Under revision. For conference submission only. The Marginal Propensity to Consume out of Credit:

Transcription

1 Under revision. For conference submission only. The Marginal Propensity to Consume out of Credit: Evidence from Random Assignment of 54,522 Credit Lines Deniz Aydın May 2017 Abstract This paper investigates the effect of credit expansions on consumer borrowing and spending decisions using comprehensive data and a test-tube exogenous shock to credit availability. I design and implement a randomized trial at a European retail bank where credit lines are deliberately varied to 54,522 pre-existing cardholders. I find that credit availability has a precisely measured and economically large effect on spending and the use of credit. Contrary to conventional wisdom, the effect of credit is not confined to a small set of constrained consumers who are up against their limits. I show how the experimental findings can be used to perform novel, often non-parametric tests of competing models of intertemporal behavior (e.g. complete markets, permanent income, spender-saver, buffer-stock, impatient). I discuss the implications of findings for modeling financial sector linkages; formulating fiscal, monetary and macroprudential policy; and evaluating welfare in the credit market. I am grateful to Doug Bernheim, John Beshears, John Cochrane and Luigi Pistaferri for invaluable guidance. I also thank Adrien Auclert, Scott Baker, Chris Carroll, Raj Chetty, Bob Hall, Ayşe İmrohoroğlu, Tullio Jappelli, Pablo Kurlat, Jonathan Levin, Davide Malacrino, Jonathan Parker, Jorge Sabat, Amit Seru, Amir Sufi, Alp Şimşek, Aiga Štokenberga, Jialan Wang and Alonso Villacorta, as well as seminar participants at WashU, Chicago, Bocconi, Insead, Oxford, Yale, Fed Board, Turkish Central Bank, NBER SI, CEPR HF Workshop, AEA and NYU HF conference, for comments and discussions. Financial support from NBER/Sloan Foundation and SIEPR is greatly acknowledged. Washington University in St. Louis. daydin@wustl.edu. Website: sites.wustl.edu/daydin. 1

2 1 Introduction How does the financial sector interact with the broader economy? In light of the dramatic financial cycle of the last decade, policymakers developed renewed interest in understanding the consumption response to credit expansions as a critical issue to modeling macroeconomic fluctuations and formulating policy to offset them. The magnitude of the response can be used to discipline models of the household sector with varying degree of financial sector linkages, ranging from Friedman s permanent income model to Keynesian rule-of-thumb behavior. The heterogeneity of the response is a central input for formulating fiscal, monetary and macroprudential policy. Finally, the composition of the response is important for understanding the state-dependence and non-linearities of the financial accelerator, as well as evaluating welfare in the credit market. In this paper, I study the consumption response to a credit expansion using comprehensive longitudinal data on income, categorized spending and the balance sheet of consumers, and a test-tube shock to credit availability, overcoming the pervasive orthogonalization problem of credit supply -that credit often expands when wealth and income increases; and when credit risk, risk aversion, uncertainty and the interest rates decreases. To overcome the difficulty in identification, I design and implement a randomized trial at a large European retail bank where I deliberately vary the credit lines of 54,522 preexisting card holders. My subject pool consists of consumer and small business cardholders pre-approved by the bank for a credit line increase. I select at random 13,438 of this group as my control, and these consumers are withheld from credit line increases for 9 months, starting September The typical cardholder in the treatment group have their credit lines extended by a median 120% of monthly income. The increases in limits are initiated by the issuer and they are unannounced. Other features of the contract, such as the interest rate, remain unchanged. Therefore, the intervention can be classified as unexpected, and, by construction, exogenous shock to only credit availability. 1 Previewing the results, I find that credit availability has a precisely measured and economically large effect on spending and the use of credit. Consumer borrowing increases by 12 cents on the dollar after three months, with this marginal propensity to consume (MPC) converging to 17 cents on the dollar after 6 months. Strikingly, the effect of credit is not confined to a small fraction of credit constrained or hand-to-mouth consumers who are up against their credit limits. While proximity to 1 Sharp research design, large sample and simple control/treatment comparisons allow for precise and robust estimates. Moreover, as the credit shocks are large -median 120% of monthly income, the welfare cost of failing to smooth would not be small. 4

