What factors contribute to explaining the heterogeneity in the reported empirical results on the globalization-spending relationship? We continue by identifying the likely sources of heterogeneity.
In line with standard techniques for meta-regression analysis (e.g., Stanley and Doucouliagos
2012), we adopt the assumption that the
ith estimate of the globalization-spending partial correlation coefficient from study
j, denoted
rij, is influenced not only by sampling error (
\(\varepsilon_{ij}\)), but by a vector of variables (
\(Z_{kij}\)) consisting of characteristics that capture differences in the underlying effect of globalization on government spending. The meta-regression model can thus be written as follows:
$$r_{ij} = \beta_{0} + \sum \beta_{k} Z_{kij} + \beta_{1} {\text{SE}}_{ij} + \varepsilon_{ij} .$$
(4)
By estimating Eq. (
4), we can account simultaneously for publication selection bias (the standard error SE is still included) and for the factors that might explain excess heterogeneity. We estimate Eq. (
4) by WLS with the inverse of the variances as optimal weights. Stanley and Doucouliagos (
2017) argue that the WLS estimator is preferable to other standard estimators that can be applied in meta-regressions. WLS is preferred since the estimates do not have the same variances and because it is important to assign more weight to those estimates that are more precise. However, we also will conduct robustness checks based on applying different estimators. As most of the studies in our meta-study database report several estimates, we also correct for potential within-study dependence by clustering the standard errors obtained from the meta-regression model at the study-level.
5.1.1 Measures of the dependent variable (government spending)
Recall that the dependent variable in the meta-analysis is a measure of public spending (see Eq.
1). We account for differences in the dependent variable by distinguishing estimates that use total government spending, government consumption, public investment, social spending, education spending, health spending and other (unspecified) types of spending (i.e., variables consisting of mixes of spending categories), respectively (see Table
2). By considering differences in the measures of government spending, we are able to test the extended versions of the two competing hypotheses regarding the impact of globalization on public expenditures introduced in Sect.
2.2 (‘efficiency’ vs. ‘compensation’).
Table 2
Variables used in the meta-regression analysis
Partial correlation | Partial correlation of the impact of economic globalization on government spending (dependent variable in the MRA) | 0.051 | 0.208 |
Government spending measures |
Total spending (used as the base) | BD = 1: Total government spending as dependent variable | 0.175 | 0.380 |
Government consumption | BD = 1: Government consumption as dependent variable | 0.190 | 0.393 |
Public investment | BD = 1: Public investment as dependent variable | 0.008 | 0.087 |
Social spending | BD = 1: Social spending as dependent variable | 0.263 | 0.441 |
Education spending | BD = 1: Education spending as dependent variable | 0.051 | 0.220 |
Health spending | BD = 1: Health spending as dependent variable | 0.044 | 0.205 |
Other spending | BD = 1: (Unspecified) dependent spending variables that consist of a mix of spending categories | 0.269 | 0.444 |
Economic globalization measures |
Trade globalization (used as the base) | BD = 1: Trade globalization variable as relevant regressor | 0.668 | 0.471 |
Financial globalization | BD = 1: Financial globalization variable as relevant regressor | 0.303 | 0.460 |
Overall economic globalization | BD = 1: Overall economic globalization variable as relevant regressor | 0.030 | 0.170 |
Data characteristics | | | |
Cross section | BD = 1: Cross sectional data used | 0.164 | 0.371 |
Mean year data | Mean year of the time sample in the data | 1987.7 | 7.384 |
Country composition | | | |
AdvancedCountries (used as the base) | BD = 1: only advanced countries included in the data | 0.437 | 0.496 |
DevelopingCountries | BD = 1: only developing countries included in the data | 0.198 | 0.399 |
Mix of Countries | BD = 1: mix of advanced and developing countries included in the data | 0.371 | 0.483 |
Econometric details | | | |
Level level (used as the base) | BD = 1: Spending variable and globalization variable in levels | 0.684 | 0.465 |
Change change | BD = 1: Spending variable and globalization variable in first differences | 0.135 | 0.341 |
Level change | BD = 1: Spending variable in levels and globalization variable in first differences. | 0.015 | 0.123 |
Change level | BD = 1: Spending variable in first differences and globalization variable in levels. | 0.166 | 0.372 |
