Social Capital and Health: A Meta-Analysis

Journal of Health Economics 72 (2020) 102317 Contents lists available at ScienceDirect Journal of Health Economics journal homepage: www.elsevier.com/locate/econbase Social capital and health: a meta-analysis Xindong Xue a,b , W. Robert Reed c,∗ , Andrea Menclova c a b c School of Public Administration, Zhongnan University of Economics and Law, China Collaborative Innovation Center for Social Governance of Hubei Province, China College of Business and Law, University of Canterbury, New Zealand a r t i c l e i n f o Article history: Received 4 January 2019 Received in revised form 2 December 2019 Accepted 12 March 2020 JEL classification: B49 C49 I10 I31 Keywords: Social capital Health Meta-analysis Mental health Physical health Self-reported health a b s t r a c t The relationship between social capital and health has received extensive attention in fields such as public health, medicine, epidemiology, gerontology and other health-related disciplines. In contrast, the economics literature on this subject is relatively small. To address this research gap, we investigate the cross-disciplinary empirical literature using meta-analysis. We analyze 12,778 estimates from 470 studies. Our analysis finds that social capital is significantly related to a variety of positive health outcomes. However, the effect sizes are consistently very small. This finding is robust across different types of social capital (e.g., cognitive, structural, bonding, bridging, linking), and for many different measures of health outcomes (e.g., mortality, disease/illnesses, depression). The small effects that we estimate cast doubt on recent initiatives to promote health through social capital such as those by the WHO, the OECD, and US Healthy People 2020. © 2020 Published by Elsevier B.V. 1. Introduction This study examines the literature on social capital and health. It has long been recognized that social capital plays an important role in economic affairs. A recent search on Web of Science identified over 1000 studies in economics journals in which social capital appears in the title. Research on social capital has been published in the American Economic Review (Guiso et al., 2004; Karlan, 2005), the Quarterly Journal of Economics (Knack and Keefer, 1997; Glaeser et al., 2000), the Economic Journal (Bowles and Gintis, 2002; Glaeser et al., 2002; Durlauf, 2002; Helliwell, 2006), and other top economics journals. The topic contin- ∗ Corresponding author. E-mail address: [email protected] (W.R. Reed). https://doi.org/10.1016/j.jhealeco.2020.102317 0167-6296/© 2020 Published by Elsevier B.V. ues to be of interest to economists (McCoy et al., 2019; Hoi et al., 2019; de Vaan et al., 2019d; Andini and Andini, 2018; and Wang, 2019). However, most of the economic research on social capital has focused on areas such as economic growth (Knack and Keefer, 1997; Algan and Cahuc, 2010), financial development (Guiso et al., 2004), political governance (Putnam, 1993; Nannicini et al., 2013) and formation of large firms and organizations (La Porta et al., 1997). While some economic research has focused on health, the literature is relatively small. This is surprising, not only because social capital has been found to be an important determinant of economic outcomes in other areas, but also because there is a voluminous literature on social capital and health outside of economics, particularly in public health, medicine, epidemiology, gerontology and other health-related disciplines. 2 X. Xue, W.R. Reed and A. Menclova / Journal of Health Economics 72 (2020) 102317 This study uses meta-analysis to aggregate empirical findings in the literature on social capital and health.1 Our goal is to assess the overall economic impact of social capital on health and determine whether it is “large” or “small”. Given the broad scope of both “social capital” and “health”, we also want to dig deeper into the overall relationship to explore whether certain types of social capital are particularly impactful for certain types of health outcomes. Our primary motivation is to determine whether this is a subject that should attract more attention from economists. If studies from other disciplines, and the extant economics literature on this topic, have found that social capital is an important determinant of health, then the dearth of economic research in this area represents a gap that should be addressed. On the other hand, if the conclusion from existing studies is that social capital is not an important determinant of health, then this justifies the relative lack of attention this subject has received in the economics discipline. Further, it would call into question recent initiatives to promote health through social capital by the WHO, the OECD, and US Healthy People 2020 (Rocco and Suhrcke, 2012; Centers for Disease Control and Prevention, 2010; Keeley, 2007). This study proceeds as follows. Section 2 reviews the economics literature on social capital. Section 3 explains our use of meta-analysis and our strategy for combining diverse estimates from different studies. Section 4 presents and discusses the data that we use in our analysis. Section 5 presents initial estimates of the “overall” effect of social capital on health, along with the results of a commonly used test for publication bias. Section 6 uses meta-regression to identify the factors that affect the size of the estimated effects of social capital on health. It also takes a closer look at finer categories of social capital and health to see if there are some kinds of social capital that may affect some kinds of health outcomes, even if the overall effect is small. Section 7 concludes by summarizing our results and drawing out implications for public policy and further study. monitoring costs. Further, social capital in the form of trust can reduce transaction costs, facilitating economic exchange and contributing to economic growth. The last mechanism has received the most empirical attention. LaPorta et al. (1997) show that trust improves the performance of large organizations. Zak and Knack (2001) present empirical evidence that trust is significantly related to economic growth. Their findings are corroborated by Beugelsdijk et al. (2004). Algan and Cahuc (2010) further substantiate the relationship between trust and economic growth by using “inherited trust” of descendants of US immigrants to establish a causal link between trust and growth. Another area that has received much attention is the relationship between trust and economic institutions. Berggren and Jordahl (2006) find that legal structure and security of property rights increase trust. Alesina, and Giuliano (2015) provide a recent summary of how culture, trust, and institutions interact to affect economic outcomes. Relatedly, Guiso, Sapienza, & Zingales (2004) and Tabellini (2010) find a strong link between culture, of which trust is a major component, and trade and economic development, respectively. Overall, there is broad consensus in the literature that the economic benefits of social capital in general – and trust in particular – are large. In contrast, the role of social capital in health is more opaque. Folland (2008) identifies a number of avenues by which social capital can affect health. Social capital may reduce stress through supportive relationships, trust, and the benefits of socializing. Stress reduction, in turn, positively affects health. Social capital in the form of socializing may make it easier to obtain (health-related) information. Further, having strong social ties may encourage individuals to make more “responsible” choices as their health and wellbeing are of importance to others whom they care about. In Folland’s (2008) formal model, social capital affects an individual’s utility U both directly, and indirectly via its effects on health: 2. Literature on the economic effects of Social Capital U = U(S, H, X) As noted above, there is a large literature that investigates the relationship between social capital and economic activity. A number of channels have been proposed for this relationship (Paldam and Svendsen, 2000). Social capital may enter the production function directly, alongside physical capital and human capital; e.g., in the form of consumer networks, reputation, and goodwill. Social capital may reduce free-riding and thus lessen the need for regulation in settings that require cooperation, reducing H = H(S, C(S)) 1 Previous meta-analyses of social capital and health have been published outside of economics (De Silva et al., 2005; Holt-Lunstad et al., 2010; Gilbert et al., 2013; Nyqvist et al., 2014). A working paper version of this manuscript describes the value-added of our analysis. Among other things, our study far exceeds these previous efforts in scope and detail. Further, previous studies have directly combined disparate estimates. In contrast, we translate estimates into partial correlation coefficients, which is a more appropriate method for comparing and combining diverse estimates. (1.a) where (1.b) and S is social capital, H is health, X are other consumption goods, and C are all other health inputs. Social capital in the form of socializing may improve health directly. For example, as noted above, supportive relationships can improve health by reducing stress, so that ∂H > 0. Social capital can also affect health indirectly. ∂S For example, strong family relationships may encourage an individual to obtain regular medical checkups, C, so that ∂H ∂C > 0. Alternatively, a desire to belong to a peer group ∂C ∂S which shares cigarettes may encourage an individual to smoke, C, ∂H ∂C < 0. As a result, the overall effect of social ∂C ∂S capital on health is theoretically ambiguous. Other formal models of the relationship between health and social capital are presented by Costa-Font and Mladovsky (2008) and Laporte (2014). Costa-Font and Mladovsky (2008) model the effects of social capital on health via the effects of peers on health-related X. Xue, W.R. Reed and A. Menclova / Journal of Health Economics 72 (2020) 102317 preferences. This corresponds to the effect of S on C, and subsequently H, in Folland’s model above. However, Costa-Font and Mladovsky (2008) extend this logic by modelling the propagation of preferences within a group through a social multiplier effect. Their approach allows for the dissemination of both health-improving (e.g., physical activity, medical checkups, vaccinations) and health-retarding (e.g., social smoking, drinking, and the obesity epidemic) behaviors within a social group. Laporte (2014) extends the classic Grossman (1972) model of investment in health to include social capital. She allows for two, complementary types of social capital, private social capital, in which the individual invests over time, and public social capital, which is exogenously determined. One consequence of the model is that increased geographical mobility on the part of health consumers can induce them to invest less in private social capital. Communities may respond by investing more in public social capital. This highlights the role of public policy in social capital. Folland at al. (2014) also explore the role of public policy in the provision of social capital. The above provides ample theoretical reasons for believing that social capital can be an important input to health. It also points out that social capital’s influence need not always be positive. Either way, the size of the economic impact of social capital on health is an empirical question. The next section explains the methodology we will use to answer this question. 