Credit Card Borrowing, Delinquency, and Personal Bankruptcy - Statistical Data Included

New England Economic Review, July-August, 2000 by Joanna Stavins

Individual Regression Results

The bankruptcy patterns in Table 1 and regional results in Table 2 are based on aggregated data and do not control for other factors. This section shows the results of regression estimation to test the effect of credit card borrowing on the probability of having delinquent loans and on the probability of having filed for bankruptcy. We use logit regressions to estimate the following equation:

delinquent = [[beta].sub.0] age [[beta].sub.1] income [[beta].sub.2] networth [[beta].sub.3] unemployed [[beta].sub.4] homeowner [[beta].sub.5] family [[beta].sub.6] married [[beta].sub.7] education [[beta].sub.8] health [[beta].sub.9] cards [[beta].sub.10] cardbal [[beta].sub.11] debtinc [[beta].sub.12] bankrupt [beta] census [omega] (1)

where:

delinquent equals 1 if the respondent was behind on payments by two months or more; age is the respondent's age; income is the respondent's annual household income; networth is the respondent's household net worth; unemployed equals 1 if the respondent was

unemployed at any time during the previous 12 months;

homeowner equals 1 if the respondent and family own their house or farm;

family is the number of people in the respondent's household;

married equals 1 if the respondent is married;

education is the highest grade of school or college completed by the respondent;

health equals 1 if the respondent's family is covered by health insurance;

cards is the number of credit cards owned by the respondent;

cardbal is the total balance still owed after last payment on credit cards;

debtinc is the ratio of total debt to annual income;

bankrupt equals 1 if the respondent had ever filed for bankruptcy;

census is a set of dummy variables denoting each of the nine Census divisions but one;

[beta] are coefficients to be estimated; and [omega] is a random error term.

As mentioned earlier, we apply weights to compensate for unequal probabilities of selection of households. The weights, provided in the SCF data, are equal to the inverse probability of observing each case, based on a comparison of each surveyed household to aggregate control totals estimated from the Current Population Survey. (5)

The regression results are reported in the first column of Table 3. Because we use logit estimation, the estimated coefficients are interpreted according to the formula:

[increment of]log P/1 - P = [beta] [increment of]x

where P is the probability of default on loans, [beta] is the estimated coeficient, and x is the variable whose effect we are trying to evaluate. Rewriting the above equation, the effect of an increase in x by 1 is:

[increment of] [approximately equal to] [beta] [P(1-P)].

Having more credit cards reduces the likelihood of delinquency, while higher unpaid balances on credit cards increases the probability of being behind on payments. A respondent holding one additional credit card ([DELTA]x = 1) has a 0.44 percent lower probability of being delinquent on payments (evaluated at the mean probability P = .06). If the respondent's unpaid credit card balance increases by $1,000, his probability of delinquency rises by 0.23 percent. Therefore, even doubling the average unpaid credit card balance of $1,817 would lead only to a 0.42 percentage point increase in the likelihood of default, not an economically significant amount, despite the statistical significance of the variable.