GMM (used as the base) | BD = 1: GMM estimator used | 0.148 | 0.355 |
OLS | BD = 1: OLS estimator used | 0.734 | 0.442 |
Random effects | BD = 1: Random effects estimator used | 0.056 | 0.230 |
Other estimator | BD = 1: Other estimator used (e.g., IV, SUR) | 0.063 | 0.242 |
Country fixed effects | BD = 1: Country dummies included | 0.689 | 0.463 |
Time fixed effects | BD = 1: Time dummies included | 0.458 | 0.498 |
Globalization interacted | BD = 1: Interaction of globalization variables with some other variables included | 0.115 | 0.319 |
Publication characteristics | | | |
Standard error | Standard error of partial correlation | 0.088 | 0.057 |
Economics journal | BD = 1: Study published in an economics journal | 0.639 | 0.481 |
Primary | BD = 1: Link globalization-spending is the primary issue of interest | 0.646 | 0.479 |
Cross author | BD = 1: Author declares receiving feedback from other authors who have published in this area | 0.315 | 0.465 |
Prior | BD = 1: Author has published previously in this area | 0.273 | 0.446 |
Macroeconomic, political and institutional control variables |
GDP growth | BD = 1: GDP growth included as control | 0.196 | 0.397 |
Unemployment | BD = 1: Unemployment included as control | 0.332 | 0.471 |
Terms of trade risk | BD = 1: Terms of trade risk included as control | 0.052 | 0.223 |
Income level | BD = 1: GDP per capita included as control | 0.771 | 0.421 |
Old age | BD = 1: Old age population included as control | 0.770 | 0.421 |
Partisan politics | BD = 1: Partisan politics as control | 0.238 | 0.426 |
Labor power | BD = 1: Labor power included as control | 0.091 | 0.288 |
Democracy | BD = 1: Democracy variable included | 0.174 | 0.380 |
5.1.2 Measures of economic globalization
As explained in Sect.
2.1, numerous indicators have been applied in capturing different aspects of globalization. We follow the typology of economic globalization indicators in Gräbner et al. (
2018), distinguishing between trade globalization, financial globalization and overall economic globalization indices; the latter combine both the trade and financial dimensions (see the appendix for more details).
5.1.3 Data characteristics
We consider whether a study relies on cross-sectional data rather than panel data. Additionally, Rodrik (
2011) suggests something special about the period of ‘hyper-globalization’ during the 1990s. Including a continuous variable for the mean year in the various published samples is an approach for testing whether the time dimension of the data matters.
5.1.4 Country composition
The globalization-spending relationship could be influenced by the underlying country sample (e.g., Rudra
2002). We thus control for whether an estimate is based on advanced countries, developing countries or a mix of the two. To distinguish between the three country groups, we make use of the IMF’s (
2018) classifications.
5.1.5 Econometric details
We test formally whether the functional form of the econometric model makes a difference, as suggested by a number of contributions to the relevant literature (e.g., Garrett
2001; Kittel and Obinger
2003; Kittel and Winner
2005). We distinguish between estimates that specify both the dependent variable and the relevant globalization regressor in levels and other functional forms of the globalization-spending relation. We identify specifications in first differences on both side of the empirical model (ChangeChange), specifications that regress the level of government spending on first differences in the relevant globalization variables (LevelChange), and specifications that regress first difference in government spending on the level of globalization (ChangeLevel).
We account for differences in the empirical estimators by coding a variable for GMM, OLS, Random Effects and other estimators (such as Instrumental Variables and Seemingly Unrelated Regressions), respectively. We also check whether differences emerge in the reported results when estimates control for unobserved country heterogeneity (CountryFixedEffects) and time-varying shocks affecting all countries (TimeFixedEffects). Furthermore, we consider whether the regression model entered an interaction term between globalization and some other variable.
5.1.6 Publication characteristics
We account for various dimensions of the publication process: differences between economics journals and other types of journals, whether a study’s primary focus is the globalization-spending relationship as opposed to entering globalization merely as a control variable, and whether the author(s) of a study previously received comments or feedback from other authors who have contributed to the relevant literature.