3. Methodology We use meta-analysis to aggregate empirical findings in the cross-disciplinary literature on social capital and health. To be included in our meta-analysis, a study must estimate the “effect” that social capital has on health. As will be discussed below, there are a variety of estimation procedures and model specifications that studies have used to do this. Conceptually, we can represent these efforts with a linear model that regresses a measure of health (H) on a measure of social capital (S), along with a set of control variables (Zk ): H= ˇ0 + ˇ1 S + K  ˇk Zk + error (1) k=2 ˆ 1i be the effect estimated by study i, and let there Let ˇ be a total of M estimates produced by multiple studies.2 ˆ 1i , become the In meta-analysis, the estimated effects, ˇ dependent variable. OLS estimation of ˛0 in the equation 2 A complication arises because increases in the health variable can mean an improvement or a decline in health, depending on how it is measured. For example, a positive estimate of ˇ1 in Eq. (1) when health is measured by mortality means something different than when health is measured by a categorical variable increasing in good health. Likewise, sometimes a measure of social capital is defined so that a larger number means an increase in social capital, but sometimes it is measured so that a larger number means a decrease in social capital. In order to get the sign of the effect consistent across studies, we standardized the signs of the estimates so that a positive estimate implied that an increase in social capital was associated with an increase in good health. 3 below produces a value equivalent to the arithmetic average of the M estimates. ˆ 1i = ˛0 + εi , i = 1, 2, . . ., M ˇ (2) If the individual estimates constitute a representative sample from the population of estimated effects, then OLS will produce an unbiased estimate of the mean true effect of social capital on health.3 However, the OLS estimate will not be efficient. OLS gives equal weight to the individual estimated effects. ˆ  s are estimated more precisely But some of the ˇ 1i than others,  as  indicated by their different standard ˆ 1i ≡ SE . An efficient estimator would assign errors, s.e. ˇ i greater weight to more precise estimates. If all estimates come from a population with the same true effect, so that the variation in εi is proportionate to the sampling error – i.e., var (εi ) = (SE i )2  2 – then Weighted Least Squares (WLS) will be efficient, with the appropriate weight being the inverse of (SE i )2 . This model of effect size heterogeneity is known in the meta-analysis literature as “Fixed Effects”, and is not to be confused with the panel data estimator of the same name. “Fixed Effects” WLS estimation of Eq. (2) is equivalent to dividing each observation by SE i and then estimating with OLS:  1  ε ˆ 1i ˇ = ˛0 + i , i = 1, 2, . . ., M. SE i SE i SE i (3) Assuming representative sampling, “Fixed Effects” WLS estimation of ˛0 will produce an unbiased and efficient estimate of the mean true effect of social capital on health. Many researchers find the “Fixed Effects” model of effect heterogeneity too restrictive. More likely, there is not a single, true effect of social capital on health, but a distribution of true effects. This model of effect heterogeneity is known in the meta-analysis literature as “Random Effects”, which again should not be confused with the panel data estimator of the same name. Let  2 represent the variation in εi due to the fact that estimated effects are drawn from populations with differing true effects. Assuming the twosources of variation in εi are independent, then var (εi ) = (SE i )2 +  2  2 = ωi  2 . The corresponding “Random Effects” WLS estimator is equivalent to estimating the following equation by OLS: 1 ε ˆ 1i ˇ = ˛0 + i , i = 1, 2, . . ., M. ωi ωi ωi (4) Note that the “Random Effects” estimator produces a more uniform distribution of weights than “Fixed Effects”, since the weighting terms include a common constant,  2 . While researchers generally agree that the “Random Effects” model most closely matches reality, there is some debate about which works best in practice (Doucouliagos and Paldam, 2013; Reed, 2015). Accordingly, our analysis uses both. A related issue concerns the weighting of estimates versus studies. The number of estimates per study can vary 3 We address issues of publication bias and endogeneity below. 4 X. Xue, W.R. Reed and A. Menclova / Journal of Health Economics 72 (2020) 102317 widely. In our sample, partly because many specifications contain multiple measures of social capital, the number of estimates per study ranges from 1 to 240, with a mean of 27. The preceding WLS estimators implicitly give greater weight, sometimes dramatically greater weight, to studies with more estimates. Accordingly, we employ an alternative weighting system that, ceteris paribus, gives equal weight to studies rather than individual estimates. In the subsequent discussion, Weight1 and Weight2 denote the weighting systems that give equal weights to individual estimates and studies, respectively. Two more issues need to be addressed. The preceding assumes that it is meaningful to average estimated effects across studies. However, in the empirical literature on social capital and health, different measures are used for both health and social capital. Further, different estimation procedures are employed – linear regression, ordered probit models, hazard models, odds ratio models and more – so that the interpretation of coefficients varies greatly, despite the fact that the respective studies are all concerned with estimating the effect of social capital on health. This is a common situation in meta-analysis, and there is a common solution: to convert the respective estimates to partial correlation coefficients (PCCs): PCC i =  ti ti2 (5) + df i where ti and df i are the t-statistic and degrees of freedom associated with the respective estimated effect. The associated standard error is given by:  SE(PCC i ) = 1 − PCC 2i df i (6) ˆ 1i is The preceding analysis still holds, except that ˇ replaced by PCC i , and SE i now stands for the standard error of PCC i , so that Eqs (3) and (4) become PCC i = ˛0 · SE i and PCC i = ˛0 ωi  1  SE i 1 ωi + + εi , i = 1, 2, . . ., M SE i εi , i = 1, 2, . . ., M. ωi (3’) (4’) Accordingly, the parameter ˛0 represents the mean true effect of social capital on health measured as a correlation. How should one assess the estimates of ˛0 ? Like any correlation, PCC takes values between -1 and 1. Cohen (1988) suggested that correlation values of 0.10, 0.30, and 0.50 in absolute value should be interpreted as “small”, “medium” and “large” effects, and his interpretation is widely accepted. However, as Doucouliagos (2011) points out, Cohen’s taxonomy refers to simple, not partial, correlations. To investigate partial correlation sizes, Doucouliagos collected over 22,000 estimates in empirical economics and converted them to PCCs. He then ranked them from smallest to largest in absolute value. He defined the 25th, 50th, and 75th percentile values as “small”, “medium”, and “large”. While there was some difference across subfields of economics, overall PCC values of 0.07, 0.17, and 0.33 corresponded to “small”, “medium” and “large” effect sizes, respectively. This establishes a scale for comparing PCC values to other PCC values in the literature, and it is the standard we will employ in interpreting our empirical work.4 4. Description of data 4.1. Selection of studies We followed the MAER-Net protocols outlined by Stanley et al. (2013) in our search for studies. To account for the multi-dimensional nature of both social capital and health, the following combination of key words was used: “social capital”, “social trust”, “social networks”, “social participation”, “social support”, “social engagement”, “social integration”, “social relationships”, “social ties”, “reciprocity”, “social cohesion”, “social connections”, “social connectedness”, “volunteering”, “health”, “mortality”, “depression” and “disease”.5 We followed a stepwise procedure to finalize a set of studies from which to draw our estimated effects. First, we excluded theoretical studies that did not report estimated effects. Second, we excluded studies that did not report sufficient information to calculate PCC values and their standard errors. Specifically, a study had to report a numerical estimate for the effect of social capital on health, and a corresponding standard error, t-statistic, confidence interval, or p-value. Third, we excluded studies that estimated the effect of social capital on well-being, welfare, quality of life, and life satisfaction. While indirectly related to health, these outcomes are not comparable to direct health effects. Fourth, we eliminated studies using interaction terms and quadratic specifications of the social capital variable 4 A remaining issue concerns the determination of the t-statistic used in calculating PCC in Eq. (5). In many cases, studies either directly report the t-statistic corresponding to the estimated coefficient, or they report the standard error, from which the t-statistic is easily calculated. However, in the social capital and health literature, the most common estimation procedure is some variant of an odds ratio model, where the estimated coefficient is the odds ratio associated with a binary outcome. In addition to the estimate, studies either report the standard error of the estimated odds ratio, or the lower and upper bounds of a 95% confidence interval. In the former case, the t (or better, z) statistic is calculated by ˆ 1i ) · ˇ ˆ 1i /s.e.(ˇ ˆ 1i ). In the latter case, one first calculates SE ORi = ti = ln(ˇ (ln(UpperBoundi ) − ln(LowerBoundi ))/(2 · 1.96), and then calculates ti = ˆ 1i /SE ORi . Calculation of PCC proceeds accordingly. A related comln ˇ plication arises when studies only report the coefficient and a set of stars to indicate the level of statistical significance: e.g., *** = significant at the 1 percent level, ** = significant at the 5 percent level, * = significant at the 10 percent level, and no stars = insignificant at the 10 percent level. In these cases, we set the p-value = 0.005, 0.025, 0.075, and 0.50, respectively, and work backwards from the inverse of the t-distribution to calculate a tvalue. We record how we calculate our t-values to see if the respective methods affect our results. 