ONLINE APPENDIX (NOT FOR PUBLICATION) Appendix A: Appendix Figures and Tables

Transcription

1 ONLINE APPENDIX (NOT FOR PUBLICATION) Appendix A: Appendix Figures and Tables 34

2 Figure A.1: First Page of the Standard Layout 35

3 Figure A.2: Second Page of the Credit Card Statement 36

4 Figure A.3: First Page of the Alternative Layout 37

5 Table A.1: Test of Randomization Balance - Low-risk Clients Prominent Interest Maturity of the Disclosure Rate Featured Plan (1) (2) (3) Credit card limit (0.000) (0.000) (0.000) Credit card balance (0.000) (0.000) (0.000) Probability of using the revolving credit line (0.015) (0.001) (0.069) Average revolved balance (0.034) (0.002) (0.154) Average monthly Interest and fees charges (0.000) (0.000) (0.000) Probability of making a late payment (0.012) (0.001) (0.054) Time with the credit card (years) (0.001) (0.000) (0.006) Revolving interest rate * (0.003) (0.000) (0.014) F(8,13295) p-value Sample size Notes: Each column represents a separate OLS regression where the LHS variable is the corresponding treatment variable. F-test corresponds to a test that the coefficients on all variables are jointly equal to zero. Sample is restricted to low-risk clients. Robust standard errors in parentheses. * significant at 10%; ** significant at 5%; *** significant at 1% 38

6 Table A.2: Test of Randomization Balance - Medium-risk Clients Prominent Interest Maturity of the Disclosure Rate Featured Plan (1) (2) (3) Credit card limit (0.000) (0.000) (0.000) Credit card balance ** 0.000* (0.000) (0.000) (0.000) Probability of using the revolving credit line (0.028) (0.002) (0.127) Average revolved balance (0.053) (0.003) (0.240) Average monthly Interest *** and fees charges (0.000) (0.000) (0.001) Probability of making a late payment (0.022) (0.001) (0.096) Time with the credit card (years) (0.003) (0.000) (0.015) Revolving interest rate (0.007) (0.000) (0.031) F(8,3198) p-value Sample size Notes: Each column represents a separate OLS regression where the LHS variable is the corresponding treatment variable. F-test corresponds to a test that the coefficients on all variables are jointly equal to zero. Sample is restricted to medium-risk clients. Robust standard errors in parentheses. * significant at 10%; ** significant at 5%; *** significant at 1% 39

7 Table A.3: Test of Randomization Balance - High-risk Clients Prominent Interest Maturity of the Disclosure Rate Featured Plan (1) (2) (3) Credit card limit (0.000) (0.000) (0.000) Credit card balance (0.000) (0.000) (0.000) Probability of using the revolving credit line (0.030) (0.002) (0.130) Average revolved balance (0.048) (0.003) (0.216) Average monthly Interest and fees charges (0.000) (0.000) (0.001) Probability of making * a late payment (0.021) (0.001) (0.093) Time with the credit card ** (years) (0.003) (0.000) (0.014) Revolving interest rate (0.007) (0.000) (0.033) F(8,3170) p-value Sample size Notes: Each column represents a separate OLS regression where the LHS variable is the corresponding treatment variable. F-test corresponds to a test that the coefficients on all variables are jointly equal to zero. Sample is restricted to high-risk clients. Robust standard errors in parentheses. * significant at 10%; ** significant at 5%; *** significant at 1% 40

8 Table A.4: Price Sensitivity of Payment Plans Demand Risk categories Full Sample Low-risk Medium-risk High-risk (1) (2) (3) (4) Average take-up (concealed interest rate) (0.002) (0.002) (0.005) (0.006) Panel i: linear probability model Prominent interest rate (0.002) (0.003) (0.007) (0.008) Interest rate *** *** * (0.054) (0.060) (0.154) (0.186) Interest rate x Prominent interest rate ** (0.078) (0.086) (0.226) (0.252) N Panel ii: Logit marginal effects Prominent interest rate (0.002) (0.003) (0.007) (0.008) Interest rate *** *** * (0.054) (0.060) (0.165) (0.177) Interest rate x Prominent interest rate ** (0.079) (0.088) (0.231) (0.251) N Notes: Panel i presents coefficients of a linear probability model of take-up on interest rate, a dummy for emphasized advertisement, and the interaction of these two variables. Panel ii presents logit marginal effects. Column 1 presents estimates for the full sample, while columns 2 to 4 present estimates separately for low-, medium-, and high-risk clients. All estimates are weighted by the inverse of the probability that the client was selected so that they represent the original population. Robust standard errors in parentheses. * significant at 10%; ** significant at 5%; *** significant at 1% 41

