Credit Riskiness of Low-Income Consumers

Donald E. Sexton, Jr., Columbia University
[ to cite ]:
Donald E. Sexton, Jr. (1975) ,"Credit Riskiness of Low-Income Consumers", in NA - Advances in Consumer Research Volume 02, eds. Mary Jane Schlinger, Ann Abor, MI : Association for Consumer Research, Pages: 197-202.

Advances in Consumer Research Volume 2, 1975      Pages 197-202


Donald E. Sexton, Jr., Columbia University

[This research was supported by the Graduate School of Business, Columbia University. The author wishes to thank Wallis Hocker, Ed Petti, and Ravinder Sharma of J.C. Penney, Morton Schwartz of Montgomery-Ward, Linden Wheeler and James Smith of Sears-Roebuck, and Robert Shay of Columbia University for their interest, cooperation and advice.]

[Donald E. Sexton, Jr. is Associate Professor, Graduate School of Business, Columbia University.]

Low-income consumers frequently experience difficulty in obtaining credit. In part, this situation is due to credit-granting procedures based on the general population. This investigation was an attempt to find variables that would identify those low-income consumers who are good credit risks. A sample of 4,119 credit histories from three national retailers was examined.

At present, the many low-income families who are unable to obtain credit must pay high prices to inner city merchants -- prices that include sizable charges for credit. If more reputable retailers can extend credit to the less affluent consumer, then perhaps the costs of consumer durables to the poor may be reduced. That was the hope with which this study was begun.


Data were obtained through the generous cooperation of J.C. Penney, Montgomery-Ward, and Sears-Roebuck. However, to preserve confidentiality, those retailers are not identified elsewhere in this report.

The effective sample of 4,119 families was randomly selected from those families, of an original sample of 38,000, for which there was complete information on the variables used in the analysis. The sample was constructed so that it would contain a relatively large number of low-income families (under $5,000 per year). All the data were gathered in 1972. The data from each of the three retailers came from a different city: Pittsburgh, Chicago, and Los Angeles.

Available for every one of the 4,119 families were the following variables.

Evaluation of payment history (good or bad)

Retailer where have account



Divorced or separated


Number of dependents


Primary monthly income (dollars)

Presence of extra income

Own home

Rent home

Have telephone

Credit investigation made

In addition, families for each of the three retailer pairs had other variables in common


Other income

Second job

Property income


Months at present job

Wife working


Mobile home dweller

Live with parents

Months at present address

Checking account

Savings account

Department store reference

Loan reference

Other reference

Number of major and minor derogatories

Previous account

Retail or catalog customer

Application date

Highest monthly payment

Current monthly payment

For retailer A, a bad credit risk was defined as a customer more than 150 days past due or one who was written off with a balance in excess of $30. For retailer B,- bad risks were those accounts in collection. For retailer C, those customers two months or more past due were labeled bad credit risks. These discrepancies in definitions were unavoidable. In the analyses, the retailer variable may be expected to remove the linear component of such variation. As a digression, there were many differences in variable definitions when the data were first received: it required eight months to make the three data sets compatible.

One major shortcoming to this study is that the data describe consumers who have already been granted credit. Therefore, the results may be biased toward a certain type of good credit risk. An ideal sample for this study would require a retailer to grant credit to everyone he could find. In the absence of such an experiment, one must use these data and extrapolate with caution.

The high-income family and low-income family samples were somewhat different in comPoSitiOn. High-income families were more likely to be married and to awn homes. On the average, families in both groups had lived at their present addresses about the same length of time, but high-income workers, on the average, had been at their current jobs approximately six years longer than low-income workers. As one might expect, relatively more high-income families had loan references, checking accounts, and savings accounts. However. the percentages of those families able to supply department store credit references and other types of credit references were the same for both high-income and low-income families


Because of the large sample size and the capacity of the computer programs available, it was necessary to employ regression analysis instead of discriminant analysis. The dependent variable was binary: 0 for a bad risk, 1 for a good risk. Separate regressions were made across all high-income families and across all low-income families (Table 1). For the high-income customers, most pairwise correlations among the independent variables were below .3. However, there were high correlations (over .7) between the married and single variables, the own home and rent variables, and phone and credit investigation made variables. In addition, the correlations indicated that store B customers were more likely to have phones and to not have had credit investigations. Among the low-income families, store B patronage, age, and monthly income were highly correlated, as were number of dependents, age, and store patronage.

In Table 1 the regression coefficients are presented in the form of beta coefficients. These are standardized coefficients that can be directly compared both with other coefficients in the same regression and with coefficients in other regressions. [See Ferber for a detailed explanation.]



Overall, the regressions in Table 1 provided good fits to the data. Moreover, the statistically significant coefficients generally had logical signs. Besides those shown in Table 1, regressions with various subsets of the independent variables were examined to avoid possible collinearity problems. However, those results were consistent with those in Table 1 and have been omitted for brevity.

Because the set of variables common to all three retailers consisted of only thirteen variables, additional regressions were made across samples composed of data from the various pairs of retailers. These analyses allowed examination of an additional twenty-five variables. All the regressions for the pairs of retailers exhibited close overall fits, but fewer than half of the regression coefficients were statistically significant and many of these were those variables found to be significant in the regressions across all three retailers. These results are available from the author, but have been omitted since they add little to the remaining discussion.

The main question of this investigation was: Do the variables associated with good credit risks differ between high-income and low-income families? Overall, the regression coefficients for the high-income families did not significantly differ from those for the low-income families. [For details of the Chow Test of coefficient differences, see Beckwith or Johnston.] However, an all-coefficient test is perhaps too broad and one might ask whether or not coefficients for individual variables differed between the high-income and low-income samples.

In fact, the individual coefficients were significantly different for few variables (Table 1). These variables consisted of: married, single, divorced or separated, credit investigation made, and presence of extra income. The married variable was strongly positively associated with good credit risks among the high-income families; the single and divorced or separated variables were strongly negatively associated with good credit risks among the low-income families. The credit investigation made variable was a useful forecaster of good credit risks for both groups, but was of more importance for low-income families. Finally, although the coefficient for presence of extra income differed between the high- and low-income samples, the beta coefficients were small, indicating it was not an important predictor variable of credit riskness.


An analysis of a large sample of retail credit histories found variables that were useful predictors of credit riskiness. However, the predictive abilities of only a few of these variables differed between high-Income and low-income families. Based on this sample and the analysis to date, one cannot conclude that the identification of good credit risks among high-income and among low-income families requires different procedures. It is possible that further analyses examining possible interactions may yet find different predictor variables for high-income and low-income consumers.


Andreasen, A. (Ed.) Improving inner city marketing. Chicago: American Marketing Association, 1972.

Beckwith, N. E. Test of equality between coefficients in several linear regressions. Columbia University. Working paper. 1972.

Durkin, T. (Ed.) Proceedings of the Arden House Conference on Consumer Credit. New York: Columbia University, 1972.

Ferber, R. Market research. New York: McGraw-Hill, 1949.

Johnston, J. Econometric models. New York: McGraw-Hill, 1963.

Sturdivant, F. (Ed.) The ghetto marketplace. Press. 1969.  New York: Free