5 The search was conducted using the search engines EconLit, JSTOR, EBSCO, Google Scholar, RePEc, SSRN, Social Science Citation Index (SSCI), Science Citation Index (SCI) and Scopus. Backward and forward citation searching was employed to leverage articles identified through the search engine process. We also manually searched academic journals that were found to have published studies on social capital and health. To be as comprehensive as possible, we also searched working papers, books, doctoral dissertations, master theses, and government reports. The search was ended in September 2017. X. Xue, W.R. Reed and A. Menclova / Journal of Health Economics 72 (2020) 102317 because of the difficulty of combining multiple coefficients to calculate a single estimated effect with standard error (cf. Gunby et al., 2017).6 For similar reasons, we dropped studies employing path analysis that use social capital as a mediator variable. Lastly, we excluded studies measuring the intergenerational effect of social capital on health; e.g., the effect of parent’s social capital on child’s health. 4.2. Study coding For each estimate in our sample, we coded data to enable construction of effect sizes and their standard errors, and to record corresponding study, data and estimation characteristics. Information included the study’s author(s), type of publication (e.g., journal), journal name, publication year, countries included in the study sample, number of observations, and data type (cross-sectional or panel). We recorded the estimated coefficient and associated statistics (such as standard errors, confidence intervals, etc.).7 We categorized the estimation procedures used to produce each estimate. Binary variables were used to indicate the following methods: OLS, FGLS, probit/logit, ordered probit/logit, Hierarchical Linear Model (HLM), and instrumental variables (IV). In addition, a set of dummy variables were used to indicate whether common control variables were included in the regression specifications (e.g., age, gender, income, etc.). Endogeneity is certainly a concern in the social capital and health literature. While most studies assume that the effect runs from social capital to health, the literature does recognize that causation can go in the opposite direction. For this reason, our analysis will pay particular attention to estimates derived from panel data or IV estimation procedures. Holding other factors constant, we will want to investigate whether these estimates differ systematically from estimates that do not address endogeneity. For each estimate, we coded the associated health outcome, type of social capital, and whether the social capital variable was individual-based or was aggregated at the community level. Table 1 provides some detail about the different measures of health employed by the studies in our sample. Measures of health consisted of measures of physical health, mental health, and measures of overall/general health. Among measures of physical health, the most common measure was mortality, usually “all-cause” mortality, but sometimes mortality due to a particular illness or disease, such as cancer or cardio-vascular disease. Studies measured mortality over given sample periods, and as the sample periods differed in length, the interpretation of mortality rates differed accordingly. 6 Note that we do not encounter problems with calculating marginal effects for nonlinear models because all we require for PCC is to be able to calculate a t-statistic. The problem arises when more than one coefficient is involved, such as when there is an interaction or quadratic term. 7 We note that due to poor reporting practices and the generally large number of observations used in the respective studies – the median number of observations for a given estimate in our study was approximately 3,500 – we substituted the number of observations for degrees of freedom in calculating PCC and SE(PCC) in Eqs. (5) and (6). 5 Table 1 Common Measures of Health. Measure Percent PHYSICAL HEALTH (4770 observations) Mortality Disease/Illness Self-Reported 51.2 17.4 20.9 MENTAL HEALTH (2934 observations) Depression Self-Reported 28.2 90.9 NOTE: Authors’ calculations. Percent values are either percent of total Physical Health observations or percent of total Mental Health observations. Table 2 Common Types of Social Capital. Measure Percent COGNITIVE / STRUCTURAL (11,776 observations) Cognitive (What people feel) Social trust Perceived social support Perceived social cohension Perceived reciprocity Sense of belonging Loneliness 13.2 7.3 2.0 1.5 1.5 1.0 Structural (What people do) Social participation Social networks Social support Social engagement Volunteering Group membership Social integration Social relationship 16.7 16.4 6.7 6.5 4.8 4.6 3.9 3.5 BONDING / BRIDGING / LINKING (3572 observations) Bonding (Horizontal ties between similar people) Bridging (Horizontal ties between dissimilar people) Linking (Vertical ties between different people) 62.0 34.2 3.9 NOTE: Authors’ calculations. Percent values are either percent of total Cognitive/Structural observations or percent of total Bonding/Bridging/Linking observations. The second most common type of physical health measure was the presence or onset of a particular illness or disease (e.g., hypertension, heart disease, diabetes). Various measures of overall health were also common. Sometimes these were indices constructed from multiple questions about a person’s health, and sometimes they were categorical measures in which a respondent’s health was characterized as good, fair, poor, etc. A substantial number of the physical health measures relied on respondents’ own assessments. Among mental health measures, depression was the most common category of mental health. Other categories included dementia, mental distress or anxiety, and measures of cognitive ability. As with physical health, many of the studies used an overall measure of mental well-being. Perhaps not surprisingly, a very large number of mental health measures relied on self-reported assessments. Table 2 gives a sense of the wide variety of social capital variables used by the studies in our sample. The most common framework employed by studies was cognitive/structural, where cognitive refers to what people feel, and structural to what people do. Social trust was the most 6 X. Xue, W.R. Reed and A. Menclova / Journal of Health Economics 72 (2020) 102317 common type of cognitive social capital, followed by the perception of social support, and then the perception of social cohesion. The key element here is the respondent’s perception of these constructs. The most common measures of structural social capital was participation in some form of social activity, followed by measures of one’s network of personal relationships. It is noteworthy that some measures of social capital mixed the two types of social capital, often by composing an index of social capital that relies on both. An alternative framework for categorizing social capital is bonding/bridging/linking. While less common than cognitive/structural, it is still widely used. Bonding refers to horizontal ties between similar people, while bridging refers to horizontal ties between dissimilar people. Linking refers to hierarchical relationships. Of these, bonding was the most commonly used social capital variable in the studies included in our sample. 5. Data Analysis: part 1 5.1. Descriptive Analysis Our initial dataset consisted of 13,040 observations gleaned from 471 studies. Calculation of PCC uses t-values and df, so the first two columns of Table 3 focus on these variables. The full sample of t-values has mean and median values of 1.67 and 1.20, respectively. We will comment on these relatively low values later. For now, we wish to note the minimum and maximum values of -59.67 and 850. This raises concern with outliers. A similar concern applies to the df variable. It has mean and median values of 29,023 and 3300, with minimum and maximum values of 5 and 2,442,948. The corresponding distribution of PCC values ranges from -0.747 to 0.998, with mean and median values of 0.030 and 0.019. Large (absolute) values of PCC are potentially a problem because of the key role that PCC plays in determining the standard error, and hence, the weights used in the empirical analysis: s.e.(PCC i ) = 1−PCC 2 i , with weights df i increasing in the absolute value of PCC. As a result, we proceed by truncating the top and bottom 1% of PCC values, leaving 12,778 observations from 470 studies. The truncated distributions of t-statistic, df, and PCC values are also reported in Table 3, immediately to the right of the full sample statistics.8 Corresponding histograms for the t-statistics and PCC values are reported in Fig. 1. The two histograms in Fig. 1 and corresponding columns in Table 3 go far in answering our first question about the size of the effect of social capital on health. The mean and median PCC values for the truncated sample are 0.028 and 0.019. Based on Doucouliagos (2011), these do not even get close to the threshold value of 0.07 that Doucouliagos sets for “small”. If social capital has an 8 As a robustness check, we also truncated the observations based on the lowest and highest one-percent of t-values and df’s. We redid all the analyses with this alternative truncation strategy. The results were very similar to those reported in this paper, and the qualitative conclusions were identical. Fig. 1. Distribution of t and PCC Values. effect on health, these values suggest that the effect is very small. The reasons for the small PCC values are not hard to identify. First, a large number of estimates in the literature are statistically insignificant. The table immediately below the histogram in the top panel of Fig. 1 reports that 56.4% of all t-values lie between -2 and 2. Compounding these relatively low t-values are relatively large sample sizes. The distribution of df values for the truncated sample ranges from 5 to 2,442,948, with a median value of 3,451. If we calculate the PCC value that corresponds to the median t and df values using Eq. (5), we obtain a value for PCC equal to 0.020. However, there are two caveats. First, the numbers in Table 3, and the values represented in Fig. 1, are unweighted. So we need to re-compute our estimate of the mean true effect, ˛0 , using the different weighting schemes described above. Second, the analysis ignores publication bias. Publication bias arises when the results reported by researchers, and/or the studies accepted for publication by journals, comprise a biased sample of the population of all estimates. Note that “publication bias” can occur X. Xue, W.R. Reed and A. Menclova / Journal of Health Economics 72 (2020) 102317 7 Table 3 Descriptive Statistics for Estimated Effects and t-statistics. t-Statistics