9 Table A.5: Interest Rate Elasticities by Interest Rate Disclosure - Conditional on Revolving Interest Rate Revolving rate = 15.99% Revolving rate < 15.99% All Low-risk Medium-risk High-risk All (1) (2) (3) (4) (5) Prominent Rate disclosure? No *** *** *** (0.163) (0.234) (0.319) (0.257) (0.257) Yes *** *** * *** *** (0.166) (0.223) (0.307) (0.253) (0.222) p-value (elasticities are equal) N Notes: Column 1 reports the interest rate elasticities when the interest rate is concealed and when it is prominently disclosed, along with the p-value of a test that these two elasticities are equal, for the sample of clients with revolving interest rate equal to 15.99%. Columns 2 to 4 present the results separately by each risk category restricting the sample to those with revolving interest rate equal to 15.99%. Column 5 presents the results for the sample of clients with revolving interest rate lower than 15.99%. The interest rate elasticities are calculated based on a Linear Probability Model, using equation 2. The interest rate elasticities are calculated based on a Linear Probability Model, using equation 2. All clients in this sample received a payment plan offer, with interest rate equal to 3.99%, 7.99%, or 11.89%. Estimates in columns 1 and 5 are weighted by the inverse of the probability that the client was selected. Robust standard errors in parentheses. * significant at 10%; ** significant at 5%; *** significant at 1% 42

10 Table A.6: Interest Rate Elasticities by Interest Rate Disclosure - Heterogeneity by Risk Profile Determinants 1 st quartile 2 nd quartile 3 rd quartile 4 th quartile (1) (2) (3) (4) i. Heterogeneity by recurrence in late payments Prominent rate disclosure? No *** *** *** (0.332) (0.282) (0.269) (0.281) Yes *** ** *** *** (0.301) (0.283) (0.271) (0.287) p-value (elasticities are equal) N ii. Heterogeneity by average revolving balance Prominent rate disclosure? No *** *** *** ** (0.545) (0.429) (0.296) (0.188) Yes ** ** *** *** (0.508) (0.462) (0.310) (0.175) p-value (elasticities are equal) N iii. Heterogeneity by average credit usage Prominent rate disclosure? No *** ** *** (0.303) (0.299) (0.289) (0.266) Yes *** ** *** ** (0.300) (0.316) (0.274) (0.254) p-value (elasticities are equal) N Notes: This table reports the interest rate elasticities when the interest rate is concealed and when it is prominently disclosed, along with the p-value of a test that these two elasticities are equal when the sample is divided in quartiles of recurrence in late payments (panel i), average revolving balance (panel ii), and average credit usage (panel iii). These variables are constructed based on information prior to the experiment. The dataset provides information of up to 19 months prior to the experiment. The interest rate elasticities are calculated based on a Linear Probability Model, using equation 2. All clients in this sample received a payment plan offer, with interest rate equal to 3.99%, 7.99%, or 11.89%. All estimates are weighted by the inverse of the probability that the client was selected. Robust standard errors in parentheses. * significant at 10%; ** significant at the 5%; *** significant at 1% 43

11 Table A.7: Test of Randomization Balance - Conditional on Enrollment Prominent Interest Maturity of the Disclosure Rate Featured Plan (1) (2) (3) Credit card limit 0.000* * (0.000) (0.000) (0.000) Credit card balance * (0.000) (0.000) (0.000) Probability of using the revolving credit line (0.080) (0.005) (0.361) Average revolved balance (0.111) (0.007) (0.502) Average monthly Interest and fees charges (0.000) (0.000) (0.001) Probability of making a late payment (0.053) (0.003) (0.232) Time with the credit card (years) (0.006) (0.000) (0.029) Revolving interest rate ** * (0.016) (0.001) (0.071) Medium-risk (0.050) (0.003) (0.221) High-risk (0.052) (0.003) (0.234) F(10,651) p-value Sample size Notes: This table replicates Table 3 for clients who enrolled in a payment plan. * significant at 10%; ** significant at 5%; *** significant at 1% 44