3 the limit is positively correlated with the MPC, this propensity remains quantitatively large even for those who are far from the limit. Indeed a large component of the average response is driven by consumers that utilize only a small fraction of their credit lines, but increase their borrowing on the margin. Analyzing spending patterns, I find that about two-thirds of the increases in borrowing that result from credit extensions are direct expenditures in durables and services. Credit, therefore, appears to mostly be used by consumers that make purchases with investment in features, and then pay down the incremental debt over time, rather than those that are under financial distress to smooth spending. I then explore the mechanism that drive sensitivity to credit within commonly used dynamic consumption models 2, and show how the experimental findings can be used to test their competing predictions -the model explicates which predictions of the commonly used models are inconsistent with the empirical evidence. It is well known that a shock to credit availability, which entails no wealth effects and does not change the interest rate, should have no effect on borrowing and spending behavior in the variants of benchmark permanent income model where consumption is proportional to wealth, or where consumers are constrained by only the natural limit of their resources. However, the deviations from the PI benchmark can not be rationalized by a simple spender-saver model, as a significant component of the average sensitivity is driven by unconstrained consumers that adjust their borrowing on the margin. Formal modeling points out that theories of precautionary savings, despite being able to account for the basic qualitative patterns, quantitatively predict responses an order of magnitude smaller than what is observed. Myopic models, calibrated to deliver high MPCs, are unable to simultaneously account for the dynamics of the response, as consumers, despite appearing very impatient in spending out of credit, are very patient when saving their way out of debt. I show that only a buffer-stock model with illiquid durables, in particular with one-sided adjustment costs -where credit shocks lower than reasonable adjustment costs trigger upward adjustment- can deliver many aspects of the MPC-C quantitatively. Related literature. This paper builds on and amalgamates a large literature in empirical macroeconomics estimating the consumption response to income shocks, 3 with the literature in finance estimating the borrowing response to credit shocks. 4 To the best of my knowledge, 2 Including Friedman (1957), Campbell and Mankiw (1989), Deaton (1991), Carroll (1997), Kaplan and Violante (2014), among others. 3 See Cochrane (1991), Blundell et al. (2008), Parker et al. (2013), Baker (2013); and Jappelli and Pistaferri (2010) for a comprehensive survey. 4 This literature, best exemplified by a seminal paper by Gross and Souleles (2002) and a recent paper by Agarwal et al. (2015), document some sensitivity of household borrowing to naturally occuring variation in 5

4 Table 2: Intertemporal models: MPC and Testable Predictions 0 MPC-C 1! Buffer-stock Spender- Prediction PI Buffer-stock w/durables Myopia saver (1) Credit affects behavior? N X X X X (2) Unconstrained respond? N X X X N (3) Delivers magnitude? N N X X N (4) Dynamics mean-reverting? N X X N N (5) Composition of spending? N N X N N Note. This figure orders a set of intertemporal models by the sensitivity to credit availability. At the left extreme is the permanent income model, where an increase in credit availability has no effect on behavior. At the right extreme is a simple rule-of-thumb model with no consumption smoothing where credit availability increases consumption one-for-one. Simulated values for the models are calculated in Section E. this is the first paper to study the consumer response to a truly exogenous shock to credit availability. In addition to testing competing intertemporal models, the experimental findings also contribute to the debate in macroeconomic modeling on the importance of consumption fluctuations on consumer welfare and macroeconomic outcomes. I aim to document cleanly identified partial equilibrium responses using microeconomic data that are crucial to map policies to equilibrium outcomes in fully specified macroeconomic models, with applications in studying dynamics of household leveraging, 5 understanding the effects of precautionary behavior on the business cycle, 6 formulation of fiscal, monetary credit limits. Although these papers control for numerous channels of potential endogeneity, apprehension remains, that some endogenous variation has not been purged. 5 MPC-C is the key moment to study the distributional consequences of financial shocks through the consumption decisions of the household sector using heterogenous agent incomplete markets models. See Hall (2011), Eggertsson and Krugman (2012) and Guerrieri and Lorenzoni (2015). 6 See Guerrieri and Lorenzoni (2009). 6