Mean Median Minimum Maximum Std. Dev. 1% 5% 10% 90% 95% 99% Obs df PCC Values Full Truncated Full Truncated Full Truncated 1.67 1.20 −59.67 850.00 8.40 −4.91 −2.66 −1.75 4.70 7.16 15.34 13,040 1.60 1.20 −17.44 48.86 3.53 −4.37 −2.58 −1.63 4.67 7.02 14.60 12,778 29,023 3300 5 2,442,948 201,648 39 190 412 23,153 44,986 271,642 13,040 29,600 3451 5 2,442,948 203,663 49 244 462 23,153 44,986 1,358,932 12,778 0.030 0.019 −0.747 0.998 0.084 −0.157 −0.061 −0.033 0.102 0.143 0.361 13,040 0.028 0.019 −0.156 0.360 0.059 −0.113 −0.055 −0.031 0.097 0.132 0.239 12,778 NOTE: The truncated sample is obtained from the Full Sample by deleting observations having the top and bottom 1% of PCC values. even in working papers that are not published in journals. This can happen if researchers choose not to write up results because the initial analyses did not produce interesting/promising findings.9 In that case, even unpublished working papers can be characterized by publication bias. Publication bias is widely recognized as a problem, with selection typically favoring estimates that are statistically significant, and/or consistent with researchers’ and journals’ preconceived beliefs (Christensen and Miguel, 2018). That being said, we note that in order for publication bias to explain the low PCC values we see in our sample, it would have to discriminate against statistically significant estimates. Most researchers would view this as unlikely. Fig. 1 and Table 3 present “overall” estimates, aggregating estimates of the effects of social capital across different kinds of health outcomes. To address concerns about the validity of combining these estimates, Fig. 2 breaks the full sample of PCC values into three subsamples, depending on whether the health outcome is physical health, mental health, or general health. A table at the bottom of the figure reports means and standard deviations for each of the three subsamples. Mean PCC values are similar, with most values lying within the range characterized by Doucouliagos (2011) as small. The similar distributions provide casual support for the legitimacy of combining estimates for the different health outcomes. We will explore this subject in greater detail below. Table 4 describes some of the other variables in our data. Approximately 98% of the estimates in our sample are drawn from peer-reviewed journals (Journal). It is common in meta-analyses to include a mix of published and unpublished studies, mostly to address publication bias. While our initial search produced a larger set of working papers, most of these were subsequently published, and were eliminated as duplicates. As it turns out, publication bias ends up not being a major concern in our study. While we find evidence of publication bias in our subsequent analysis, its estimated influence is very small, a result foreshadowed by the small PCC values reported above. The variable PubYear 9 Franco et al. (2014) report that the main source of publication bias is failure of researchers to write up results that are not significant or interesting. reports that the studies in our sample were published in the window from 1985 to 2017. Social capital can be constructed on an individuallevel (Glaeser et al., 2002) or group-level (Coleman, 1990; Putnam, 1995) basis. 86.5% of the estimates in our sample are associated with individual-level measures (IndividualSC). Our analysis will address whether individual- and group-level measures of social capital have different effects on health. Our sample covers a diverse set of countries, with the largest number of estimates using data from Western or Northern Europe (33.9%). East Asia (20.4%) and the USA (20.9%) were also common areas to study, but a substantial number of estimates came from countries other than those above. Of the three categories of health outcomes, most studies examined the effect of social capital on a general measure of health (40.4% of estimates), closely followed by physical health (37.3%) and then mental health (23.0%). Two thirds of the estimated effects were based on self-reported assessments of health. Almost all of the estimates were derived using social capital variables categorized as cognitive and structural social capital (92.2%). Another perspective, which includes bonding, bridging, and linking social capital, was less commonly employed (28.0%). It is quite common for studies to include an array of social capital variables in a single regression equation (NumberSCVariables). In our sample, these ranged from 1 to 28 separate social capital variables in a single specification, with a mean of 6.5 social capital variables per equation. We also tracked the control variables that were commonly included in studies of social capital and health. 85.2% of estimates came from a regression specification including an age variable; 84.5% included a gender variable; 59.9% included an education variable; 38.9% included a marital status variable; and 40.3% included an income variable. A large number of different estimation procedures were used to produce the estimated effects in our sample. The most common procedure involved the odds ratio model, such as logistic regression or hazard model estimation (55.2%). The next two most common procedures were hierarchical linear modelling (17.9%) and ordinary least squares 8 X. Xue, W.R. Reed and A. Menclova / Journal of Health Economics 72 (2020) 102317 Table 4 Description of Variables. Variable Description Mean Min Max Study type Journal PubYear = 1, if study is a journal article Year study was published/appeared 0.977 2009.3 0 1985 1 2017 Data characteristics IndividualSC Panel Cross-sectional* = 1, if estimate based on individual-level social capital variable = 1, if estimate based on panel data = 1, if estimate based on cross-sectional data 0.865 0.444 0.556 0 0 0 1 1 1 Countries EastAsia USA* WestNorthEurope HighIncome OtherCountry = 1, if country studied is located in East Asia = 1, if country studied is USA = 1, if country studied is located in Western or Northern Europe = 1, if country studied is a high income country not included above = 1, if country studied is none of the above 0.204 0.209 0.339 0.094 0.155 0 0 0 0 0 1 1 1 1 1 Health measure PhysicalHealth MentalHealth GeneralHealth* SelfReported = 1, if health variable measures physical health = 1, if health variable measures mental health = 1, if health variable measures overall health = 1, if health variable is self-reported 0.373 0.230 0.404 0.679 0 0 0 0 1 1 1 1 Social capital measure CognitiveStructural BondBridgeLink NumberSCVariables = 1, if social capital is cognitive or structural = 1, if social capital is bonding, bridging, or linking Number of social capital variables included in the regression 0.922 0.280 6.49 0 0 1 1 1 28 Control variables Age Gender Education MaritalStatus Income = 1, if an age variable is included in the regression = 1, if a gender variable is included in the regression = 1, if an education variable is included in the regression = 1, if a marital status variable is included in the regression = 1, if an income variable is included in the regression 0.852 0.845 0.599 0.389 0.403 0 0 0 0 0 1 1 1 1 1 Estimation method OLS ORHazard HLM FGLS* ProbitLogit* OrderedProbitLogit* IV* OtherEstimation* SENonspherical = 1, if estimation method is OLS = 1, if estimation method is Odds Ratio or Hazards Ratio = 1, if estimation method is Hierarchical Linear Modelling = 1, if estimation method is FGLS = 1, if estimation method is Probit or Logit = 1, if estimation method is Ordered Probit or Logit = 1, if estimation method is Instrumental Variables = 1, if estimation method is none of the above = 1, if standard error estimation assumes nonspherical errors 0.139 0.552 0.179 0.017 0.053 0.031 0.019 0.010 0.255 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 Calculation of t-statistic tNormal tCalculatedBypValue tCalculatedByCI* = 1, if t-statistic is calculated as ratio of coefficient to standard error = 1, if t-statistic is calculated from p-value = 1, if t-statistic is calculated from confidence interval 0.215 0.178 0.606 0 0 0 1 1 1 NOTE: When the grouped variables include all possible categories, the categories omitted in the subsequent analysis (the benchmark categories) are indicated by an asterisk. (13.9%). Other procedures included probit and logit models, ordered probit and logit models, feasible generalized least squares, and instrumental variable estimation. We note that only a very low percentage of estimates employed instrumental variable procedures (1.9%). The standard error of the estimated effect plays an important role in weighting observations. As a result, we recorded the procedure used to estimate the standard error, categorizing these as either assuming spherical errors (homoskedasticity, error independence) or not. 25.5% of the estimates calculated standard errors assuming some form of nonspherical error behavior. Finally, three different methods were used to derive t-values. In 21.5% of the cases, t-statistics were either directly reported, or could be calculated by dividing the estimated coefficient by its reported standard error (tNor- mal). In 17.8% of the cases, all we had was a p-value, and we worked backwards from the inverse t-distribution function to obtain a t-value (tCalculatedBypValue). However, in most cases (60.6%), the t-statistic was calculated using the reported confidence interval or from the log of the estimated odds ratio (tCalculatedByCI).10 While we are aware of no study that compares the frequency of “Fixed Effects” and “Random Effects” estimators in the meta-analysis literature, our sense is that “Fixed Effects” is generally preferred by researchers. Table 5 identifies a concern with “Fixed Effects”. It calculates a “study weight” for each study in our sample, weighting the individual estimates of that study by the respective weighting 10 See footnote #4 for more details. X. Xue, W.R. Reed and A. Menclova / Journal of Health Economics 72 (2020) 102317 9 Table 5 Study Weights. Mean Median 1% 5% 10% 90% 95% 99% Maximum Top 3 Top 10 Studies Fixed Effects Random Effects 0.21% 0.01% 0.0001% 0.0003% 0.0007% 0.1622% 0.3251% 1.9023% 45.2% 67.4% 77.3% 470 0.21% 0.23% 0.0467% 0.0959% 0.1399% 0.2576% 0.2607% 0.2636% 0.26% 0.79% 2.63% 470 NOTE: The methodology for calculating “study weights” is described in Footnote #11 in the text. 45.2%.12 The top 3 studies (out of 470!) account for 67.4% of the total weight, and the top 10 studies comprise almost 77.3%. Thus the “Fixed Effects” estimate will disproportionately weight a small number of select studies that have large PCC values and/or use a large number of observations (df). In contrast, the large size of  2 in the “Random Effects” estimator swamps the size of the individual (SE i )2 terms, so that the “Random Effects” estimator weights the individual estimates much more uniformly.13 The median value is 0.23%, compared to a mean value of 0.21%. The maximum weight any single study receives is 0.26%, and the top 10 studies account for 2.63%. This arguably overcompensates the extreme skewness of the “Fixed Effects” estimator. In what follows, we report both “Fixed Effects” and “Random Effects”, with a mild preference for the latter because it is not so heavily dependent on a small number of studies. As a practical matter, all of our key findings are robust across estimators. 