Additional information

Notes on contributors

Adam Mayer

Adam Mayer is a PhD student in sociology at Colorado State University in Fort Collins, CO where he is affiliated with the Center for Disaster and Risk Analysis and the Cooperative Institute for Research in the Atmosphere. Starting in the Fall of 2017, Adam will be an assistant professor in the department of Human Dimensions of Natural Resources at Colorado State University.