Linear panel model

5.2.1 Model specification

Considering the availability of loan data across both time and recipient countries, the application of panel data estimators is the logical first step. The goal of such an analysis is to predict the expected loan amounts that each type of coalition would obtain from each of the four types of creditors.

I therefore estimate a random effects model of the following form

y_it = α + γy_it−1+ βx_it+ ψz_it+ _it (5.1) where t covers the time period from 2004 through 2010. The choice of t is motivated by theory. I argue that governments choose between four types of creditor based on the preferences of the dominant societal coalition. However, BRICs are emerging creditors that have not been active for a long period of time. To apply my theoretical framework onto a time period where BRICs were still inactive would result in biased estimates. For this

reason, I identified the first instance of a BRIC lending more than $1 billion US dollar, which was a Chinese loan to Angola in 2004. I therefore assume that from 2004 onward BRIC creditors were sufficiently active to be considered by recipient governments as a viable alternative. With the data obtained from the World Bank ending in 2010, the panel is characterized by t = 7 and N = 129.

A common concern with panel data is the possible presence of serial correlation across time. In short, it might be the case that existing debt would have an effect on a cur-rent year’s loans. However, the dependent variable measures the new loans obtained in a given year, not the overall debt stock already obtained from a particular creditor. The dependent variable therefore captures the first differences rather than the level of debt, which should moderate concerns about serial correlation. Nevertheless, having established a good working relationship with a particular creditor might provide governments with the incentive to favor that particular creditor over time. To account for this possibility of serial correlation, I follow the common practice of including a lagged dependent variable yit−1 to further alleviate such concerns.

In addition to concerns about autocorrelation, my estimates must also address the problem of selection effects. Przeworski and Vreeland (2000) and Vreeland (2003a) show with respect to the IMF that the selection into IMF programs may be non-random. In other words, the same variables that explain loan size may also explain the likelihood of obtaining a loan in the first place. Such reasoning applies not only to the IMF. It is reasonable to assume that the likelihood of obtaining a BRIC loan also affects the size of the BRIC loan obtained. To address this selection problem I follow Copelovitch (2010b) and employ propensity score matching – albeit modified to take into account the possibility of borrowing from four separate creditors as opposed to Copelovitch’s specification that only considers one creditor. I therefore create propensity scores z₁, z₂ and z₃ that capture the predicted probability of obtaining a BRIC loan, DAC loan or private loan, respectively. The propensity score for IFIs is excluded to avoid the dummy variable trap. Essentially, these

propensity scores measure the probability of obtaining one of the respective loans given the covariates of the observation (i.e. each country-year). The inclusion of these propensity scores implements the idea of matching each observation with a ‘control’ observation for which are values of the explanatory variables are as close to identical as possible.

Including z1−3 in the panel regression minimizes selection bias and enables the use of standard, parametric regression techniques (Ho et al., 2007). As such, for example, it allows me to use cluster-robust Huber/White/sandwich estimator that results in an error term  that is identically and independently distributed over the clusters, but correlated within each cluster. Without accounting for this correlation the OLS point estimates would be consistent, while the standard errors of the variance-covariance estimation would not.

5.2.2 Results

The results of the panel regressions are displayed in Table 5.1. As noted above, I estimate Equation 5.1 in four separate specifications. In model 1, the dependent variable is the logged loan amount obtained from BRICs. Consequently, the lagged dependent variable refers to the lagged log loan amount acquired from BRICs. To account for the cost of loans, I include the corresponding grant element for the BRIC loans obtained. In contrast, model 2 uses the logged loan amount from DACs along with the corresponding lagged dependent variable and grant element. The dependent variable of model 3 are logged IFI loans, while model 4 is estimated with the logged loan amount obtained from private creditors. However, other than the differences in dependent variable, lagged dependent variable and corresponding grant element, the model specification is identical across the four regressions.

The results indicate that Capital coalitions borrow less from BRICs than Corporatist coalitions, while the opposite is the case with private creditors. The same conclusion can be drawn when comparing a Consumer coalition to a Corporatist coalition, as both the magnitude of the coefficients and their statistical significance are remarkably similar.