12 Table A.8: Comparison of Main and Larger Scale Samples Full Sample High-risk clients Main Larger Scale Main Larger Scale Experiment Experiment Experiment Experiment (1) (2) (3) (4) Credit card limit 1, , (17.0) (9.2) (17.9) (7.2) Credit card balance (7.4) (4.1) (12.9) (7.1) Probability of using the revolving credit line (0.003) (0.002) (0.009) (0.003) Average revolved balance (0.002) (0.001) (0.006) (0.002) Average monthly Interest and fees charges (0.5) (0.3) (1.8) (0.9) Probability of making a late payment (0.003) (0.002) (0.008) (0.003) Time with the credit card (years) (0.03) (0.01) (0.06) (0.02) Revolving interest rate (0.01) - (0.02) - Sample size Notes: Column 1 presents averages of the corresponding variable for the main experiment sample, while column 2 presents averages of the corresponding variable for the larger scale experiment described in Appendix B. Sample is restricted to high-risk clients in columns 3 and 4. Robust standard errors in parentheses. 45

13 Table A.9: Test of Randomization Balance - Larger Scale Experiment Offered i = 6.40% Offered i = 9.59% Control group (1) (2) (3) Credit card limit (0.000) (0.000) (0.000) Credit card balance 0.000* (0.000) (0.000) (0.000) Probability of using the revolving credit line (0.007) (0.006) (0.008) Average revolved balance (0.012) (0.011) (0.015) Average monthly Interest and fees charges (0.000) (0.000) (0.000) Probability of making a late payment (0.004) (0.004) (0.005) Time with the credit card *** 0.002*** (years) (0.001) (0.001) (0.001) Medium-risk (0.004) (0.004) (0.005) High-risk (0.005) (0.004) (0.005) F(9,103106) p-value Sample size Notes: This table presents a test of randomization balance for the larger scale experiment presented in Appendix B. Each column represents a separate OLS regression where the LHS variable is the corresponding treatment variable. F-test corresponds to a test that the coefficients on all variables are jointly equal to zero.. * significant at 10%; ** significant at 5%; *** significant at 1% 46

14 Table A.10: Causal Effects of Enrolling in a Payment Plan on Probability of Default Full Sample Low-risk Medium-risk High-risk Reduced 2SLS Reduced 2SLS Reduced 2SLS Reduced 2SLS Form Form Form Form (1) (2) (3) (4) (5) (6) (7) (8) Offered i=6.39% 0.005* 0.006* (0.003) (0.003) (0.007) (0.007) Offered i=9.59% 0.006** 0.006** * (0.003) (0.003) (0.007) (0.007) Accepted i=6.39% 0.073* 0.110* (0.041) (0.059) (0.077) (0.059) Accepted i=9.59% 0.124** 0.149** * (0.050) (0.073) (0.095) (0.069) Control mean (0.002) (0.003) (0.006) (0.006) Sample size Notes: This table presents the reduced form and 2SLS estimates of the main effect of enrolling in a payment plan on the probability of default in the following 12 months, using the larger scale experiment carried out by the firm. Offered i=6.39% is a dummy variable equal to 1 if client received a payment plan offer with interest rate equal to 6.39%, while Offered i=9.59% is a dummy variable equal to 1 if client received a payment plan offer with interest rate equal to 9.59%. The omitted group received no payment plan offer. All estimates are weighted by the inverse of the probability that the client was selected so that they represent the original population. Robust standard errors in parentheses. * significant at 10%; ** significant at 5%; *** significant at 1% 47

15 Table A.11: Comparison of Compliers in the Main and in the Larger Scale Experiments Larger scale experiment Main Experiment Clients who enrolled at 11.89% because Sample Clients who enrolled at 9.59% the interest rate information was not prominently disclosed (1) (2) Credit card limit ,135.0 (37.4) (340.0) Credit card balance 1, ,073.5 (41.1) (269.0) Probability of using the revolving credit line (0.007) (0.088) Average revolved balance (0.007) (0.083) Average monthly Interest and fees charges (3.5) (26.2) Probability of making a late payment (0.007) (0.089) Time with the credit card (years) (0.10) (1.66) Revolving interest rate (0.00) (0.41) Notes: Column 1 presents averages of the corresponding variable for clients in the larger scale experiment (described in Appendix B) who were offered and enrolled in a payment plan with interest rate equal to 9.59%. Column 2 presents averages of the corresponding variable for clients in the main experiment who enrolled in a payment plan with interest rate 11.89% because the interest rate information was not prominently disclosed. The average for this group can be calculated by X = X D p ( X H X D ), where p = take-up when prominently disclosed take-up when not prominently disclosed, XD is the average for those who enrolled when information was prominently disclosed, and X H is the average for those who enrolled when information was not prominently disclosed. Robust standard errors in parentheses. 48

