Reestimating the Impact of Private Credit

In reexamining the Cournede-Denk results, a crucial first question is thus whether to include country fixed effects. Classic early studies of cross­country growth typically did not include country fixed effects.

The most important driver of growth in cross-country tests has tra­ditionally been the convergence factor captured by the logarithm of ppp income per capita. This variable becomes irrelevant when country fixed effects are applied, and Cournede and Denk instead apply the (lagged) log­arithm of each country's own national real per capita income. Important variation is sacrificed as a consequence.^[227]

Whether to use country fixed effects is related to whether growth conver­gence is “absolute” or “conditional.” In absolute convergence, poorer coun­tries would tend to grow faster than richer countries, and by implication all countries would eventually tend to converge to the same real per capita income. The early cross-country growth literature applying ppp income estimates when they became available instead confronted the paradox that for 1960-90, poor countries were not growing more rapidly than rich countries, and indeed were growing more slowly (Barro 1996, Sala-i-Martin 1996).

It turns out, however, that beginning in the 1990s growth per capita in emerging-market and developing countries rose increasingly above that in advanced countries, such that the evidence shifted toward absolute rather than conditional convergence. Thus, per capita growth was only 1.9 percent in developing countries versus 2.4 percent in advanced countries in 1981-90 (which included Latin America's “lost decade” from the debt crisis). In suc­cessive decades, however, this comparison swung to 2.2 versus 2.0 percent, respectively, in 1989-98; 4.1 versus 2.1 percent in 1997-2006; and, in the Great Recession and its aftermath, 4.1 versus 0.27 percent in 2008-14 (IMF 1999, 169; 2015b, 170). With less need to seek conditional rather than abso­lute convergence, the case for sacrificing variation in order to capture unob­served country-specific influences has presumably declined.

In the specific case of the Cournede-Denk estimates for the OECD, moreover, the countries included in the sample are much more homoge­neous than in most cross-country tests, providing an important additional reason for excluding country fixed effects. An additional consideration is that in cross-country growth equations, country fixed effects are found in Monte Carlo experiments to exaggerate the speed of conditional conver­gence and thereby reduce the magnitudes and statistical significance of co­efficients of explanatory variables (Hauk and Wacziarg 2009, 105).

In at least one regard, a plausible case can be made for including rather than excluding country fixed effects: The finance variable does move sub­stantially, providing some potential basis for obtaining discrimination using within-country-only variance. Nonetheless, because the key influence of cross-country difference in real per capita income is thrown out when country fixed effects are used, and because this influence is surely the most fundamental in the cross-country growth literature, this reason alone is sufficient to prefer tests omitting country fixed effects over tests including them. The first major decision in specification of the new tests conducted here, then, is to exclude country fixed effects.

Another decision concerns the time period. The tests here end in 2007 growth, to avoid distortions from the Great Recession. Another key ques­tion is whether to use annual data or period averages. The main estimates here use five-year averages (the same approach adopted in both Cecchetti and Kharroubi 2012 and Arcand, Berkes, and Panizza 2012). The use of annual data instead will tend to introduce a bias toward a negative influ­ence of finance on growth from cyclical patterns. Namely, in a recession, the magnitude of debt in the finance variable numerator will tend to rise from accumulation of unpaid balances, whereas the GDP denominator will tend to decline because of lower output.

Another question is whether to include investment as an explanatory variable. Because investment directly drives growth, presumably the most interesting question is whether greater financial depth benefits growth in­directly through facilitating higher investment. The tests here omit invest­ment because otherwise there will be a tendency to understate the influence of finance working through facilitation of investment.

Table 6.2 reports the results of applying tests of per capita GDP growth on the same variable for finance as used by Cournede and Denk: the level of credit to the private sector as a percent of GDP, as well as the same variable

Table 6.2 Cline estimates for four specifications of regressions for per capita GDP growth^a

a.

b. Dependent variable. I ndependent variable lagged two years.

c. Year fixed effects.

Source: Author's calculations.

they use for human capital (years of schooling).^[228] The logarithm of ppp per capita income is taken from the Penn World Table (Feenstra, Inklaar, and Timmer 2015a, 2015b).^[229]8

The two preferred tests in table 6.2 are shown in columns A and B. In these tests, the independent variables are five-year (nonoverlapping) aver­ages beginning in 1961 and the dependent variable (per capita growth) is the corresponding average for the period two years later (e.g., growth in 2003-07 regressed on 2001-05 independent variables). In column A, the finance variable turns out to have a positive coefficient, albeit not a signifi­cant one. As expected, the coefficient for the logarithm of lagged ppp per capita income is negative and highly significant.^[230] The number of school years also has the correct sign and is highly significant.^[231]

For purposes of the “too much finance” debate, the key finding is that the coefficient on the financial depth variable is positive, rather than nega­tive. Column B reports the same test but with this variable stated as the natural logarithm of private credit as a percent of GDP. This time the /-sta­tistic on the variable is considerably higher (but still below the level needed for even a 10 percent level of significance). The estimated coefficient is again positive. The size of the coefficient indicates that as credit to the private sector doubles from 50 to 100 percent, the per capita growth rate would be expected to increase by 0.19 percentage point.^[232] It would take another dou­bling, to 200 percent, to boost the growth rate by another 0.19 percentage point, showing diminishing returns to additional finance, but not negative returns. The higher /-statistic and (albeit only slightly) R² than in the linear case (column A) support this common-sense finding of diminishing returns.