While the interpretation of marginal effects provides some insights, my theory suggests that the absolute amount of loans obtained from each creditor should vary across coalitions.

The quantity of interest is therefore the predicted loan amount that each coalition obtains from the four creditors. I therefore estimate the predicted loan amount by coalition. To facilitate the interpretation of the results I use simulation techniques to visualize the sub-stantive results of interest. This information is displayed in Figure 5.1. The location of each curve’s highest point indicates the point prediction of the loan amount that the respective coalition is estimated to obtain on average. The shape of the curve provides information about the certainty of the estimate, with wider curves indicating larger uncertainty.

With respect to the top panel, Corporatist coalitions are estimated to borrow more heavily from BRICs than either Capital or Consumer coalitions. Translating the loga-rithmic scale into absolute amounts reveals that Corporatist coalitions are estimated to borrow an average of 268 million US$ per year, while Capital coalitions are expected to obtain about 46 million US$ and Consumer coalition around 54 million US$. I test whether these differences are statistically significant by conducting a series of difference-in-means tests¹ . The difference between the Corporatist and the Capital coalition yields a χ² of 4.85 which implies a p-value of 0.028. The probability that a Capital coalition borrows a larger amount from BRICs than a Corporatist coalition is therefore less than 3%. The t-test for the difference between Corporatist and Consumer coalition is also statistically significant (χ² = 4.73, p = 0.030). Lastly, the difference between BRIC loans obtained by Capital and Consumer coalition is not statistically significant (χ² = 0.07. p = 0.788). In short, my theoretical expectations with respect to BRIC loans are confirmed by the data.

DAC loans, however, do not appear to be determined by social coalitions. The second panel indicates that Corporatist coalitions, on average, obtain lower loan amounts from DACs than either Capital or Consumer coalitions. However, these differences are not

1 I calculate two-tailed difference-in-means tests. While I could achieve statistical significance for additional estimates reported in this section if I were to use one-tailed t-tests, I deem their use inappropriate in this case. After all, I need to account for the possibility that the BRIC loan amount by a Capital coalition is either lower or higher than that of the Corporatist coalition.

2 4 6 8 Logged loan amount

Corporatist Capital Consumer

BRIC loans

2 3 4 5 6

Logged loan amount

Corporatist Capital Consumer

DAC loans

4.5 5 5.5 6 6.5 7

Logged loan amount

Corporatist Capital Consumer

IFI loans

4 5 6 7 8

Logged loan amount

Corporatist Capital Consumer

Private loans

Figure 5.1: Predicted loan amounts by creditor and coalition. Estimates based on Panel OLS Model.

statistically significant, as the difference-in-means test for the Corporatist-Capital pair results in a χ² of 2.04 (p = 0.154) and that for the difference between Corporatist and Consumer coalition in a χ² of 0.74 (p = 0.389). The predicted DAC loan amounts for

Consumer and Capital coalition do not differ either (χ² = 0.39, p = 0.534).

A similar picture is revealed in the degree to which governments obtain loans from IFIs. Neither of the three coalitions appears to obtain significantly larger amounts from IFIs than the respective other coalitions. Difference-in-means tests confirm the estimates displayed in the third panel of Figure 5.1. The difference between Corporatist and Capital coalition is insignificant (χ² = 0.16, p = 0.688) as is the difference between the Corporatist and Consumer coalition (χ² = 0.02, p = 0.897) and the Capital and Consumer coalition (χ² = 0.40, p = 0.529).

In contrast, the predicted loan amount that the three coalitions are expected to ob-tain from private creditors differ markedly. The bottom panel of Figure 5.1 shows that governments of countries characterized by a Corporatist coalition utilize private creditors less extensively than those characterized by both Capital and Consumer coalition. On average, the model predicts Capital coalitions to obtain 433 million US$ while Consumer coalitions would borrow 521 million US$ per year. In contrast, Corporatist coalitions are predicted to borrow only 216 million US$ per year. The p-value for the difference between Corporatist and Capital coalition is 0.046 (χ² = 3.97), and that for the difference between Corporatist and Consumer coalition is 0.005 (χ² = 7.85). In contrast, the difference be-tween Capital and Consumer coalition is not statistically significant (χ² = 0.20, p = 0.652).