16 Appendix B: Effects of Information Disclosure on Subsequent Financial Outcomes for High-Risk Clients The results in section 4.2 reveal that the degree of interest rate disclosure significantly affects high-risk clients decisions. In particular, high-risk clients are 2.8 percentage points less likely to enroll in a payment plan with a high interest rate when the interest rate information is prominently disclosed. But does this affect their future financial position in any meaningful way? While the scale of the experimental design in this paper is not sufficient to estimate how this change in behavior translates into subsequent financial outcomes, such as default probability 1, it is possible to use a larger scale experiment carried out by the same firm with another sample of clients to estimate the causal effect of enrolling in a payment plan. The credit card company designed this larger scale experiment to measure the impact that enrolling in a payment plan has on default rates. The larger scale experiment had two differences in the sample selection relative to the main experiment analyzed in this paper. First, the company selected only clients with the highest revolving interest rate (15.99%) in the larger scale experiment. Second, the main experiment restricted the sample to 4 different credit card due dates, while the larger scale experiment did not have this restriction. Appendix Table A.8 presents averages for baseline variables in the main and in the larger scale experiments. The two samples are similar with respect to observables, especially considering only high-risk clients 2. The credit card company carried out this larger scale experiment in July of 2010 with 103, 116 clients, randomly allocated in three groups. In the first group, 34, 743 clients were offered a menu of payment plans with interest rate equal to 6.39%, in a second group, 49, 573 clients were offered plans with interest rate equal to 9.59%, and finally a control group of 18, 800 clients did not receive any payment plan offer. Clients in the treatment group received a payment plan offer for 3 consecutive months. Appendix Table A.9 presents regressions of each treatment group (offered i = 6.40%, offered i = 9.59%, and no offer) on clients characteristics at baseline. The p-value of a test that all coefficients in these regressions are jointly equal to zero is smaller than 0.1 in one case (p-value=0.087 for the regression using a dummy for control group as LHS variable). Most coefficients in these regressions are not statistically different from zero, with exception of time with the credit card. All results based on this larger scale experiment are not sensitive to including all baseline variables as controls. In order to estimate the impact of enrolling in a payment plan on default rates, I use the following specification: Y i = α + ρ 6.39 E 6.39,i + ρ 9.59 E 9.59,i + ε i (1) where Y i is default in the 12 months after the offer, and E r,i is equal to 1 if client i enrolled in a payment plan with interest rate equal to r. The default variable is equal to 1 if the client does not meet the minimum payment for 3 consecutive months in the 12 months after the payment plan offer 3. The default variable is defined for the credit card account as a whole, and not for the payment plan in particular. 1 In a reduced form regression of information disclosure on default for high-risk clients offered a high interest rate payment plan, the standard error on the information disclosure coefficient is equal to 2.3 percentage points. Although the estimate is not statistically different from zero, it would not be possible to reject that information disclosure has large effects on default, given the large standard errors. 2 Given the differences in sample selection explained above and the large sample size, the differences in baseline variables in the two experiments are statistically significant for most variables. 3 This is how the credit card company defines that a client is on default. When this happens, the company cancels the account, and the client will never be able to apply for this credit card again. 1

17 Coefficient ρ r is the causal effect of enrolling in a payment plan with interest rate r on outcome Y. Note that this causal effect can depend on the interest rate offered. In order to estimate ρ 6.39 and ρ 9.59, offers of payment plans with rates 6.39% and 9.59% are used as instruments for payment plan enrollment at these two rates (the omitted group received no payment plan offer). The 2SLS estimates should, therefore, be interpreted as the local average treatment effect (LATE) of enrolling in a payment plan with interest rate r for those who enroll in such plans, relative to a scenario where they had only the revolving option. Note that this strategy uses payment plan offer (not payment plan interest rate) as instrumental variable for enrollment. Therefore, these estimates are not affected by adverse or advantageous selection. These estimates would capture, however, ex-post incentives on repayment as those would be part of the causal effect of enrolling in a payment plan (Karlan and Zinman (2009)). The results presented in Table A.10 indicate that, on average, enrolling in a payment plan at either of these two interest rates significantly increase the probability of default in the following 12 months. Considering only the high-risk clients, enrolling in a payment plan induces more clients to default when the interest rate is higher (9.59%). Extrapolating these results, enrolling in a payment plan with an even higher interest rate should also increase the probability of default. Since prominent rate disclosure reduces the demand of high-risk clients for payment plans when the interest rate is equal to 11.89%, this information treatment likely reduces the probability of default for clients that change their payment plan enrollment decisions because of the treatment. It is important to point out that reducing the probability of default would not necessarily mean an increase in welfare, as clients might be exercising strategic default. However, a survey with credit card clients in Mexico reported in Seira and Elizondo (2014) shows that over 9/10 said that defaulting would decrease their welfare taking the benefits of defaulting into account. Although these estimates suggest that high-risk clients who did not enroll in a payment plan when offered the full information treatment are less likely to default, another estimate of interest is the reduced form impact of information disclosure on default probability. Assuming that information disclosure only affects subsequent financial outcomes through the enrollment decision, consider the model: 2