5.2. Publication bias Fig. 2. Distribution of PCC Values by Measures of Health. scheme (“Fixed Effects”/“Random Effects”) and then aggregating the weights at the study level. In this way, each study receives a weight, the sum of which equals 100%.11 If the 100% weight was divided equally across studies, given 470 studies, each study would receive a weight of 0.21%. Against this benchmark, “Fixed Effects” produces a highly skewed weighting distribution. The median weight is 0.01%, and the maximum weight for a single study is 11 Study weights were calculated by wi /  or wi = 1/ (SE i )2 +  2  2 wi , where wi = 1/(SE i ) depending on whether Fixed Effects or Random Effects were being used (cf. Rinquist, 2013, page 128). Publication bias represents a serious challenge to the validity of meta-analysis. If the estimates in the literature are disproportionately large and significant, then averaging them will preserve this bias, producing a distorted estimate of the mean true effect. Methods to identify and correct publication bias remains an active research area in the meta-analysis literature (Andrews and Kasy, 2017; Stanley et al., 2017; Alinaghi and Reed, 2018). A common test for publication bias is given by the Funnel Asymmetry Test (FAT). The FAT is carried out by adding the standard error variable, SE, to the constantonly regressions above. It is designed to capture the idea that publication bias introduces a systematic relationship between the effect size (PCC) and its standard error (SE): PCC i = ˛0 + ˛1 SE i + εi (7) 12 Our ID for this study is 177 (Blakely et al., 2006). It has 72 estimated effects of social capital on health, and the number of observations in the respective samples ranges from 2,306,760 to 2,442,948. 13 The corresponding I2 value is 89.9%, which indicates that  2 is approximately nine times as large as the average SE 2 value. 10 X. Xue, W.R. Reed and A. Menclova / Journal of Health Economics 72 (2020) 102317 The FAT consists of testing the significance of SE. Rejection of H0 : ˛1 = 0 is taken as evidence that publication bias exists. If we divide Eq. (7) by the weights SE and ω, we obtain: PCC i = ˛0 SE i  1  SE i ε +˛1 + i . SE i (3”) and PCC i = ˛0 · ωi 1 ωi + ˛1 ·  SE  i ωi + εi ωi (4”) 6. Data Analysis: part 2 6.1. Meta-regression analysis The preceding section has been concerned with estimating the mean true effect of social capital on health. In this section, we investigate factors that affect the size of that effect. To do that, we add potential moderator variables Xk , k = 1, 2, . . .K, into the specification of Eq. (7): PCC i = ˛0 + ˛1 SE i + K  ˛k+1 Xki + εi (8) k=1 It is standard practice when the meta-analysis sample consists of multiple estimates from the same study to correct for non-independence of the error terms by using cluster robust standard errors. Inclusion of the SE variable also serves to control for the influence of publication bias on the estimate of mean true effect, ˛0 . Thus, the estimate of ˛0 in the specification of Eq. (7) represents the bias-adjusted estimate of the mean true effect of social capital on health. A test of the significance of ˛ ˆ 0 is known as the Precision Effect Test (PET). Rejection of H0 : ˛0 = 0 is taken as evidence that the mean true effect of social capital on health is nonzero. Table 6 reports the FAT/PET results, with the FAT results reported on the first row and the PET results on the second. The first four columns report the various combinations of “Fixed Effects”/“Random Effects” and weighting by individual estimate(“Weight1”)/weighting by study(“Weight2”). Across all four columns, we reject H0 : ˛1 = 0 at the 1 percent level of significance, indicating the existence of publication bias. The positive coefficient indicates positive publication bias, suggesting sample selection that favors the publication of positive estimates of the effect of social capital on health. In three of the four cases we also reject H0 : ˛0 = 0, with the corresponding estimates of ˛0 significant at the 1 percent level. The exception is the “Fixed Effects(Weight1)” regression. Thus, the PET results generally support the finding that social capital is positively and significantly related to health. However, the sizes of the coefficient estimates indicate that this effect is very small. Bias-adjusted estimates of the mean true effect of social capital on health range from 0.004 to 0.022, substantially below the value that Doucouliagos (2011) identifies as being “small”. 5.3. Meta-regression analysis Columns (5) and (6) in Table 6 report the weighted average estimates of mean true effect, uncorrected for publication bias, using the “Random Effects(Weight1)” and “Random Effects(Weight2)” estimators. The associated estimates are 0.024 and 0.032, which are close to the unweighted value of 0.028 reported in Table 3. These fall to 0.013 and 0.022, respectively, when SE is added to the specification to control for publication bias. Thus, while publication bias positively inflates estimates of the effect of social capital on health, it does not inflate them very much. The coefficient ˛k+1 measures the change in the effect of social capital on health due to Xk , where a positive coefficient indicates that studies/regressions that have characteristic Xk estimate a larger effect of social capital on health. The specification of Eq. (8) is known as a metaregression. Given the large number of study, data, and estimation characteristics included in our dataset (cf. Table 4), we are concerned that multicollinearity may disguise significant relationships. Accordingly, we adopt a model selection algorithm to select a “best” specification. We use a backwards stepwise procedure that is designed to select the model specification with the smallest Bayesian Information Criterion value (Lindsey and Sheather, 2010). In the first round, all the variables are included in the regression equation: SE, PubYear, IndividualSC, Panel, EastAsia, WestNorthEurope, HighIncome, OtherCountry, PhysicalHealth, MentalHealth, SelfReported, NumberSCVariables, Age, Gender, Education, MaritalStatus, Income, OLS, ORHazard, HLM, IV, SENonspherical, tNormal, and tCalculatedbypValue. At each subsequent round, the algorithm drops the variable that causes the largest decrease in BIC. It continues to do that, one variable at a time, until the BIC can no longer be reduced. We then re-estimate the final, best model in order to obtain cluster robust standard errors.14 We do this for each of the four estimation procedures (“Fixed Effects(Weight1)”, “Fixed Effects(Weight2)”, “RandomEffects(Weight1)”, and “RandomEffects(Weight2)”. We forced five variables to be retained in each step of the selection process: two variables to indicate the type of health outcome, PhysicalHealth and MentalHealth, with the omitted category being general health; a variable indicating that the respective social capital variable was individual-level as opposed to a community average (IndividualSC); and two variables that represent attempts to deal with endogeneity, Panel and IV. 14 We encountered a problem when using the stepwise regression algorithm with the “Random Effects” estimators. Note that there is no constant term in the weighted specification of equation (4”), as the constant term in the original equation is divided by ωi . The Stata program that we used, vselect, does not allow one to drop the constant term. Our workaround was to estimate the “best” model with a constant term, and then estimate that same variable specification, but without the constant term. Note that was not a problem for the “Fixed Effects” estimators, because the publication bias variable, SE, reduces to the constant term when divided through by SE (see equation 3”). In this case, the “constant” term is actually the coefficient on SE, ˛1 . X. Xue, W.R. Reed and A. Menclova / Journal of Health Economics 72 (2020) 102317 11 Table 6 The Funnel Asymmetry Test (FAT) and Precision Effect Test (PET). Including Publication Bias Term FAT PET Observations Excluding Publication Bias Term Fixed Effects (Weight1) (1) Fixed Effects (Weight2) (2) Random Effects (Weight1) (3) Random Effects (Weight2) (4) Random Effects (Weight1) (5) Random Effects (Weight2) (6) 1.269*** (5.93) 0.004 (1.45) 12,778 1.491*** (6.98) 0.007*** (2.87) 12,778 0.599*** (5.33) 0.013*** (5.15) 12,778 0.551*** (4.34) 0.022*** (7.49) 12,778 ---0.024*** (12.79) 12,778 ---0.032*** (16.17) 12,778 NOTE: The FAT and PET results in Columns (1) through (4) come from estimating ˛1 and ˛0 , respectively, in Eq. (7) in the text using Weighted Least Squares (WLS). The four WLS estimators (Fixed Effects-Weight1, Fixed Effects-Weight2, Random Effects-Weight1, and Random Effects-Weight2) are described in Section 2 of the text. The PET results in Columns (5) and (6) are taken from estimates of Eq. (4’). All of the estimation procedures calculate cluster robust standard errors. *, **, and *** indicate statistical significance at the 10-, 5-, and 1-percent level, respectively. Table 7 Meta-Regression Analysis. Variables Fixed Effects (Weight1) (1) Fixed Effects (Weight2) (2) Random Effects (Weight1) (3) Random Effects (Weight2) (4) SE PhysicalHealth MentalHealth IndividualSC Panel IV PubYear EastAsia WestNorthEurope HighIncome OtherCountry SelfReported NumberSCVariables Age Education MaritalStatus OLS HLM SENonspherical tNormal tCalculatedBypValue Adjusted R-squared Hypothesis Test: Physical = Mental = 0 Observations 0.468*** (3.22) −0.008** (−2.37) −0.006* (−1.95) 0.013*** (4.84) −0.006** (−2.41) −0.000 (−0.06) ---0.006 (1.60) 0.009*** (2.91) 0.003 (0.78) ---0.007** (2.18) −0.001*** (−3.92) −0.003* (−1.96) −0.002* (−1.74) −0.003 (−1.36) 0.006 (1.02) ---−0.003 (−1.10) −0.010*** (−3.47) −0.023*** (−7.72) 0.206 F = 3.15 (p = 0.044) 12,778 0.641*** (4.77) −0.003* (−0.59) −0.004 (−0.99) 0.020*** (5.50) −0.006** (−2.23) 0.002 (0.50) −0.000 (-1.15) 0.006* (1.69) 0.010*** (3.07) 0.006 (1.44) ---0.015*** (2.97) −0.002*** (−4.73) ---−0.004* (−1.92) −0.005** (−2.08) 0.006 (0.86) 0.006 (1.51) −0.007 (−1.66) −0.011** (−3.61) −0.027*** (−6.93) 0.226 F = 0.53 (p = 0.59) 12,778 — −0.012*** (-3.61) −0.006* (-1.93) 0.015*** (3.62) −0.014*** (−4.52) −0.007 (−1.38) ------0.009*** (2.66) ---−0.005 (-1.24) ---−0.001*** (−3.93) −0.004 (−1.06) −0.006** (−2.09) ------— −0.007** (−2.11) ---−0.023*** (−6.39) 0.275 F = 7.13 (p = 0.001) 12,778 — −0.006 (−0.90) −0.003 (−0.72) 0.018*** (4.52) −0.012*** (−2.85) −0.000 (−0.04) −0.000 (−1.37) 0.006 (1.12) 0.007 (1.59) 0.006 (0.81) ---0.010 (1.61) −0.002*** (−5.38) ---−0.010** (−2.44) −0.004 (−1.11) 0.007 (0.98) — −0.006 (−1.46) −0.007 (−1.61) −0.034*** (−6.58) 0.302 F = 0.48 (p = 0.621) 12,778 NOTE: The table reports the results of estimating Eq. (8) in the text. The top value in each cell is the coefficient estimate, and the bottom value in parentheses is the associated t-statistic. The variables PhysicalHealth, MentalHealth, IndividualSC, Panel, and IV were locked into each regression specification. Remaining control variables were selected using a backwards stepwise regression procedure that chooses variables to minimize the Bayes Information Criterion. The four WLS estimators (Fixed Effects-Weight1, Fixed Effects-Weight2, Random Effects-Weight1, and Random Effects-Weight2) are described in Section 2. All four estimation procedures calculate cluster robust standard errors. *, **, and *** indicate statistical significance at the 10-, 5-, and 1-percent level, respectively. The hypothesis test reports the results of testing whether there is no difference in mean PCC values for the three different health outcomes after controlling for the effects of other variables. The results are reported in Table 7. The publication bias term SE is positive and statistically significant at the 1 percent level in two of the four estimation procedures.15 However, the estimated coefficients suggest that the size of the bias is small.16 Thus, while the social capital and health literature may be affected by publication bias, it is of little practical consequence. Neither of the two health outcome variables are consistently significant. PhysicalHealth is significant at the 5-percent level in two of the regressions (FixedEffects- 15 The SE variable is dropped by the selection algorithm for the Random Effects(Weight2) regression. 16 Doucouliagos and Stanley (2013) determine that “the literature suffers from substantial selectivity” if the estimated publication bias coefficient has an absolute value greater than 1. Note that the SE coefficient is less than 1 in all three of the regressions in which it appears. Weight1, RandomEffects-Weight1). MentalHealth never achieves significance. This confirms the casual observation from Fig. 2 regarding the similarity of the PCC distributions for the three different health outcomes. Nevertheless, to be cautious, our subsequent analysis of the social capital variables will – in addition to pooling the PCC values – also divide the full sample into subsamples based on the nature of the health outcomes. In contrast, we find strong evidence that social capital variables based on individual-level measures are more effectual for positive health outcomes than community average measures. However, while the IndividualSC estimates are significant at the 1-percent level in all four regressions, the sizes of the estimated differences in PCC are small, ranging between 0.013 and 0.020. We find weak evidence that correcting for endogeneity diminishes the estimated effect of social capital on health. 12 X. Xue, W.R. Reed and A. Menclova / Journal of Health Economics 72 (2020) 102317 The coefficients for Panel are consistently negative across the four regressions. While significant at the 5-percent level in every case, the sizes of the estimated coefficients are small. The IV estimates are even weaker – very small in size and never attaining significance at even the 10-percent level. As we proceed to the other variables included in the meta-regressions of Table 7, we limit our discussion to those that are significant at the 5 percent level in at least three of the four regressions. The coefficient for WestNorthEurope is positive, indicating that social capital is estimated to have a larger effect on health for residents of Western and Northern Europe compared to other countries. The negative coefficient on NumberSCVariables indicates that the estimated effect of a given social capital variable tends to be smaller when more social capital variables are included in the regression. This is to be expected because the estimated effect variable, PCC, is a function of the t-statistic in the original study. One would expect that the more social capital variables there are in the equation, the more likely collinearity will reduce the significance of any given social capital variable. The last variable in our meta-regression specification is tCalculatedBypValue. For approximately one-fifth of the estimated effects in our sample, the only information related to coefficient standard errors reported by the original study was stars; e.g. three stars to indicate significance at the 1 percent level, two stars to indicate significance at the 5 percent level, and so on. In our original coding, we assigned p-values for each number of stars.17 When the coefficient was insignificant and no stars were reported, we set the p-value equal to 0.50. The negative coefficient estimates for tCalculatedBypValue suggest that this was not a good approximation. Estimated effects calculated from the resulting t-values were significantly lower than those calculated following standard procedures. Further, compared to the other effects for binary variables reported in Table 7, the coefficient sizes are relatively large in absolute value, ranging from −0.023 to −0.034. The takeaway from this is that our estimates of the unconditional mean true effect are downwardly biased by the inclusion of estimated effects using these t-values. However, even after compensating for this, the estimated mean true effect fails to reach the 0.07 value that Doucouliagos identifies as “small”. 6.2. Examination by type of social capital The last part of our analysis carries out a closer examination of the social capital variables included in our sample. The top panel of Table 8 provides some statistical detail about the different kinds of social capital variables that studies have employed. Most studies in our sample use social capital variables that fit within the cognitivestructural framework. Of the 12,778 estimates of the effects of social capital on health in our sample, 11,776 use social capital variables that can either be classified as cognitive or 17 See Footnote #4. structural – approximately 92%. Of these, the great majority are structural: 69.4%, versus 30.6% that are cognitive. In contrast, only 3,572 of the estimated effects fit within the bonding–bridging–linking framework. Indeed, many of these can be cross-classified as either cognitive or structural. The breakdown for the bonding-bridging-linking social capital variables are 62.0% bonding, 34.2% bridging, and only 3.9% linking. The bottom panel of Table 8 reports that there are 3,603 estimated effects that are based on a cognitive social capital variable; 8,173 effects are based on a structural social capital variable; 2,213 are based on a bonding social capital variable; and so on. For each subsample, we calculate the mean, unweighted PCC value unadjusted for publication bias. Mean PCC values for the Structural = 1, Bonding = 1, and Bridging = 1 subsamples are 0.022, 0.023, and 0.025, respectively – very close to the mean value of 0.028 for the entire sample. In contrast, the Cognitive = 1 and Linking = 1 subsamples are different, with the Cognitive subsample showing a higher (0.041), and the Linking subsample a lower (0.010) mean PCC value. To investigate these differences further, we employ the same meta-regression analysis we used in Table 7, only this time we add social capital variables.18 In the first set of exercises, we add Cognitive, so that the omitted social capital variable is Structural. In the next set of exercises, we add Bonding and Bridging for the full sample analysis, with Linking being the omitted variable; and Bonding for the subsample analysis, with Bridging and Linking being the omitted variables19 We then implement the backwards stepwise regression procedure described above. Table 9 displays the results for the Cognitive/Structural framework. In the interests of brevity and to focus attention, we only report estimated coefficients for the social capital variables. The top panel uses the full sample of 11,776 estimates that are derived from either cognitive or structural social capital variables. Across all four regressions, the coefficient for Cognitive is positive and statistically significant at the 1 percent level. This indicates that, in general, estimates of social capital on health that rely on cognitive social capital variables will find larger effects than those that rely on structural social capital variables. The next three panels stratify the sample by the three health outcomes: physical health, mental health, and general health. The Cognitive coefficient is significant twice (at at least the 5-percent level) in the general health subsample, three times in the physical health subsample, and in all four of the regressions for the mental health subsample. However, the estimated values are still small by Doucouliagos’ (2011) standards. 18 Another difference is that we do not force any variables into the equation other than a constant term and the respective social capital variables. In the weighted regressions of equations (3’) and (4’), the “constant term” is actually the slope coefficient on the 1 SE i and 1 ωi terms, respectively. We forced these variables into the equation so that there would always be a constant term in the final regression. 19 We could not add both Bonding and Bridging in the subsample analysis because there were too few observations of Linking. X. Xue, W.R. Reed and A. Menclova / Journal of Health Economics 72 (2020) 102317 13 Table 8 Social Capital Variables. A. Sample Statistics for Social Capital Variables Type of Social Capital Obs Mean Min Max Cognitive Structural Bonding Bridging Linking 11,776 11,776 3,572 3,572 3,572 0.306 0.694 0.620 0.342 0.039 0 0 0 0 0 1 1 1 1 1 B. Distribution of PCC Values by Type of Social Capital Type of Social Capital Obs Mean Min Max Cognitive = 1 Structural = 1 Bonding = 1 Bridging = 1 Linking = 1 3,603 8,173 2,213 1,220 139 0.041 0.022 0.023 0.025 0.010 −0.155 −0.157 −0.157 −0.150 −0.073 0.360 0.360 0.347 0.234 0.308 NOTE: Values in the table report sample statistics of the social capital variables for two sets of observations. The first set of 11,776 observations consist of estimates on health of social capital variables using the cognitive/structural framework. The second set of 3,572 observations consist of estimates based on the bonding/bridging/linking framework. NOTE: Values in the table report conditional mean, minimum, and maximum values of the PCC variable for observations satisfying the condition in the leftmost column. Table 9 Meta-Regression Analysis: Cognitive/Structural. Variables Fixed Effects (Weight1) (1) Fixed Effects(Weight2) (2) Random Effects(Weight1) (3) Random Effects (Weight2) (4) Full sample Cognitive Observations 0.012*** (4.54) 11,776 0.013*** (4.03) 11,776 0.011*** (3.65) 11,776 0.017*** (4.00) 11,776 Physical health = 1 Cognitive Observations 0.012** (2.02) 4,421 0.014*** (3.91) 4,421 0.011** (2.41) 