8
5.1.7 Macroeconomic, political and institutional control variables
Finally, we consider whether the inclusion of potentially relevant control variables has an impact on the reported relationship between globalization and government spending. Several studies account for macroeconomic factors, represented by GDP growth and the unemployment rate, respectively. We also check whether studies control for income levels (proxied by GDP per capita). Some authors have argued that government spending plays an important risk-reducing role, particularly in economies that are exposed to substantial external risks (e.g., Rodrik
1998; Epifani and Gancia
2009); hence, we consider whether the sampled studies control for terms-of-trade risk. An important argument is that demographic change affects government spending, with particular importance for an aging population (e.g., Shelton
2008); we thus check whether it matters whether or not the share of the aged population is accounted for. The political science literature has emphasized that the identities of the parties running the government matter, since their ideological leanings may lead to quite different spending policies, even in the context of globalization (e.g., Cusack
1997; Belke
2000; Kittel and Obinger
2003). We thus also ask whether entering controls for partisan politics affects reported partial correlations. The political science literature likewise highlights the salience of labor’s institutional powers in the context of government spending regressions (e.g., Iversen and Cusack
2000). Finally, democracy potentially is a relevant moderator variable, as more democratic political systems may be generous government spenders, especially on social security and welfare. One explanation for that relation could be that the median voter prefers redistributive policies, providing incentives for governments to cater to those preferences (e.g., Meltzer and Richard
1981).
A few preliminary remarks are in order regarding the interpretation of the coefficients in the estimates of multivariate meta-regression models presented below. The models always omit one category (as the reference category) from each group of mutually exclusive and jointly exhaustive dummy variables (e.g., government spending or economic globalization measures). That is necessary to avoid perfect multicollinearity implying that the intercept \(\beta_{0}\) cannot be interpreted as the “true” effect of globalization on government spending because it incorporates the effects of the reference groups. Other specifications would yield different estimates of the intercept term. Our specification provides an estimate of the impact of trade globalization on total government spending, based on data from advanced countries only—with all variables entered in their levels.
Note that the choice of the omitted (reference) categories in no way influences any of the other estimated coefficients, but it shifts the empirical model’s intercept
\(\beta_{0}\).
9 Hence, the coefficients of the moderator variables from each group of mutually exclusive and jointly exhaustive dummy variables allow us to make predictions regarding the impact of economic globalization on government spending based on the chosen reference categories. For example, the estimated average partial correlation of social spending compared to total government spending (which is the reference category) can be predicted by adding up the value of the intercept
\(\beta_{0}\) and of the social spending coefficient. Also be aware that we adopt a general-to-specific estimation approach, which is the suggestion of prominent guidelines for meta-analysis (e.g., Stanley and Doucouliagos
2012, p. 105). In particular, we eliminate the variable that returns the largest
p value when all coded moderators in Table
2 are entered and repeat that step until the
p values of all variables are smaller than 0.1.
10 The advantage of the general-to-specific approach is “that model construction proceeds from a very general model in a more structured, ordered and (statistically valid) fashion, and in this way avoids the worst of data mining” (Charemza and Deadman
1997, p. 78).
Table
3 shows the general-to-specific modeling results from the multivariate meta-regression analysis.
11 Column (1) applies WLS with the standard errors clustered at the study level. The social spending variable returns a negative and statistically significant coefficient, indicating that—compared to estimates entering total government spending as the dependent variable (which is excluded as the reference category in the group of government spending measures)—estimates that model social spending alone report larger negative effects of globalization on total government spending. Indeed, our model predicts that globalization has a small-to-moderate negative impact on social spending of − 0.1 (obtained by adding the social spending coefficient to the intercept). While the finding of a negative and statistically significant coefficient on the social spending variable in column (1) clearly contradicts the extended version of the ‘compensation’ hypothesis, which predicts that globalization has a differential impact on various components of government spending and, in particular, will push up spending on social protection to compensate for the greater risks associated with international market integration (e.g., Rodrik
1998), it lends support to the extended version of the ‘efficiency’ hypothesis, which predicts the opposite, namely that social spending will be under more pressure from globalization than other components of public expenditures (e.g., Schulze und Ursprung 1999; see also Sect.
2.2). Notably, however, the size of the impact of globalization on social spending predicted by the model estimated in column (1) is small-to-moderate when its interpretation is based on the guidelines proposed by Doucouliagos (
2011), which suggest that a partial correlation coefficient between zero and 0.07 must be considered to be small. In other words, the regression results do not predict a large negative impact of economic globalization on social spending, on average, but rather small-to-moderate downward pressure.