18 E r,i = α 1 + β r D r,i + ε 1,i Y i = α 2 + ρ r E r,i + ε 2,i (2) where β r is the causal effect of information disclosure on payment plan enrollment, and ρ r is the causal effect of payment plan enrollment on outcome Y i. Note that these effects can vary with r. Combining these two equations: Y i = α + π r D r,i + ε i, where π r = (β r ρ r ) (3) Therefore, the reduced form effect of prominent rate disclosure would be the product of the effect of prominent rate disclosure on enrollment (β r ) and the effect of enrollment on the outcome variable (ρ r ). This strategy is similar in spirit to Angrist and Krueger (1992) Two-Sample IV. However, instead of using estimates from the reduced form and from the first stage in order to back out the structural parameter, the strategy used here combines the first-stage estimates from my experiment, and the structural parameter estimates from the firm experiment in order to produce an estimate of the reduced form. That is, the overall effect of prominent rate disclosure on default. An implicit assumption in this strategy is that the treatment effect for those who enrolled in a payment plan with interest rate 9.59% is a good approximation to the treatment effect for those who enrolled in a payment plan with interest rate 11.89% because the interest rate information was not prominently disclosed. This hypothesis might be problematic, as evidence from section 4.2 reveals important heterogeneities in treatment effects. While this hypothesis can not be directly tested given the experimental design, it is possible to check whether these two groups are similar in terms of observable variables 4. This comparison, presented in Appendix Table A.11, suggests that these two groups are similar with respect to their observable covariates. Another assumption for this strategy is that information disclosure does not affect the probability 4 I can calculate average characteristics for the sample that enrolls in a payment plan with interest rate 9.59% directly by conditioning on clients who were offered and enrolled in a payment plan with this rate in the larger scale experiment. For those who enrolled in a payment plan with interest rate 11.89% because the interest rate information was not prominently disclosed, it is necessary to assume that all clients who enrolled in a payment plan when the information was prominently disclosed would have also enrolled if the information was in the fine print. In this case, for each observable variable X, the average for this group can be calculated by X = X D p ( X H X take-up when prominently disclosed D ), where p = take-up when not prominently disclosed, XD is the average for those who enrolled when information was prominently disclosed, and X H is the average for those who enrolled when information was not prominently disclosed. 3

19 of default through channels other than enrollment in the payment plan 5. In any case, this exercise would still capture the effects of information disclosure on default that are mediated through changes in payment plan take-up rates. The first experiment provides an estimate for β ( ˆβ = 0.028, s.e ), while the second experiment provides an estimate for ρ 9.59 (ˆρ 9.59 = 0.121, s.e ). Assuming that ρ 9.59 ρ 11.89, combining these two estimates yields the reduced form effect of information disclosure on probability of default for high-risk clients ˆπ = (0.12) ( 0.028) = (s.e ) 6. Note that the standard error is around 10 times smaller than the standard error of the reduced form effects of prominent rate disclosure based on the main experiment (0.25 vs 2.3 percentage points), allowing even modest effects of information disclosure on probability of default to be ruled out. Therefore, prominent rate disclosure has a significant effect in reducing default for clients that are induced to enroll in a payment plan with a high rate because the interest rate information is concealed in the fine print. However, since it affects only a small proportion of consumers, the aggregate effect of such policy would be small even for the population of high-risk clients. 5 This assumption would be violated if, for example, clients realize that the credit card company charges high interest rates when this information is disclosed, and become more prone to default regardless of enrolling in a payment plan. 6 Standard error is calculated using the Delta Method. 4

20 References Angrist, Joshua D. and Alan B. Krueger, The Effect of Age at School Entry on Educational Attainment: An Application of Instrumental Variables with Moments from Two Samples, Journal of the American Statistical Association, 1992, 87 (418), Karlan, Dean and Jonathan Zinman, Observing Unobservables: Identifying Information Asymmetries With a Consumer Credit Field Experiment, Econometrica, November 2009, 77 (6), Seira, Enrique and Alan Elizondo, Are Information Disclosure Mandates Effective? Evidence from the Credit Card Market, Social Science Research Network Working Paper Series, May