«Research Division Federal Reserve Bank of St. Louis Working Paper Series Understanding the Accumulation of Bank and Thrift Reserves during the U.S. ...»
In all other cases, the share held by thrifts was substantially smaller. While there is a chance that all of the CPP injection was channeled into the thrift, we think this is an unlikely event.
a random sample from the population, as would be the case in an experimental setting.
If better-performing banks were awarded funds from the CPP, CPP status becomes endogenous, invalidating the use of simple correlation estimation.
We use PSM techniques to control for endogeneity.34 The basic idea is to construct control and treatment groups, where receiving CPP funds is the treatment. Our goal is to ﬁnd a set of control banks that are a priori equally likely to receive a capital injection as those banks that ultimately did receive one. PSM is then combined with a diﬀerence-in-diﬀerences approach to measure the average divergence in the performance paths between the BHCs in the CPP group and those in the non-CPP group.
We match individual bank identiﬁers to BHC identiﬁers. Information available at the BHC level is assumed to carry over to individual banks within the BHC group.
For example, (i) we collect information on whether BHCs are publicly traded from a publicly available dataset at the Federal Reserve Bank of New York and construct a dummy variable equal to 1 for each bank in the publicly traded BHC, and (ii) we use the BHC identiﬁer to match each bank with our proxy for management quality.35 To formalize the PSM procedure, we deﬁne the cash-to-assets ratio that we would like to evaluate as Y. Let Y 1 and Y 0 denote cash-to-assets ratios for the BHCs in the CPP group and the non-CPP group, respectively. Let CP P be a binary variable indicating whether a BHC received CPP support. The aim of the analysis is to estimate
the following causal eﬀect of CPP funds on the outcome Y :
E Y 1 − Y 0 |CP P = 1 = E Y 1 |CP P = 1 − E Y 0 |CP P = 0, (11) which is the diﬀerence between the dynamic path of the cash-to-assets ratio for BHCs that received CPP funds (ﬁrst term) and the analogous outcome for the same BHCs had they not been granted CPP funds (second term).
The PSM technique is used to approximate the unavailable counterfactual by drawing comparisons conditional on the observables, X (see Dehejia and Wahba 2002 for a discussion). We thus assume that, conditional on the observable characteristics relevant to the CPP decision, the mean of the outcome for the BHCs in the CPP group, had they not been granted CPP funds, should be the same as the mean for those in 34 The high dimensionality of the observable characteristics increases the diﬃculty of ﬁnding exact matches for each BHC in the CPP group. Conditioning on a vector of variables requires a choice regarding which dimensions should be used to match across units or which weighting scheme to apply. Rosenbaum and Rubin (1983) and Dehejia and Wahba (2002) demonstrate that the propensity score provides a natural weighting scheme that yields unbiased estimators of the treatment impact. Thus, conditioning on the propensity score is equivalent to conditioning on all variables in the treatment model, hence reducing the dimensionality issue.
35 The CRSP-FRB Link dataset is available at http://www.newyorkfed.org/research/banking_research/datasets.
the non-CPP group:
that is, the selection bias is removed, conditional on X.
The propensity score is the probability that a BHC receives CPP funds conditional on a set of covariates, denoted as p(X). We deﬁne
To ensure the timing is correct for the pre-TARP and post-TARP quarters outlined above, we use information about the CPP from the U.S. Treasury and media sources.
When the CPP was announced in October 2008, a number of applications were submitted to the U.S. Treasury. However, at this point, and despite various lawsuits under the Freedom of Information Act of 1966, the U.S. Treasury has not disclosed the list and the timing of applications. Thus, we must rely on informal evidence for the application timing and the pool of applicants. We have two pieces of information that can assist us with the timing. First, Treasury oﬃcials revealed that “thousands of applications” for funds were received, but only a few hundred BHCs qualiﬁed for funds through the CPP based on their CAMELS scores.36 Second, the United States Department of Financial Stability (2010) stated that the rate at which applications were submitted declined rapidly in early 2009. The report cites three key reasons for this decline. (i) In February 2009, Congress adopted more restrictive executive compensation requirements for all TARP recipients. (ii) Many banks felt there was a stigma associated with participation in the program. (iii) The impact of the crisis on DIs started to appear less dramatic.