My theoretical predictions are therefore strongly supported, as Corporatist coalitions ob-tain significantly fewer resources from private creditors than either Capital or Consumer coalitions.

In sum, the panel estimates presented in this section provide support for the hypotheses derived in Section 2.5. Even when controlling for a host of methodological issues (selection effect, temporal autocorrelation) and alternative explanations, social coalitions appear to determine the choice of creditor. There are significant differences between the Corporatist coalition, which excludes Finance, as compared to either the Capital or the Consumer coalition which both include Finance. While the former tend to favor borrowing from

BRICs, the latter prefer to obtain loans from private creditors. In other words, Capital and Consumer coalition borrow about 10 times as much from private creditors than BRICs per year. In contrast, Corporatist coalitions borrow more from BRICs than private creditors.

5.2.3 Appraisal

Linear panel data models are valuable because they allow for estimating the loan amount that each coalition is expected to acquire. However, recall that my theory suggests an interdependency among creditors. Faced with the constraint of a maximum amount of loans that can be obtained, a government’s choice for one creditor is simultaneously a choice against another. However, the approach with panel data requires me to estimate the predicted loan amount obtained separately for each type of creditor. Panel models are therefore inherently unable to account for the interdependencies I derive in Chapter 2.

To adequately test my theory, I therefore need to turn to estimation methods that allow for the borrowing decisions of governments among the different creditors to be estimated jointly.

BRICs DACs IFIs Private

Capital coalition -1.759** 0.687 0.134 0.695**

(0.799) (0.482) (0.333) (0.349)

Consumer coalition -1.598** 0.461 -0.044 0.881***

(0.735) (0.534) (0.341) (0.314)

intransitive situation -0.087 0.560 0.039 0.388

(0.557) (0.484) (0.370) (0.325)

Taiwan recognized -0.856 -0.150 -0.124 -0.456

(2.073) (0.948) (0.405) (0.497)

Total natural resources rents (% of GDP) -0.067 -0.007 -0.026* 0.009

(0.047) (0.025) (0.014) (0.017)

S-score of G5 -3.215 2.084 0.926 2.538***

(2.948) (1.295) (0.844) (0.837)

1.IMF Membership (dropped) (dropped) (dropped) (dropped)

. . . .

Principal arrears, official creditors (US$) -0.000 -0.000 -0.000 -0.000

(0.000) (0.000) (0.000) (0.000)

Interest arrears, official creditors (US$) -0.000 0.000* 0.000 0.000

(0.000) (0.000) (0.000) (0.000)

Principal arrears, private creditors (US$) -0.000 0.000** -0.000 -0.000

(0.000) (0.000) (0.000) (0.000)

Interest arrears, private creditors (US$) -0.000 -0.000 0.000 0.000

(0.000) (0.000) (0.000) (0.000)

GDP per capita -0.535 0.149 -0.006 0.271

(0.755) (0.286) (0.194) (0.306)

Revenue, excluding grants (% of GDP) 0.063 -0.041 0.009 -0.088***

(0.061) (0.035) (0.023) (0.028)

Gov’t consumption expenditure (% of GDP) -0.155* 0.005 -0.013 0.087***

(0.093) (0.047) (0.027) (0.028)

Imports of goods and services (% of GDP) -0.081 -0.010 -0.036*** -0.008

(0.069) (0.021) (0.013) (0.016)

Exports of goods and services (% of GDP) 0.095 0.018 0.005 0.004

(0.061) (0.029) (0.015) (0.014)

External debt stocks (% of GNI) 0.011 -.006 0.003 -.008

(0.024) (0.007) (0.003) (.003)

z1 0.158 -0.236 -0.636 -0.000

(1.526) (1.098) (0.802) (0.926)

z2 1.682 2.241 0.370 1.001

(2.240) (1.841) (1.106) (1.080)

z3 0.546 -0.926 0.330 -1.119**

(0.686) (0.615) (0.257) (0.458)

z4 -4.962e+07 -1.896e+07 443924.560 -4.868e+07

(34619973.900) (24545706.585) (7327629.744) (33640166.470)

Table 5.1: Model results of OLS panel regression.