5 and macroprudential policy, 7 estimating the demand for consumer durables 8 and design of credit markets. Layout. Sections 2 describes the data and the institutional environment, and Section 3 details the experimental implementation. Section 5 builds a dynamic consumption model, and derives testable predictions to distinguish the models. A reader familiar with these models may directly skip to Section 4, which presents the empirical results. Section 6 concludes by discussing implications. 2 Data and Environment For the purposes of the field experiment, I collaborate with a large retail bank. The financial institution is a top ten credit card platform in Europe with more than 5 million customers, holds a 20% market share in the state retail market, and its customer base is representative of the local banked population. It offers a multitude of financial products, including credit cards, deposit and investment accounts, consumer loans, payment services, and insurance. I merge five types of data on bank s retail customers. First, I use credit card data, which includes card limits, within month transactions, end-of-month balances, payments made toward balances, and debt that is carried over between statement periods, of all credit cards a customer has at the bank. The expenditure data is comprehensive and each transaction is categorized into 18 subcategories (e.g., groceries, appliances, health expenses etc.) using a unique retailer point-of-sale machine identifier. Second, account balance sheet data contains information on the end-of-month balances on all the liquid assets (e.g. checking, savings, money market accounts, investments) and liabilities at the particular financial institution. I supplement this data with consumer credit bureau records, which contain information record total balances on all liabilities at other banks, as well as credit scores. Fourth, income information is also available for a subset of consumers with direct de- 7 In environments where demand determines output, optimal policies relate the covariance between the consumers marginal propensity to consume and balance sheet positions, either ex-ante via macroprudential policy, or ex-post as stabilization policy. See Jappelli and Pistaferri (2014), Auclert (2015), Farhi and Werning (2013) and Korinek and Simsek (2015). The magnitude of appropriate stimulus depends on obtaining an accurate measurement of the state-dependence and non-linearities of the multiplier. See Parker (2011). 8 See Carroll (1997), Mankiw (1982), Bernanke (1984), Grossman and Laroque (1990), Eberly (1994),Berger and Vavra (2015). 7

6 posit. It contains only after tax labor income and does not include financial income (e.g., interest, dividend, other capital income) or government transfers (e.g., benefits and social security income). Total income can be decomposed to its components (e.g. overtime, bonus, severance pay). Finally, there exists a rich set of demographic data, including age, gender, marital status, education, city of residence and profession. Table 3 reports summary statistics for a subset of relevant variables on the experiment participants and a random subsample of the universe of all customers of the financial institution. The data is of monthly frequency and the unit of analysis is an individual. Due to its administrative nature, it has far fewer problems with attrition, non-response, and measurement error than survey data sources. If an individual has multiple accounts, then accounts are matched using a unique citizenship number, and verified using a customer identification number, ensuring perfect match quality. Information regarding balance sheet variables are end-of-month calculations; however, credit card variables are endof-billing-cycle calculations. Institutional details. Credit cards and formal loans are the marginal source of credit for most consumers in the sample economy, and are used pervasively to finance consumption expenditures within pay-periods, as well as to transfer resources across pay-periods. 29% of individuals in the local economy reported using a credit card in the past year, and 20% report having borrowed from a financial institution -compared with 34% and 16% respectively for the Euro area. 9 Focusing on experiment participants with information on labor income, I find that the median participant spends in excess of 30% of their post tax monthly labor income using credit cards at the bank, and has credit card debt equaling 34% of monthly income. Credit card market. The credit card market under study has three notable distinctions from the US market. First, the maximum interest rate that can be charged on any credit card card is capped by the central bank at 24% APR, and this upper limit is binding for almost all customers. This allows me to ignore any pecking order on credit cards with potentially different rates, as well as allowing me to focus on the notion of credit constraints as quantity constraints. Second, credit card limits of individuals are capped by the regulatory authority to a maximum of four months of income. Third, the credit card contract allows for financing purchases with installments. Installment credit is available works just like a mortgage or an auto loan, and payments are 9 See Appendix Table 10 for details on the prevalence of formal loans. 8

7 Table 3: Summary Statistics RCT participants All cardholders (1) (2) (3) (4) Variable Mean s.d. Median Mean Age Gender Married Schooling Customer for months Credit card variables (at bank) Number of cards Limit 5,450 5,875 3,400 8,581 I(D Limit > 0) D Limit D Limit > 0 Spending 1,034 1, ,458 Debt 1,236 2, ,229 Debt Debt>0 Revolving debt 307 1, Installment debt 928 1, ,540 Debt/Limit (at bank) Credit card variables (outside of bank) Number cards Limit 6,179 16,198 1,000 16,675 Spending , ,991 Debt 2,370 8, ,937 Spending at bank/total spending Debt at bank/total debt Assets 5,755 41, ,794 Checking 1,604 10, ,229 Time deposits 3,044 34, ,102 Investments 1,106 12, ,463 Other Debt (excluding credit card) 5,684 17, ,109 Mortgages 2,702 15, ,211 Vehicle loans 167 2, Personal loans 2,814 6, ,530 Monthly Wage 2,977 5,032 1,801 2,576 Credit score 1, ,506 1,460 Non-performing loans Note. Columns (1), (2) and (3) are based on a 50,000 random sample of all credit card holders in August Column (4) is based on 52,267 consumer cardholders in August Unit of analysis an individual. Nominal variables expressed in local currency of Labor income information for the subset of customers with direct deposit. See text for definitions of variables. 9

