The impact of federal deposit insurance on savings and loan failures: reply.

I. Introduction

My recent study [2] finds, within a variety of models based extensively on Barth [1], that raising the ceiling level of federal deposit insurance acts to raise the savings and loan (S&L) failure rate. This finding (which is based on OLS estimates using the Cochrane-Orcutt correction for first-order autocorrelation) is obtained whether the ceiling level on federal deposit insurance is expressed in current or constant dollars.

It is alleged by Saltz [4] that there is an econometric flaw in my analysis. In particular, it has been suggested that there exists a simultaneity bias in my basic model involving the average S&L cost of funds, which is a statistically significant variable in my study [2]. Further, it is argued that, after allowing for this problem (alleged problem), the conclusion in my empirical study that federal deposit insurance affects S&L failures is invalidated.

This study examines the simultaneity issue within the context of my model. First, I describe the original basic model. Then, I expressly address the possibility of simultaneity bias involving explanatory variables in the model. It is shown that (1) there is no simultaneity problem involving the S&L cost of funds variable, (2) there is evidence of a possible simultaneity involving the S&L mortgage portfolio yield, but (3) once the latter is appropriately allowed for, the conclusions in my original study [2] remain unchanged. The comment [4] is useful, nevertheless, for its pointing out the need to allow formally for the possibility of simultaneity.

II. The Model

The most basic model estimated in Cebula [2] is the following reduced-form equation:

[PSL.sub.t] = a + b [RINS.sub.t-2] + c [VOL.sub.t] + d [RP.sub.t-2] + e [COST.sub.t-2] + f [MORT.sub.t] + g [CAP.sub.t-2] + u (1)

where:

[PSL.sub.t] = the percentage of federally insured S&Ls that failed in year t;

a = constant;

[RINS.sub.t-2] = the FSLIC insurance ceiling per account on deposits at S&Ls in year t - 2, expressed in thousands of 1982 dollars;

[VOL.sub.t] = the variance in year t of the yield on new home mortgages (FHLBB series);

[RP.sub.t-2] = the average price per barrel of imported crude oil in year t - 2, expressed in 1982 dollars;

[COST.sub.t-2] = the average cost of funds to S&Ls in year t - 2, expressed as a percent per annum;

[MORT.sub.t] = the average S&L mortgage portfolio yield in year t, expressed as a percent per annum;

[CAP.sub.t-2] = the average required S&L capital-to-asset ratio in year t - 2, expressed as a percent;

u = stochastic error term.

This system was estimated for the period 1965-1989.

Within this model, it has been alleged that there is a simultaneous-equation bias involving the S&L cost of funds, [COST.sub.t-2]. However, it seems almost implausible that the S&L failure rate, [PSL.sub.t], and the cost of funds variable, [COST.sub.t-2], could manifest a two-way causality. This is because the cost of funds variable is lagged two years behind the S&L failure rate. While it is plausible that the cost of funds two years earlier could affect the S&L failure rate, it seems implausible that the S&L failure rate could affect an event that occurred two years earlier! Alternatively expressed, how could the two-year lagged [COST.sub.t-2] variable be correlated with the error term? Answer: it realistically cannot. Interestingly, if the S&L cost of funds variable were altogether unlagged, i.e., if it were simply [COST.sub.t] rather than [COST.sub.t-2], then the Hausman [3] specification test indicates that there would have been a simultaneity problem!

The issue of simultaneity nevertheless does deserve attention. The Hausman [3] specification test reveals evidence that there is a possible simultaneity between [PSL.sub.t] and [MORT.sub.t], which are contemporaneous. To allow for this issue, equation (1) is estimated using an instrumental variables (IV) technique, with the instrument being the three-year lag of the federal funds rate, [FF.sub.t-3]. The choice of instrument is based on the finding that [FF.sub.t-3] systematically explains [MORT.sub.t], whereas [FF.sub.t-3] is not correlated with the error term.

Estimating equation (1) thusly by IV, using the Cochrane-Orcutt procedure to correct for first-order autocorrelation and the White [5] procedure to correct for heteroskedasticity, yields:

[Mathematical Expression Omitted]

where terms in parentheses are t-values.

In equation (2), all of the estimated coefficients exhibit their expected signs and are statistically significant at the four percent level or beyond. The coefficient on the federal deposit insurance variable is positive and significant at nearly the one percent level.

The conclusions in the original study [2] are further supported by numerous alternative IV estimates. For example, dropping variable [VOL.sub.t] from the system while adding (as in Cebula [2]) a dummy (binary) variable to the system, REC, to represent the 1981-1982 recession, and a second dummy (binary) variable, TXR, to represent those full years when the Tax Reform Act of 1986 was in effect during the sample period, provides a model that also generates robust results. This IV estimate is given by

[Mathematical Expression Omitted].

As shown in equation (3), the estimated coefficients are all significant at beyond the three percent level. Of greatest relevance is the coefficient on the federal deposit insurance variable, which is positive and significant at beyond the one percent level. This kind of outcome for the deposit insurance variable was obtained in a wide variety of IV and OLS estimations.(1)

III. Conclusion

It has been alleged that there is a simultaneity problem involving the cost of funds ([COST.sub.t-2]) in my recent study [2]. Although this alleged correlation between [COST.sub.t-2] and the error term in the system can be readily dismissed, there is modest evidence of a possible simultaneity involving the S&L mortgage portfolio yield, [MORT.sub.t]. I have allowed expressly for such a possibility by using IV estimates. Such IV estimates reveal that there is a strong and positive relationship between the S&L failure rate and the ceiling real level of federal deposit insurance. These findings are entirely consistent with those in my original study [2]. Nevertheless, the comment by Saltz [4], a former colleague, was useful in refining the results that were generated.

Richard J. Cebula Georgia Institute of Technology Atlanta, Georgia

1. These results will be provided upon written request.

References

1. Barth, James R. The Great Savings and Loan Debacle. Washington, D.C.: The AEI Press, 1991.

2. Cebula, Richard J., "The Impact of Federal Deposit Insurance on Savings and Loan Failures." Southern Economic Journal, July 1993, 620-28.

3. Hausman, Jerry A., "Specification Tests in Econometrics." Econometrica, October 1978, 1125-51.

4. Saltz, Ira S., "Impact of Federal Deposit Insurance on Savings and Loan Failures: Comment." Southern Economic Journal, July 1995.

5. White, Halbert, "A Heteroskedasticity-Consistent Covariance Matrix Estimator and a Direct Test for Heteroskedasticity." Econometrica, May 1980, 817-38.