Based on this information, we treat the entire population of banks, with the exception of foreign banks, which were excluded from receiving funds under the program, as the pool of applicants. We also exclude new commercial banks (credit card companies and investment banks) for reasons explained in Section 4. We conjecture that the majority of applications were submitted in the fall of 2008. Based on this assumption, we estimate the probability of receiving CPP funds based on observable characteristics measured at the end of 2008:Q3.
5.4 An empirical model of the capital purchase program
The ﬁrst part of our approaches relies on a reduced-form empirical model of CPP participation. CPP beneﬁciaries diﬀer from non-CPP banks along many dimensions (Contessi and Francis, 2011) and, in fact, our data reveal substantial dissimilarities in terms of capital ratios, size, and loan composition. We observe banks becoming CPP beneﬁciaries along with a matrix of observable indicators. A natural approach to model this event is to use a probit model for the probability a bank received CPP funds based on a set of observable characteristics. We assume that local economic 36 A bank’s CAMELS score is a conﬁdential regulatory bank rating metric based on six factors: C, capital adequacy;
A, asset quality; M, management quality; E, earnings; L, liquidity; S, sensitivity to Market Risk.
conditions, along with key bank-level characteristics, aﬀect the probability of applying for and being granted CPP funds. The explanatory variables are measured at the end 2008:Q3, as the application process opened in 2008:Q4, and according to U.S. Treasury documents, most applications were received by the beginning of 2009.37 The results for the probit estimates are listed in Table 11. We estimate the probit model using three groups of regressors: a set of standard ﬁnancial indicators for banks, geographic variables meant to capture changes in demand, and other variables likely to aﬀect selection into the program. We use the following speciﬁc variables in our
speciﬁcation of the probit model:
• Capital adequacy: We use three measures of capital adequacy: The ratio of total equity to total assets, the Tier 1 capital ratio in levels, and squared.
• Asset size and composition: We use the logarithm of total assets; commercial and industrial loans as a share of total assets; cash and reserves as a share of total assets; and all “other securities” (quarterly average) as a share of total assets.
• Bad loans: We use loan loss reserves, loan losses provisions as a share of earning assets, loan losses as a share of equity, and net loan charge-oﬀs as a share of total loans.38
• Composition of liabilities: We use deposits as a share of total assets and borrowed funds with maturities longer than one year as a share of total assets.
• Other variables: We include a measure of leverage (the ratio between total loans and deposits), a dummy variables for whether a BHC is publicly traded or a top 40 BHC ranked by assets, as well as a dummy variable equal to 1 if any of the managers of a BHC is also on a regional Federal Reserve Bank Board in the fall of 2008. Hypothetically, a BHC may be more likely to receive CPP funds if its political connection is stronger. Duchin and Sosyura (2012), for example, argue that political connections — as measured by contributions to House members on ﬁnance committees and representation at the Federal Reserve as Board members — have signiﬁcant positive marginal eﬀects on the probability of a bank being granted CPP (TARP) funds. Alternatively, a BHC could have been excluded from CPP funding because a bank manager did not apply for them due to his or 37 An alternative route is to estimate a probit model based on observables measured at the end of the quarter in which CPP funding was granted; however, anecdotal evidence suggests that a large number of applications were submitted in the ﬁrst few months of the program.
38 The sum of net loan charge-oﬀs and the loan loss provision is deﬁned as gross charge-oﬀs.
her strong anti-government intervention beliefs (CNNMoney.com, 2010). These types of unobservable determinants of CPP funding are likely to be time invariant and can be eliminated by the diﬀerence-in-diﬀerences approach.39
• Management quality: We construct a proxy of management quality using the number of corrective actions taken against bank management by its regulator in the 2006-09 period.40
• Earnings: We use the the ratio of pretax net income and total earning assets (the sum of total loans and total securities) and the ratio of net income to operating income to capture earnings.