8 Table 4: Summary Statistics: Selection, Sampling and Randomization Group Count Age Income Limit Debt Credit score (1) Census (2) All cardholders >5m ,488 8,664 2,261 1,460 (3) Participants 54, ,785 5,452 1,236 1,480 P( P = All ) < 0.01 < 0.01 < 0.01 < 0.01 < 0.01 (4) Treatment 41, ,870 5,887 1,208 1,505 (5) Treatment - US 27, ,047 6,704 1,044 1,555 (6) Treatment 13, ,434 4,203 1,548 1,402 (7) Control 13, ,494 4,121 1,523 1,401 P( T = C ) P( T + US = C ) < 0.01 < 0.01 < 0.01 < 0.01 < 0.01 Note. Table entries are group means using data from June to August Row (1) based on a 50,000 random sample of all credit card holders. Row (2) is based on 52,267 experiment participant consumer cardholders. Definitions of Rows (3), (4) and (5) are defined in Section 3. Income, credit limit and credit card debt variables expressed in local currency. Labor income information for the subset of customers with direct deposit. made periodically over a predetermined horizon, typically 2 to 12 months. 10 The credit contract therefore allows to separate two distinct types of debt contracts with distinct consumer motives for borrowing on them. Total credit card debt is the sum of this installment debt and revolving debt, and always indicates only balances carried across pay-periods, not within month transactions that are paid off. 3 Experimental Design Experiment participants are 54,522 pre-existing consumer and small business cardholders approved for a limit extension between June 2014 and August These extensions are automatic i.e. initiated by the issuer, and reflect the credit supply function, indicated and discussed in Appendix Table 8. The eligibility criteria for selection includes profitability 10 The installment contract works as follows. Suppose the consumer buys a refrigerator that costs $2000, with 4 installments. Her debt balance then increases by $2000 plus the interest charges. The first payment of $500 plus the interest charges show up in the end-of-month balance in the month of purchase. The remaining $1500 plus the interest charges shows up under the installments balance. If the consumer chooses to not pay any part of this $500 installment payment due at the end of the month, it would start bearing additional interest charges, and would show up under revolving balances. 10

9 and risk criterion, and the recency of the last limit increase. Rows (1) and (2) of Table 4 compares the 54,522 experiment participants to a 50,000 random subsample of the universe of cardholders. I find that experiment participants are, on average, 4 years younger, earn 12% more labor income, and have 1.5% lower credit scores, where all differences are statistically significant. The salient difference is that a typical experiment participant has 40% lower credit line than that of the typical cardholder. As the magnitude of the designated automatic limit increase for the typical participant increases her credit line to the level of the typical cardholder, the experiment participants appear to be cardholders that are catching up with the typical consumer in terms of credit line magnitude. I select at random 13,438 of this 54,522 as the control group, using a three step procedure. First, I stratify the participants into non-overlapping and exhaustive bins with respect to their end-of-month balances over limits. Second, using a random number generator, I draw a random sample n b from each bin, totaling up to 26,854 cardholders. 11 Finally, I distribute the randomly selected 26,854 cardholders to control and treatment groups using a coin-flip procedure: within each bin, I divide the cardholders 50/50 equally across the two groups. This randomization procedure leads to a ternary classification of participants. The two identical halves of the 26,854 are denoted the control (C) and treatment (T) groups, and the remaining 27,938 are denoted the undersampled (U). The distribution of these three classes is displayed in Figure 6 and the corresponding summary statistics are given in Table 4. The control group of 13,438 is withheld from additional credit limit increases for T = 9 months starting September The treatment group of 13,416 and the undersampled (T-US), are pushed for the limit increases. Subjects in the treatment group then learn about the limit extension through their credit card statement -they do not know that they are participants in an experiment. 11 In order to maximize statistical power, the ratio of the sample sizes n b are proportional to the standard deviation of the outcome variable. See See List et al. (2011). Further details on the sampling and randomization are given in Appendix B. 11

