Forbearance regression without forbearance

In this model, I assume that the regulator does not “forbear” from closing down a bank. But the researcher observes \( X \), which is the regulators signal, \( Z \), plus noise (i.e., due to opacity). I assume the researcher can observe “opacity'' for each firm, which is simply the standard deviation of the noise component.

Set parameters

N <- 1000
mean_Z <- -3
Z_star <- 0.2
sigma_min <- 0.2
sigma_max <- 1.8

Simulate the data

The regulator observes \( Z \), which is the probability of failure, and simply shuts the bank down if \( Z > Z^{*} = 0.2 \). There is no "forbearance” in this model. Some banks that are not shut down, fail anyway. We don't observe the distinction between a regulator-ordered closing and a failure.

# Regulator-observed score
z <- rnorm(N, mean_Z)
Z <- exp(z)/(1 + exp(z))
closed <- Z >= Z_star

# Measurment error in earnings
sigma <- runif(N, min = sigma_min, max = sigma_max)

# Earnings is true signal plus noise
x <- z + rnorm(N, mean = 0, sd = sigma)
X <- exp(x)/(1 + exp(x))

# Banks can be closed or can fail
failed <- rbinom(N, size = 1, prob = Z) | closed
table(failed, closed)
##        closed
## failed  FALSE TRUE
##   FALSE   892    0
##   TRUE     50   58

Empirical analysis

We can measure forbearance as in Gallemore (2013) and then estimate whether this measure is associated with opacity.

# Measure of forbearance
forbear <- ifelse(failed, 0, X)

# Regression of
summary(lm(forbear ~ sigma))
## 
## Call:
## lm(formula = forbear ~ sigma)
## 
## Residuals:
##     Min      1Q  Median      3Q     Max 
## -0.1068 -0.0649 -0.0312  0.0220  0.8132 
## 
## Coefficients:
##             Estimate Std. Error t value Pr(>|t|)    
## (Intercept)  0.03530    0.00848    4.16  3.4e-05 ***
## sigma        0.03981    0.00757    5.26  1.8e-07 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Residual standard error: 0.112 on 998 degrees of freedom
## Multiple R-squared:  0.027,  Adjusted R-squared:  0.026 
## F-statistic: 27.6 on 1 and 998 DF,  p-value: 1.78e-07