Table 3
Multivariate regression results
Estimator | WLS | Random Effects | Robust regression | WLS | Robust, statistically significant |
Constant | − 0.046 | − 0.018 | 0.032 | − 0.048 | No |
| (0.040) | (0.050) | (0.058) | (0.043) | |
Standard error | 0.604** | 0.613** | 0.510* | 0.710** | Yes |
| (0.278) | (0.272) | (0.267) | (0.312) | |
Government spending measures |
Social spending | − 0.051** | − 0.068*** | − 0.081** | − 0.058*** | Yes |
| (0.021) | (0.024) | (0.032) | (0.022) | |
Public investment | 0.159 | 0.096 | 0.029 | 0.242 | No |
| (0.107) | (0.117) | (0.087) | (0.191) | |
Government consumption | − 0.013 | − 0.027 | − 0.043 | − 0.018 | No |
| (0.025) | (0.025) | (0.033) | (0.027) | |
Education spending | − 0.039 | − 0.040 | − 0.048 | − 0.046 | No |
| (0.027) | (0.027) | (0.030) | (0.029) | |
Health spending | − 0.021 | − 0.029 | − 0.041 | − 0.026 | No |
| (0.022) | (0.026) | (0.029) | (0.024) | |
Other spending | − 0.010 | − 0.021 | − 0.028 | − 0.012 | No |
| (0.022) | (0.023) | (0.026) | (0.023) | |
Economic globalization measures |
Financial globalization | 0.041*** | 0.050*** | 0.058*** | 0.041*** | Yes |
| (0.013) | (0.016) | (0.015) | (0.014) | |
Overall economic globalization | 0.054*** | 0.057** | 0.060 | 0.057** | No |
| (0.022) | (0.027) | (0.039) | (0.023) | |
Country composition |
Developing countries only | 0.121*** | 0.124*** | 0.126*** | 0.134*** | Yes |
| (0.039) | (0.033) | (0.040) | (0.044) | |
Mix of countries | 0.024 | 0.020 | 0.026 | 0.021 | No |
| (0.019) | (0.021) | (0.026) | (0.020) | |
Econometric details | | | | | |
Change change | − 0.109*** | − 0.116*** | − 0.113*** | − 0.117*** | Yes |
| (0.022) | (0.025) | (0.028) | (0.024) | |
Level change | − 0.115*** | − 0.125** | − 0.142** | − 0.125*** | Yes |
| (0.034) | (0.051) | (0.068) | (0.037) | |
Change level | − 0.101*** | − 0.113*** | − 0.123*** | − 0.108*** | Yes |
| (0.019) | (0.023) | (0.029) | (0.021) | |
OLS | 0.081** | 0.073*** | 0.053** | 0.088** | Yes |
| (0.035) | (0.036) | (0.026) | (0.039) | |
Random effects | 0.069 | 0.062 | 0.046 | 0.075 | No |
| (0.070) | (0.060)) | (0.043) | (0.077) | |
Other estimator | − 0.011 | − 0.018 | − 0.045 | − 0.011 | No |
| (0.030) | (0.033) | (0.037) | (0.032) | |
Country fixed effects | − 0.054*** | − 0.064*** | − 0.083*** | − 0.058*** | Yes |
| (0.020) | (0.018) | (0.024) | (0.020) | |
Publication characteristics as well as macroeconomic, political and institutional control variables |
Economics journal | 0.070*** | 0.068*** | 0.061** | 0.075*** | Yes |
| (0.020) | (0.020) | (0.026) | (0.022) | |
Prior | − 0.040*** | − 0.044*** | − 0.047** | − 0.044*** | Yes |
| (0.013) | (0.015) | (0.020) | (0.013) | |
GDP growth | − 0.039*** | − 0.049** | − 0.054** | − 0.042*** | Yes |
| (0.014) | (0.019) | (0.027) | (0.015) | |
Unemployment | 0.040** | 0.048*** | 0.050** | 0.042** | Yes |
| (0.016) | (0.018) | (0.023) | (0.017) | |
Terms of trade risk | − 0.125*** | − 0.125*** | − 0.129*** | − 0.132*** | Yes |
| (0.020) | (0.038) | (0.045) | (0.020) | |
Democracy | − 0.047* | − 0.048* | − 0.044 | − 0.054** | No |
| (0.025) | (0.026) | (0.038) | (0.027) | |
Observations | 1,182 | 1,182 | 1,182 | 1,182 | |
R squared (adjusted) | 0.258 | 0.357 | 0.267 | 0.253 | |
In fact, no government expenditure variable other than social spending carries statistical significance. The coefficients of the financial globalization and the overall economic globalization variables, however, are both positive and statistically significant. Those findings suggest that, compared to models that control for trade globalization, models that include financial globalization or overall economic globalization report larger (i.e., more positive) relationships between globalization and government spending. The meta-regression results suggest that the choice of the economic globalization variable indeed does matter and the ‘compensation’ effects reported in the literature tend to be stronger when the dimension of financial globalization is considered.