4,421 0.016* (1.96) 4,421 Mental health = 1 Cognitive Observations 0.024*** (4.14) 2,638 0.028*** (4.01) 2,638 0.026*** (4.41) 2,638 0.028*** (3.40) 2,638 General health = 1 Cognitive Observations 0.011*** (3.72) 4,810 0.011*** (3.42) 4,810 0.005 (1.15) 4,810 0.007 (1.60) 4,810 0.008** (2.30) 0.004 (0.45) 0.024** (2.26) 11,776 0.005 (1.18) 0.005 (0.85) 0.022*** (3.27) 11,776 0.008* (1.69) 0.009 (0.85) 0.028*** (2.91) 11,776 Full sample with interaction terms 0.010*** (3.55) Cognitive −0.001 (−0.17) Cognitive*Physical 0.015** (2.22) Cognitive*Mental 11,776 Observations NOTE: The table reports the results of estimating Eq. (8) in the text for different samples of estimates, using different estimation procedures. Only the coefficient estimates for the social capital variables are reported. The top value in each cell is the coefficient estimate, and the bottom value in parentheses is the associated t-statistic. In addition to a constant term, the variable Cognitive was locked into the regressions in the first four sets of regressions (Full Sample, Physical Health = 1, Mental Health = 1, General Health = 1). The last set of regressions (Full Sample with Interaction Terms) also locked in the interaction terms Cognitive*Physical and Cognitive*Mental. The remaining control variables were selected using a backwards stepwise regression procedure that chooses variables to minimize the Bayes Information Criterion. The four WLS estimators (Fixed Effects-Weight1, Fixed Effects-Weight2, Random Effects-Weight1, and Random Effects-Weight2) are described in Section 2. All four calculate cluster robust standard errors. *, **, and *** indicate statistical significance at the 10-, 5-, and 1-percent level, respectively. The last panel of Table 9 again uses the full sample, but includes interaction terms for cognitive social capital and physical health (Cognitive*Physical), and cognitive social capital and mental health (Cognitive*Mental). The omitted health category is GeneralHealth. The coefficients should be interpreted as follows: The coefficient on Cognitive represents the difference between the mean estimated effect of cognitive social capital on general health, and the mean estimated effect of structural social capital on any kind of health. The coefficient on Cognitive*Physical (or Cognitive*Mental) represents the difference between the mean estimated effect of cognitive social capital on physical health (or mental health) compared to the mean estimated effect of structural social capital on any kind of health. Across the four regressions in Table 9, only the coefficient for Cognitive*Mental is consistently significant at the 5 percent level. The mean partial correlation of cognitive social capital and health is estimated to be approximately 0.02 to 0.03 larger than the mean partial correlation of structural social capital on any kind of health. Together with the previous results, these estimates indicate that cog- 14 X. Xue, W.R. Reed and A. Menclova / Journal of Health Economics 72 (2020) 102317 Table 10 Meta-Regression Analysis: Bonding/Bridging/Linking. Variables Fixed Effects (Weight1) (1) Fixed Effects (Weight2) (2) Random Effects (Weight1) (3) Random Effects (Weight2) (4) Full sample Bonding Bridging Ho : Bonding = Bridging = 0 Observations 0.002 (0.80) 0.000 (0.04) F = 0.32 p-value = 0.724 3,572 0.002 (1.42) −0.000 (−0.07) F = 1.02 p-value = 0.362 3,572 0.010* (1.74) 0.011* (1.84) F = 1.81 p-value = 0.166 3,572 0.017** (2.38) 0.016** (2.45) F = 3.33 p-value = 0.038 3,572 Physical health = 1 Bonding Observations −0.001 (-0.50) 1,023 0.004 (1.43) 1,023 0.009 (1.50) 1,023 0.015** (2.31) 1,023 Mental health = 1 Bonding Observations 0.008 (0.96) 1,004 0.003 (0.24) 1,004 0.005 (0.81) 1,004 −0.002 (−0.37) 1,004 General health = 1 Bonding Observations 0.001 (0.10) 1,597 0.003 (0.37) 1,597 −0.005 (−1.13) 1,597 0.000 (0.08) 1,597 0.005 (0.64) −0.003 (-0.38) −0.009 (-0.83) 3,572 −0.006 (−1.50) 0.014* (1.79) 0.010 (1.35) 3,572 0.004 (0.56) 0.003 (0.28) −0.007 (−0.68) 3,572 Full sample with interaction terms Bonding −0.001 (−0.15) 0.004 (0.47) Bonding*Physical 0.003 (0.56) Bonding*Mental 3,572 Observations NOTE: The table reports the results of estimating Eq. (8) in the text for different samples of estimates, using different estimation procedures. Only the coefficient estimates for the social capital variables are reported. The top value in each cell is the coefficient estimate, and the bottom value in parentheses is the associated t-statistic. In addition to a constant term, the following variables were locked into the regressions: Bonding and Bridging for the set of Full Sample regressions; Bonding for the set of Physical Health = 1, Mental Health = 1, General Health = 1 regressions; and Bonding, Bonding*Physical and Bonding*Mental for the Full Sample with Interaction Terms set of regressions. The remaining control variables were selected using a backwards stepwise regression procedure that chooses variables to minimize the Bayes Information Criterion. The four WLS estimators (Fixed Effects-Weight1, Fixed EffectsWeight2, Random Effects-Weight1, and Random Effects-Weight2) are described in Section 2. All four calculate cluster robust standard errors. *, **, and *** indicate statistical significance at the 10-, 5-, and 1-percent level, respectively. The hypothesis test in the Full Sample set of regressions reports the results of testing whether there is any difference in the estimated effects on health for the Bonding, Bridging, and Linking social capital variables. Table 11 Mean PCC Values and the Distribution of t-values for Finer Categories of Social Capital and Health. Obs Mean PCC Distribution of t-values (1) (2) t < -2.00 (3) −2.00 ≤ t ≤ 2.00 (4) t > 2.00 (5) Sub-categories of Cognitive Social Capital Social trust 1,553 863 Perceived social support Perceived social cohesion 237 179 Perceived reciprocity 173 Sense of belonging 114 Loneliness 0.045 0.038 0.043 0.053 0.038 −0.010 3.0 5.0 3.8 0.6 4.6 28.9 40.9 62.1 61.2 60.9 53.8 48.2 56.1 32.9 35.0 38.5 41.6 22.8 Sub-categories of Structural Social Capital 1,965 Social participation 1,931 Social networks 785 Social support 769 Social engagement 567 Volunteering 537 Group membership 456 Social integration 411 Social relationship 0.034 0.015 0.020 0.019 0.024 0.022 0.007 0.020 5.3 9.7 8.1 5.7 4.4 9.7 15.6 17.0 51.3 68.8 69.8 57.5 51.8 61.4 57.9 49.1 43.3 21.5 22.0 36.8 43.7 28.9 26.5 33.8 Sub-categories of Health Mortality Disease/Illness Depression 0.020 0.014 0.028 9.5 5.5 8.6 62.3 75.9 59.8 28.2 18.5 31.6 2,441 830 826 nitive social capital is particularly salient for mental health, as opposed to physical or general health. Table 10 performs a similar set of exercises for the Bonding/Bridging/Linking social capital variables. The top panel pools all the estimates that are based on these social capital variables. We include dummy variables for Bonding and Bridging social capital variables. With two exceptions (Bonding/Random Effects-Weight2; Bridging/Random EffectsWeight2), the associated coefficients are all insignificant at the 5-percent level. Further, when we test the joint hypothesis that both coefficients equal zero, we fail to reject it in three of the four regressions. These results suggest that X. Xue, W.R. Reed and A. Menclova / Journal of Health Economics 72 (2020) 102317 there is little difference between bonding, bridging, and linking social capital variables with respect to their effect on health. The next three panels break the full sample into three subsamples based on health outcomes. Due to the small number of estimates that use linking social capital variables in each of the subsamples, we are forced to combine linking and bridging social capital as a single omitted category, so that the only included social capital variable is Bonding. Across the three subsamples, Bonding is significant only once, in column (4) for the physical health subsample. When the three subsamples are combined again in the bottom panel, with interaction terms for Bonding and Physical Health, and Bonding and Mental Health, the respective interaction terms are everywhere small in size and statistically insignificant. This supports the finding from the top panel in Table 10 that there is no systematic difference in how bonding, bridging and linking social capital affect health. Table 11 provides one last deep dive into the finer categories of the social capital and health variables. Reported are the number of estimates associated with the respective subcategory (Column 2), the mean PCC value for those estimates (Column 3), and the percent of associated t-statistics that are (i) less than −2.00, (ii) between −2 and 2, and (iii) greater than 2.00 (Columns 3–5). Not a single subcategory of social capital or health rises to the level of “small” using Doucouliagos’ PCC (2011) guidelines. The social capital sub-categories with the largest mean PCC values are Social trust (0.045) and Perceived reciprocity (0.053). With only two exceptions (Social trust and Loneliness), over half of all estimated coefficients are statistically insignificant (−2.00 ≤ t ≤ 2.00). This is all the more noteworthy given the large sizes of the samples underlying these estimates (cf. Table 3). 7. Conclusion The relationship between social capital and health has received extensive attention in fields such as public health, medicine, epidemiology, gerontology and other healthrelated disciplines. Hundreds of studies have reported that social capital is significantly associated with a variety of positive health outcomes. In contrast, the economics literature on this subject is relatively small. The primary aim of this study has been to determine whether the effect of social capital on health is “large” or “small”. If “large”, then the dearth of economic research in this area represents a gap that should be addressed. If “small”, then this justifies the relative lack of attention this subject has received in the health economics discipline. We use meta-analysis to analyze the effects of 12,778 estimates from 470 studies. To place estimated effects on a common scale, we transform the individual estimates into partial correlation coefficients and compare these against the “size” guidelines established by Doucouliagos (2011). Our analysis finds that social capital is significantly related to a variety of positive health outcomes. However, the effects are consistently very small. This result is robust across a wide variety of different types of social capital (e.g., cognitive, structural, bonding, bridging, linking), and 15 for many different measures of health outcomes (e.g., mortality, disease/illnesses, depression). One of the most compelling findings from our analysis is that the majority of estimates in the social capital/health literature are statistically insignificant. This is striking given the typically large sample sizes of the studies included in our analysis: Median sample size for the studies in our sample was 3,451. Over 