For the probit estimation, we perform both forward and backward stepwise procedures and select the model speciﬁcation with the highest pseudo-R-squared. All coeﬃcients reported in Table 11 are signiﬁcant at the 5 percent level or better. We compute the predicted probability (i.e., propensity score), based on the parameter estimates in the selected model, and match the BHCs in the TARP group with those in the non-TARP group using one-to-one nearest neighbor matching on the propensity score. The average TARP eﬀect on the TARP group is then calculated using a diﬀerence-in-diﬀerences approach described in Section 5.2.
First, we consider the BHC characteristics that increased the probability of receiving TARP-CPP funds. We ﬁnd that larger banks (higher log of total assets) and banks with more commercial lending were more likely to receive TARP funds. Banks with a higher loan loss provision—indicating signiﬁcant default risk in their loan portfolio— a larger percentage of loans in non-current status, higher loan loss ratios, and loan charge-oﬀs as a percent of total loans were less likely to receive TARP funds. However, banks with larger loan loss reserves were more likely to receive TARP funds. A loan loss provision is taken when the risk of loan default is higher. Loan loss reserves are based on a risk assessment of the loan portfolio but could be a measure of prudence.
We ﬁnd evidence that publicly traded and larger (e.g., BHCs in the top 40 BHCs by asset size) banks were more likely to receive funds. BHCs with higher real estate 39 We include this variable because a banker who is a member of a regional Board may be more likely to know about TARP, perhaps because of better information on the program.
40 As in Duchin and Sosyura (2012), who generously provided the raw data, we have a total of 1,681 orders issued to 961 commercial banks. Enforcement actions include prohibitions from further participation in banking activities, orders to cease and desist, and orders to pay civil monetary penalties.
exposure or in states with more unemployment insurance claims were also more likely to receive TARP funds.
The most signiﬁcant variables for understanding the allocation of TARP funds relate to capital adequacy. We have no clear prior about the sign of the coeﬃcient on the capital adequacy variables. A positive coeﬃcient suggests that the decision to grant CPP funds was geared toward reinforcing the capital position of healthy banks. A negative coeﬃcient, on the other hand, suggests that funds predominantly supported relatively weaker banks. As the relationship may be nonlinear, we introduced a quadratic term.
We ﬁnd that the coeﬃcient on Tier 1 capital is negative and the coeﬃcient on the quadratic term is positive, indicative of a convex relationship between receipt of CPP funds and capital adequacy. This result could be interpreted as supportive of the spirit of the CPP legislation. Banks with weaker capitalization, but still above a threshold Tier 1 capitalization, separating healthy from unhealthy (or likely to fail) institutions, were more likely to apply for and be granted a capital injection.
Assuming that our propensity scoring exercise created a well-matched set of treated and control BHCs by removing observable diﬀerences, we can now use diﬀerence-indiﬀerences estimation to consider how cash-to-assets ratios of the treated and control groups diﬀered, removing unobservable ﬁxed diﬀerences between the two groups.
First, we ﬁnd our matching procedure performs well as our matched pairs of BHCs are only 0.07 percentage points apart in terms of the propensity score. Moreover, the cash-to-assets ratio for the TARP group in the pre-TARP and TARP quarters is larger than for the non-TARP groups (ﬁrst two rows and columns of Table 12).
Second, we report the average treatment eﬀects (average treatment on the treated) in two diﬀerent ways as described in Section 5.2. While AT T 1, which compares cashto-assets ratios in the target quarter with those in the previous quarter, does not produce a general pattern, AT T 2, which compares cash-to-assets ratios in the target quarter with those in the pre-TARP quarter, does exhibit a notable pattern. The fact that cash-to-assets ratios experienced a rising trend during this period is a possible explanation for the diﬀerent patterns between the two measures.