Furthermore, the results in column (1) of Table
3 indicate that, on average, estimates that rely on samples of developing countries find stronger relationships between economic globalization and government spending than those based on samples of advanced countries. That finding supports the argument that the composition of the country group moderates the impact of globalization on government spending in the sense that the ‘compensation’ mechanism, on average, is stronger in developing countries than in advanced countries. Note again, however, that the average ‘compensation’ effect in developing countries predicted by the model estimation in column (1) is rather small. Based on our categorization of reference groups, the model predicts that globalization has a small positive impact on total government spending in developing countries (− 0.05 + 0.12 = 0.07).
Furthermore, the results in model (1) of Table
3 suggest that the choice of functional form influences the reported partial correlations. The omitted reference category is the level level specification, i.e., when the spending and globalization variables both are entered in levels. It can be seen that estimates using functional forms other than level level report larger negative relations between globalization and government spending. Garrett (
2001, p. 21) argues in favor of “an important analytic difference between the extent to which a given country is integrated into international markets in a given period of time (i.e., the level of integration) and the rate at which integration has increased between two periods (i.e., changes in market integration). The changes measure is much closer to conventional understandings of globalization, but the levels measure has been used in the best econometric work on the globalization and government spending relationship around the world.”
Our meta-regression analysis shows that the evidence reported in the existing literature indeed leans meaningfully more towards supporting the ‘compensation’ hypothesis when both the government spending and the globalization variable are specified in levels. In particular, the model in column (1) of Table
3 predicts a medium-sized negative impact of globalization on total government spending (–0.05 + –0.11 = –0.16) when the model is specified in first differences, which rejects the predictions of the ‘compensation’ theory clearly.
Several additional findings from column (1) of Table
3 merit highlighting. First, estimates based on OLS estimators lean more towards supporting the ‘compensation’ hypothesis than other econometric specifications. In contrast, the estimated coefficients for Random Effects and other estimators (e.g., IV, seemingly unrelated regressions) are statistically insignificant. Second, entering country fixed effects moderates the impact of globalization on spending: if country dummies are included, the reported partial correlation tends to lean more towards the side of zero. Third, we find that estimates published in economics journals report stronger (i.e., more positive) relationships between globalization and government spending. In other words, the economics literature reveals a tendency for reporting positive globalization-spending effects. Fourth, scholars who have contributed to the relevant literature previously report partial correlations that tend to be more negative. Fifth, three regressors seem to moderate the impact of globalization on government spending. On the one hand, the two macroeconomic controls (GDP growth and unemployment) return statistically significant coefficients, suggesting that empirical researchers wanting to model the relationship between globalization and government spending should pay particular attention to macroeconomic controls. On the other hand, estimates based on models in which terms-of-trade risk also is controlled for report larger negative effects of globalization on government spending. The coefficients on democracy are significant in three of the four models reported in Table
3. Finally, we can ask whether the evidence for publication selection bias reported in Table
1 is confirmed when we control for additional moderator variables. Indeed, the standard error coefficients in Table
3 have the same signs as the coefficients reported in Table
1; the size of those coefficients is similar. That finding can be interpreted in the sense that the multivariate meta-regression analysis confirms that some publication selectivity bias exists in the globalization-spending literature, but the magnitude of the bias is relatively small (Doucouliagos and Stanley
2013, p. 320). Finally, our meta-regression results do not find substantial evidence for time-varying effects in the globalization-spending relationship: the general-to-specific approach eliminates the continuous variable capturing the mean year of the data’s underlying structure because the coefficient on that variable is estimated to have very low precision.
12
In columns (2)–(4) of Table
3, we test the robustness of the meta-regression results reported in column (1). We do so both by applying different estimators and by transforming the dependent variable. More specifically, the Random Effects model in column (2) introduces an additional between-study variance term to account for differences in the globalization-spending estimates that go beyond sampling error and differences captured by the moderator variables (e.g., Schmidt and Hunter
2014). The robust regression estimator applied in column (3) downweighs observations with larger absolute residuals and, hence, is less fragile to the influence of outlier observations. Finally, column (4) uses Fisher’s
z-transformed partial correlation coefficients to account for the distribution of the partial correlations potentially not being normal when their values are close to − 1 and + 1. The results reported in columns (2)–(4) show that our baseline results largely remain unaffected when those robustness checks are introduced. The sizes of some of the coefficients and their standard errors are subject to variation. However, two cases arise in which a variable whose coefficient was estimated to be significant in column (1) turns out not to be significant in all three additional models: the coefficients on both overall economic globalization and democracy lose their significance when we apply robust regression methods. Besides that, the results in columns (2)–(4) of Table 3 indicate robustness.