75% of the estimates came from regressions with more than 1000 observations. Large sample sizes with high rates of statistical insignificance are indicative of small effect sizes. How could such a large literature grow from such small effects? We suspect several factors are responsible. First, estimated coefficients are difficult to interpret. Social capital variables are typically measured using Likert scale type questions. As a result, small effect sizes are not recognized as such. Second, and perhaps as a consequence of the first factor, researchers place undue emphasis on statistical significance. This is compounded by the fact that studies typically include multiple social capital variables in the same regression. The mean number of social capital variables in a regression in our sample was 6.5. The combination of large sample sizes with multiple social capital variables makes it relatively easy to find at least one social capital variable that is statistically significant. Studies rarely apply Bonferroni-type corrections when assessing the statistical significance of multiple estimates. This explains how most studies can conclude that social capital has a significant effect on health, even while the majority of estimates are insignificant. Our study is the most extensive analysis of the extant empirical literature on social capital and health ever undertaken. As such, it can be viewed as establishing a Bayesian prior belief that the effect of social capital on health is very small. The results suggest that economics is justified in not following the leads of other disciplines in devoting substantial research resources to the study of social capital and health. Our study also calls into question initiatives to promote health through social capital such as those by the WHO, the OECD, and US Healthy People 2020 (Rocco and Suhrcke, 2012; Centers for Disease Control and Prevention, 2010; Keeley, 2007). In a recent systematic review, VillalongaOlives et al. (2018) note the “enthusiasm of policy makers to implement social capital interventions to manipulate health outcomes for the better” (p. 217). Our results suggest that this enthusiasm may not translate into meaningful improvements in public health. CRediT authorship contribution statement Xindong Xue: Conceptualization, Investigation, Project administration, Writing - review & editing. W. Robert Reed: Methodology, Formal analysis, Supervision, Data curation, Validation, Writing - original draft. Andrea Menclova: Investigation, Writing - review & editing. Acknowledgements Xindong Xue acknowledges financial support from the National Social Science Fund of China (Grant No.: 16 X. Xue, W.R. Reed and A. Menclova / Journal of Health Economics 72 (2020) 102317 14BRK013), the Humanities and Social Sciences Fund from Chinese Ministry of Education (Grant No.: 20YJC630011) and the Youth Innovation Research Project from Hubei Province (Grant No.: T201932) W. Robert Reed acknowledges financial support from the Czech Science Foundation, Grant 18-02513S. We are grateful to Xiaoping Ruan, Lina Wang, Weilun Wu, Yi Xia, Junyong Yuan, and Kunpeng Zhao for their excellent research assistance. We thank Tom Coupé, Lorenzo Rocco, and participants at the 2019 MAERNet Colloquium for insightful comments and suggestions. All remaining errors are our own. Appendix A. Supplementary data Supplementary material related to this article can be found, in the online version, at https://doi.org/10.1016/j. jhealeco.2020.102317. References Alesina, A., Giuliano, P., 2015. Culture and institutions. J. Econ. Lit. 53 (4), 898–944. Algan, Y., Cahuc, P., 2010. Inherited trust and growth. Am. Econ. Rev. 100 (5), 2060–2092. Alinaghi, N., Reed, W.R., 2018. Meta-analysis and publication bias: How well does the FAT-PET-PEESE procedure work? Res. Synth. Methods 9 (2), 285–311. Andini, C., Andini, M., 2018. Social capital and growth: causal evidence from Italian municipalities. J. Econ. Geogr. 19 (3), 619–653. Andrews, I., Kasy, M., 2017. Identification of and Correction for Publication Bias. NBER Working Paper 23298. National Bureau of Economic Research. Berggren, N., Jordahl, H., 2006. Free to trust: economic freedom and social capital. Kyklos 59 (2), 141–169. Beugelsdijk, S., de Groot, H.L.F., van Schaik, A.B.T.M., 2004. Trust and economic growth: a robustness analysis. Oxf. Econ. Pap. 56, 118–134. Blakely, T., Atkinson, J., Ivory, V., Collings, S., Wilton, J., Howden-Chapman, P., 2006. No association of neighbourhood volunteerism with mortality in New Zealand: a national multilevel cohort study. Int. J. Epidemiol. 35 (4), 981–989. Bowles, S., Gintis, H., 2002. Social capital and community governance. Econ. J. 112 (483), F419–F436. Centers for Disease Control and Prevention, 2010. Healthy People 2020, https://www.healthypeople.gov/2020/topics-objectives/topic/ social-determinants-of-health. accessed at 2018/11/28. Christensen, G., Miguel, E., 2018. Transparency, reproducibility, and the credibility of economics research. J. Econ. Lit. 56 (3), 920–980. Cohen, J., 1988. Statistical Power Analysis in the Behavioral Sciences, 2nd ed. Hillsdale: Erlbaum. Coleman, J.S., 1990. Foundations of Social Theory. Harvard University Press, Cambridge, MA. Costa-Font, J., Mladovsky, P., 2008. Social capital and the social formation of health-related preferences and behaviours. Health Econ. Policy Law 3 (4), 413–427. De Silva, M.J., McKenzie, K., Harpham, T., Huttly, S.R., 2005. Social capital and mental illness: a systematic review. J. Epidemiol. Community Health 59 (8), 619–627. de Vaan, M., Frenken, K., Boschma, R., 2019d. The downside of social capital in new industry creation. Econ. Geogr., 1–26. Doucouliagos, H., 2011. How Large Is Large? Preliminary and Relative Guidelines for Interpreting Partial Correlations in Economics. Economics Series 2011/5. Deakin University, Faculty of Business and Law, School of Accounting, Economics and Finance. Doucouliagos, H., Paldam, M., 2013. The robust result in meta-analysis of aid effectiveness: a response to Mekasha and Tarp. J. Dev. Stud. 49 (4), 584–587. Doucouliagos, H., Stanley, T.D., 2013. Are all economic facts greatly exaggerated? Theory competition and selectivity. J. Econ. Surv. 27 (2), 316–339. Durlauf, S.N., 2002. On the empirics of social capital. Econ. J. 112 (483), F459–F479. Folland, S., 2008. An economic model of social capital and health. Health Econ. Policy Law 3 (4), 333–348. Folland, S., Kaarbøe, O., Islam, K., 2014. Chapter 3: how Do We invest in social Capital? An exploration of an economic model of social Capital and health. In: Folland, S., Rocco, L. (Eds.), The Economics of Social Capital and Health. World Scientific Publishing. Franco, A., Malhotra, N., Simonovits, G., 2014. Publication bias in the social sciences: unlocking the file drawer. Science 345 (6203), 1502–1505. Gilbert, K.L., Quinn, S.C., Goodman, R.M., Butler, J., Wallace, J., 2013. A meta-analysis of social capital and health: a case for needed research. J. Health Psychol. 18 (11), 1385–1399. Glaeser, E.L., Laibson, D.I., Scheinkman, J.A., Soutter, C.L., 2000. Measuring trust. Q. J. Econ. 115 (3), 811–846. Glaeser, E.L., Laibson, D., Sacerdote, B., 2002. An economic approach to social capital. Econ. J. 112 (483), F437–F458. Grossman, M., 1972. On the concept of health capital and the demand for health. J. Polit. Econ. 80 (2), 223–255. Guiso, L., Sapienza, P., Zingales, L., 2004. The role of social capital in financial development. Am. Econ. Rev. 94 (3), 526–556. Gunby, P., Jin, Y., Reed, W.R., 2017. Did FDI really cause Chinese economic growth? A meta-analysis. World Dev. 90, 242–255. Helliwell, J.F., 2006. Well-being, social capital and public policy: what’s new? Econ. J. 116 (510), C34–C45. Hoi, C.K.S., Wu, Q., Zhang, H., 2019. Does social capital mitigate agency problems? Evidence from Chief Executive Officer (CEO) compensation. J. Financ. Econ. 133 (2), 498–519. Holt-Lunstad, J., Smith, T.B., Layton, J.B., 2010. Social relationships and mortality risk: a meta-analytic review. PLoS Med. 7 (7), e1000316. Karlan, D.S., 2005. Using experimental economics to measure social capital and predict financial decisions. Am. Econ. Rev. 95 (5), 1688–1699. Keeley, B., 2007. Human capital: how what you know shapes your life. In: OECD Insights. OECD publishing, Paris. Knack, S., Keefer, P., 1997. Does social capital have an economic payoff? A cross-country investigation. Q. J. Econ. 112 (4), 1251–1288. LaPorta, R., Lopez-de-Silanes, F., Shleifer, A., Vishny, R., 1997. Trust in large organisations. American Economic Review Papers and Proceedings 87, 333–338. Laporte, A., 2014. Chapter 4: social capital: an economic perspective. In: Folland, S., Rocco, L. (Eds.), The Economics of Social Capital and Health. World Scientific Publishing. Lindsey, C., Sheather, S., 2010. Variable selection in linear regression. Stata J. 10 (4), 650–669. McCoy, S.J., McDonough, I.K., Roychowdhury, P., 2019. The impact of terrorism on social capital: evidence from the 2015 Charlie Hebdo Paris shooting. Oxf. Bull. Econ. Stat., forthcoming. Nannicini, T., Stella, A., Tabellini, G., Troiano, U., 2013. Social capital and political accountability. Am. Econ. J. Econ. Policy 5 (2), 222–250. Nyqvist, F., Pape, B., Pellfolk, T., Forsman, A.K., Wahlbeck, K., 2014. Structural and cognitive aspects of social capital and all-cause mortality: a meta-analysis of cohort studies. Soc. Indic. Res. 116 (2), 545–566. Paldam, M., Svendsen, G.T., 2000. An essay on social capital: looking for the fire behind the smoke. Eur. J. Polit. Econ. 16 (2), 339–366. Putnam, R.D., 1993. Making Democracy Work: Civic Traditions in Modern Italy. Princeton University Press, Princeton, NJ. Putnam, R.D., 1995. Bowling alone: america’s declining social capital. J. Democr. 6, 65–78. Reed, W.R., 2015. A Monte Carlo analysis of alternative meta-analysis estimators in the presence of publication bias. Econ. Open-access Open-assessment E-journal 9 (2015–30), 1–40, http://dx.doi.org/10.5018/economics-ejournal.ja.2015-30. Rocco, L., Suhrcke, M., 2012. Is Social Capital Good for Health? A European Perspective. WHO Regional Office for Europe, Copenhagen. Stanley, T.D., Doucouliagos, H., Giles, M., Heckemeyer, J.H., Johnston, R.J., Laroche, P., Nelson, J.P., Paldam, M., Poot, J., Pugh, G., Rosenberger, R.S., 2013. Meta-analysis of economics research reporting guidelines. J. Econ. Surv. 27 (2), 390–394. Stanley, T.D., Doucouliagos, H., Ioannidis, J.P., 2017. Finding the power to reduce publication bias. Stat. Med. 36 (10), 1580–1598. Tabellini, G., 2010. Culture and institutions: economic development in the regions of Europe. J. Eur. Econ. Assoc. 8 (4), 677–716. Villalonga-Olives, E., Wind, T.R., Kawachi, I., 2018. Social capital interventions in public health: a systematic review. Soc. Sci. Med. 212, 203–218. Wang, S., 2019. Social capital and rotating labor associations in rural China. China Econ. Rev. 53, 243–253. Zak, P.J., Knack, S., 2001. Trust and growth. Econ. J. 111 (470), 295–321.