10 Figure 1: The Event Study Note. The top figure displays the average credit card debt for the treatment and control group. The x-axis is time in months and t = 0 corresponds to September The y-axis is the debt level, values normalized by pre-intervention values. The dashed lines indicate theconfidence intervals for the estimate of the mean. The bottom figure displays the cumulative response of credit card debt to a dollar increase in credit lines, obtained using the distributed lag equation (1). Note that seasonality due to new-years affects both groups of participants. 12

11 4 Event Study I begin my empirical analysis in by plotting the average credit card debt for the 52,267 consumer cardholders, by treatment and control groups, between March 2014 and June 2015, in the top panel of Figure 1. The y-axis displays the change in debt relative to its level at the onset of the experiment, D t D 1, which is tied to cumulative spending within the same period from the intertemporal budget constraint. The dashed lines indicate the 95% confidence intervals for the estimate of the mean. The figure shows stable debt levels for the treatment and the control groups prior to the intervention, but a sharp and significant increase in credit card borrowing by the treatment group after the intervention -the difference is the causal effect of credit availability. The Marginal Propensity In order to obtain a value interpretable as the marginal propensity, I normalize the change in debt by change in limits, using a standard distributed lag form, DD it = T=t  f j DL it j + # it (1) j=0 The dependent variable is the change in credit card debt DD it = D it D it 1. The explanatory variable is the change in credit lines DL it = L it L it 1. For individuals that did not experience an increase at time t, the corresponding DL it equals zero. The error # it accounts for debt growth due to factors other than credit line extensions, such as shifts in preferences or income shocks. The coefficient f 0 measures the contemporaneous effect of a unit increase in credit lines on the dependent variable, and the partial autocorrelation coefficients f 1,...,f t measure the additional responses after t months. The object of interest is then the cumulative response after t months MPC(t) = t j=0 f j. Estimating Equation (1) by ordinary least squares identifies the effect of credit availability from variation in both the assignment to control group and magnitude of the limit increase. However, although the assignment of credit lines I(DL it > 0) is random, the variation in the magnitude of the limit change DL it DL it > 0, although possibly uncorrelated with the error # it, is not random. Therefore, I instrument the magnitude of the limit change with an indicator variable I(DL it > 0), which is orthogonal to # it by construction, and limits the amount of variation utilized to whether any credit line increases was received as a part of the experiment. 13

12 The sampling and randomization also ensures orthogonality between the assignment of credit lines and all other variables, in particular the residual # it, and allows to avoid potential omitted variable bias or other confounding factors. I therefore do not control for any cross-sectional or time variables. I focus on the subsample of 52,267 non-business experiment participants, using data from T = 9 months between September 2014 to June Use of survey weights and calculation of standard errors are explained in Appendix 3. The bottom panel of Figure 1 displays the impact of a unit change in credit lines on total credit card debt after t months, using results from estimation of equation (1). For brevity, I only report the cumulative response, which gives the impulse response to credit availability. A dollar increase in credit limits increases credit card debt by 12 cents after 3 months and by 17 cents after 9 months. The response is highly statistically significant. Debt rises sharply and significantly in the first quarter following a credit limit increase, with convergence happening only after four months. Subsequent marginal coefficients decline in magnitude and significance, and there is no evidence that debt changes beyond the first four months following the limit change, with marginal coefficients f 5,...,f 9 jointly insignificant. The cumulative response of the treatment group does not exhibit a reversal, implying that credit does not only shift the timing of the spending by pulling it forward from the future, but on the contrary has long-run effects. When the credit limit is relaxed, consumers appear to monotonically lever up to a new target debt level. As there is little evidence of non-linear dynamics, a static specification that allows for only an immediate debt response, with consumers permanently levering up to a new debt level, is well suited to rationalize the impulse response. The magnitude of the MPC-C is statistically and economically consistent across a broad range of specifications that use different forms of variation, or various controls. I begin in Table??, which reports estimation results from specifications that utilize different samples and forms of variation. The first row displays the baseline results of estimating equation (1) by instrumental variables, where I instrument the magnitude of the limit change with an indicator variable I(DL it > 0), which limits the amount of variation utilized to whether any credit line increases was received as a part of the experiment. The second row displays the results of estimating equation (1) by ordinary least squares, which identifies the effect of credit availability from variation in both the assignment to control group and magnitude of the 14