In contrast, columns C and D represent “misleading” results, either because they include country fixed effects or because they use annual rather than longer-period average data. In column C, inclusion of country fixed effects turns the coefficient of the finance variable negative using the five- year average data.^[233] In column D, the use of annual rather than five-year data also turns the coefficient on financial depth negative, even without country fixed effects.

It is useful to apply tests that are more in keeping with what the authors truly think: tests with specifications that allow the influence of finance on growth to diverge between countries at lower levels of finance and those at higher levels, rather than the one-size-fits-all negative linear coefficient. For this purpose, the most reasonable dividing point is a ratio of private credit to GDP of 60 percent or less. This is the threshold at which their supple­mentary tests show that inclusion of observations with greater financial depth begins to turn the coefficient on finance from positive to negative (Cournede and Denk 2015, 29).^[234] A first approach for this purpose is simply

Table 6.3 Growth results for low versus high financial depth^a

a.

b. ln (pc): logarithm of credit to private sector as percent of GDP.

Source: Author's calculations.

to estimate the regression equation for two separate samples: one including only country-periods with private credit less than 60 percent of GDP, and the other including all others. An alternative means of conducting this test is to include all observations, but to add a dummy variable for those cases with the finance variable less than 60 percent of GDP and allow it to in­teract with the finance variable.^[235]

As shown in table 6.3, when these tests are conducted, and using the preferred model specification of column B in table 6.2, the results are much closer to what one would expect. The influence of additional finance is posi­tive and statistically significant in the below-60 group (column A). The in­fluence turns negative but is statistically insignificant in the above-60 group (column B). In the combined test, the influence of the logarithm of finance is again negative but statistically insignificant. However, the interaction term shows that this coefficient turns sizable and positive if the country­period has private credit below 60 percent of GDP, although the coefficient is not significant at the 10 percent level.^[236]

When the exact variables applied by Cournede and Denk (as in table 6.1) are applied to the above 60 percent subsample using five-year averages, despite the problems with these specifications as just discussed, the nega­tive linear coefficient of growth on private credit as a percent of GDP is confirmed and is highly significant. The log of lagged (national) GDP per capita has the right sign and is significant. But investment is not significant; schooling has the wrong sign and is significant; and population growth has the wrong sign. For the below-60 observations, however, private credit still has a negative sign, although the coefficient is small and insignificant. Schooling, population growth, and lagged national GDP per capita all have the wrong signs. Only investment has the right sign.^[237]

Two broad patterns can be seen in these tests. First, the results for the below-60 group are extremely weak, again strongly suggesting that the use of country fixed effects and the use of lagged national rather than ppp GDP per capita income rob the explanatory role of just about all influences (in­cluding credit) except investment. Second, the poor results on the other variables raise the question of why one should trust the significant negative effect on credit in the above-60 group.

In summary, the main results reported in Cournede and Denk (2015) do not provide a solid basis for concluding that “at current levels” OECD finance is so excessive that it depresses growth, because as the authors rec­ognize it does not apply to lower levels of finance and therefore its test sta­tistics are not reliable. Separate tests that distinguish between low and high levels of finance find a significant positive effect at low levels but do not find a statistically significant effect at high levels. Moreover, even if the full set of data is examined without distinguishing between low and high finance observations, the sign of the finance variable switches back to positive rather than being negative if an arguably more appropriate specification is applied.^[238] This specification involves (a) incorporating the traditional work­horse variable for cross-country growth regressions: logarithm of lagged ppp per capita income; (b) excluding rather than including country fixed effects; and (c) applying five-year averages rather than annual data. Their finding that more finance depresses growth is thus not robust to these three relatively basic alternatives (or, I would say, improvements).

Finally, it should be emphasized that even if a negative linear effect were robust to estimation just for the above-60 percent of GDP private credit group, applying the logarithm of lagged ppp per capita income and omit­ting country fixed effects, there would still be the problem of likely reverse causation. The authors do attempt to address causality by constructing an instrumental variable for private credit, based on changes in national finan­cial regulatory requirements. However, not only do they again apply the analysis to the full sample rather than just observations above 60 percent of GDP in private credit, but in addition they do not include lagged per capita income in their growth equations at all (neither ppp nor national), a strange test considering the primacy of this convergence variable in the cross-country growth literature.