13 limit increase. These estimates are slightly smaller, with the quarterly MPC-C being 7 cents on the dollar, and the long-run response 10 cents on the dollar. The composition of the MPC remains stable. The third panel replaces DL it with I(DL it > 0), a dummy variable indicating whether a limit increase occured in period t, as the main explanatory variable, and measures the average dollar increase in expenditures caused by the increase in credit line. Consumers on average increased their credit card debt by 447 after 3 months, and 582 after months in the local currency, with p-value< I also examine the response of the experiment participants debt outside the bank using credit bureau data, in order to ensure that increases in credit card debt does not represent mere balance shifting. I replace the left-hand side variable in Equation (6) with credit card debt outside the bank. 12 Consumers do substitute away debt outside the bank, by decreasing their debt outside the bank by 0.01 cents after 3 months and 0.02 cents after 9 months, confirming the results that the increase in debt does not represent mere balance shifting. For brevity, I report results on balance shifting and defaults in the Appendix Section??. First, using credit bureau information, I estimate the effect of a change in credit lines on card debt outside the financial institution, and verify that borrowing on other cards remain stable. Therefore the results do not represent mere balance shifting. Second, I report that no detectable difference between the treatment and control group exists by the end of the randomized trial with respect to late payments or non-performing loans. Therefore, the magnitude of the MPC-C is statistically and economically significant across a broad range of specifications that use different forms of variation, or various controls, indicating that a shock to credit availability has substantial effects on the consumption behavior, rejecting complete markets models, and models where consumption is determined by permanent income. Therefore, the results across a variety of specifications confirm a statistically and economically significant effect of credit availability on credit card borrowing. 12 The financial institution gathers credit bureau information on cardholders on an approximately quarterly basis. Therefore I am unable to plot the full impulse response of debt outside the bank, but I am able to measure the cumulative response after 3 and 9 months. 15

14 Table 5: Borrowing Response to the Credit Shock (A) (B) (C) (D) (E) X S C MPC 3 MPC 9 MPC9 Inst MPC9 Rev MPC9 Other f 1 f 2 f 3 f 4 Â t>4 f t (1) DL IV (0.009) (0.013) (0.008) (0.01) (0.002) (2) DL OLS (0.009) (0.014) (0.011) (0.010) (0.015) (3) I(DL) OLS (35) (42) (28) (33) (51) N R2 f i f t f s Table 6 (1) (2) (3) (4) (5) Specification X MPC 3 MPC 9 MPC9 Inst MPC9 Rev MPC9 Other (A) DL IV (0.009) (0.013) (0.008) (0.01) (0.002) (B) DL OLS (0.009) (0.014) (0.011) (0.010) (0.015) (C) I(DL) OLS (35) (42) (28) (33) (51) f i f t f s N R2 16

15 Heterogeneity Average responses mask substantial heterogeneity and do not distinguish whether all consumers respond equally or whether the response is driven by the part of the sample. If the cardholders are facing a binding credit constraint, or maintaining a buffer of resources as precautionary savings, then low level of liquidity can indicate that consumption is depressed and the consumer has a higher propensity to spend credit on arrival. In order to analyze heterogeneity of the response by disposable resources, I sort individuals into bins based on pre-intervention values, and estimate Equation (1) separately for each bin. Such binning allows me to see whether individuals with lower ex-ante liquid resources exhibit larger sensitivities to credit shocks. I use credit line utilization, ratio of credit card debt across pay periods to credit lines, as my primary measure of disposable resources. This utilization measure directly identifies cardholders for whom the credit limit constraint is binding and yields sharper results. When binning, I use the average values in the year prior to the onset of the experiment. This value is pre-determined with respect to decisions taken thereafter. Figure 2 plots the cross-sectional distribution of the responses by credit line utilization bins. First, Figure 2 points to substantial heterogeneity in MPC-C based on a consumers disposable resources. Unsurprisingly, consumers closer to the constraint give larger responses. For example, consumers with credit line utilization exceeding 90% accumulate more than 80 cents of debt per unit increase in limits, in excess of five times the average response. These are consumers whom are at a corner solution to their intertemporal problem, for whom the the value of borrowing and spending significantly exceeds that of keeping available credit. They are hand-to-mouth, and their MPC should be approximately equal to 1. Second, and strikingly, I find that consumers with substantial resources exhibit economically very large responses to an increase in credit availability. For example, consumers that utilize less than 10% of their credit lines at the onset of the experiment, accumulate 8 cents of debt per unit of credit line increase. Similarly, consumers with utilization rates between 20% and 30% -median utilization lies in this interval- accumulates in excess of 10 cents of debt per unit of credit line increase. All these measures are highly statistically significant. Third, the response of the unconstrained consumers, considered jointly with the histogram of credit line utilization, suggests that not only unconstrained consumers give economically and statistically large responses to a change in credit availability, but they indicate that a significant component of the average response is driven not by a small 17

16 Figure 2: Heterogeneity Note. Figure plots the cross-sectional distribution of the marginal propensity to consume out of credit, after 3 months. Consumers in the first bin had an average credit line utilization rate of zero to 10% in the year before the intervention; and those in the second bin had an average credit line utilization between 10% to 20% of their credit lines and so on. Estimates are obtained using the distributed lag Equation (1) on a sample of 52,267 consumer cardholders. The right axis reports the cumulative coefficients MPC t = Â t j=0 f j after 3 months. x-axis is credit line utilization, defined as the ratio of interest bearing debt to credit line. Also displayed the histogram of credit line utilization, with fraction in each bin indicated on the left. 18

17 fraction of credit constrained consumers. On the contrary, the average response is due to a substantial precautionary motive through which credit constraints affect the consumption behavior. I interpret Figure 2 as showing that a shock to credit availability has substantial effects on the consumption behavior of unconstrained consumers, rejecting a simple spendersaver model. Qualitatively, these findings are consistent with models featuring concave consumption rules such as the buffer-stock model, as well as previous research that documents a lack of consumption smoothing, in particular among consumers with low assets. However, as discussed Section 5, quantitatively, the MPC-C of unconstrained consumers are incompatible any realistic calibration for the discount factor of the buffer-stock model. Composition What are consumers doing with borrowed money? Understanding the composition of spending could be very helpful in understanding why the consumers exhibit a high MPC, as well as how the MPC will vary over the business cycle. The comprehensive nature of the expenditure data allows me to analyze transaction patterns using two different types of data. First, I can decompose the total credit card debt to its subcomponents: change in revolving debt and change in installment debt. Revolving debt consists of end-of-month balances that are not paid off in full. Installment balances are debt directly incurred at the time of purchase to finance expenditures. Second, I can decompose total credit card transaction volume to decreasingly aggregated measures of categorical consumption expenditures: expenditures in durables, nondurables, and services, as well as sectoral spending in each of the 18 categories. In order to estimate the effect on categorical transactions, I replace the left-hand side variable in Equation (1) with the change in spending, Note that spending is a flow variable linked to the change in debt via an accounting identity, therefore is analyzed in levels. I then calculate the share of spending in each category, as a fraction of total increases in spending For the rest of the paper, I assume that the increase in credit card debt represents an increases in expenditure, for three reasons. First, foreshadowing my results, I find that two-third of the additional debt is accumulated to finance direct expenditure on durables and services. Second, I verify from transaction data that consumers in the treatment group increase their expenditures, not just revolve debt keeping spending constant. Third, payday loans, and other forms of relatively expensive unsecured borrowing, is not a feature of the credit market under consideration. 19

18 Figure 3: Impulse response of credit card debt, by debt type Note. Figures plot the marginal propensity to consume out of credit. Estimates are obtained using the distributed lag equation (1) in Section 4 on a sample of 52,267 consumer cardholders. The figure on the left represents total credit card debt. The figures in the middle and on the right represent installment debt and revolving debt respectively. Figure 4: Spending patterns in installments Note. Figure plots the categorical composition of the spending response to credit. First, I estimate the response of categorical expenditure for each of 18 spending categories, using the distributed lag equation (1) in Section 4 on a sample of 52,267 consumer cardholders. The figure then plots the share of each category on the total additional spending done by the treatment group. For example, the groceries bar indicates that 18% of the additional spending done by the treatment group has gone to grocery spending. The red bars represent the total share of spending in nondurables, durables and services respectively. 20

19 Figure 3 reports the response of credit card debt by this debt type. Across a broad range of specifications, I find that roughly-two-thirds of the total debt accumulated comes in the form of installments. Increases in installments represent direct consumption expenditures, used to finance the purchase of durables and services. Therefore, unlike the literature on finance that measures the marginal propensity to borrow, I can bound the marginal propensity to consume out of credit at 0.1. Moreover, threequarters of the final installment debt is accumulated only after a quarter. Therefore the immediate response of a typical consumer to an increase in credit availability is to finance consumption of durable goods with an investment nature, such as furniture, appliances, health, and education. Installment debt increases throughout the second quarter of the experiment, however there is no evidence of any increases after the second quarter. (p = 0.77). Note that installment debt and revolving debt are distinct debt contracts with different interactions to the consumer income and balance sheets. In the baseline intertemporal consumption model with a single liquid asset, theory predicts that revolving debt will be used as a self-insurance device, comoving negatively with income shocks. Figure 4 then displays the contribution of each spending category in the total change in spending. I find that spending on consumer durables (furniture, clothing, electronics, appliances, jewellery), and services (insurance, tourism, health, education) represents two-thirds of the additional spending. Only a quarter of the increase in spending is due to nondurables (groceries, restaurants, hobbies, cosmetics, retail, recreation). Spending on groceries and restaurants constitute almost all of the increase in nondurable spending. Finally another 10% of the additional spending is taken out as cash advances. These findings complement the results that use information in debt type that, expenditures in durables and services represent a large share of the increase in spending. Commodities usually classified as services have substantial amount of durability. In addition, consumer durables have some collateral value, give flow utility, and spending on services such as education, health and insurance constitutes investment in human capital, and have substantive effects on well-being and labor market outcomes. Therefore the increase in borrowing does not appear to finance a short lived increase in nondurable consumption. Moreover, the spending on consumer durables and services have a lumpy nature and exhibit highly cyclical patterns. This is in contrast to spending in nondurable consumption, for which the consumers have a preference for smoothness. 21

20 Figure 5: Intertemporal models: Consumption Note. Figure illustrates the consumption rule for the permanent income, buffer-stock and rule-of-thumb models. MPC-C is the difference in consumption with respect to a relaxation of the credit limit dl. Dynamics 5 Tests of Competing Models In this section I study a discrete time infinite horizon model of intertemporal consumption behavior. This specification nests the spectrum of commonly used macroeconomic models. I use this framework to guide the empirical findings and derive testable economic predictions to distinguish the models. Consumers have expected utility form preferences over nondurable consumption C t, with flow utility u(c t ), and discount factor b. Labor income Y t is uncertain and follows an exogenous Markov process P(Y t ) over time. Consumers smooth income fluctuations using a single liquid asset with interest rate R. Holdings of this liquid asset is given by A t, borrowing on which is possibly bounded above by a limit L. The well known optimality condition for the intertemporal consumption problem 14 gives 14 The consumer problem is given by, Ṽ(A 0, Y 0 ; L) = max E0 P C t, A t+1 " Â b t u(c t ) t=0 # s.t. C t + DA t apple Y t ; Y t+1 P(Y t ); A t+1 L (2) When the model does not permit analytical solutions under general conditions, I obtain numerical results. 22

21 the credit constrained Euler equation, u c (C t )=br E P t [u c (C t+1 )] + µ (3) where µ is the Lagrange multiplier on the credit constraint. The solution C 0 (A 0, Y 0 ; L) is a non-decreasing function of net assets, income shock and the credit limit. The consumption response to an increase in credit availability, marginal propensity to consume out of credit (MPC-C), then equals DC 0 DL. Figure 5 highlights the relevant features of the consumption rule under various models, and Table 2 orders a set of intertemporal models by their predicted values of the magnitude of the marginal propensity to consume out of credit. Test 1: Permanent income model In the benchmark version of the permanent income 15 (PI) model consumption admits an explicit formula, " # C0 PI = R 1 A 0 + Y 0 + R Â R t E0 P [Y t ] t=1 (4) and is an affine function of asset holdings and future income. It s slope -MPC out of wealth- is equal to the annuity factor and constant at all levels. Consumption depends only on the first moment of the value of endowment, therefore precautionary motives are absent. Importantly, consumption is a martingale, and debt is a unit root process. Under the PI model, a shock to credit availability that does not result in a change in the borrowing rate, and that does not entail any wealth effects have no effect on consumption behavior -the MPC-C is zero; H0 PI DC0 : DL = 0. Any positive effect of credit availability on spending constitutes a rejection of the permanent income model, making it a high powered test. See Section E for the details of parametrization and calibration. 15 In particular, the certainty equivalent case where u c (C) =C, b = R 1 and L =. See Friedman (1957). Similar conclusions also apply if markets are complete or if the borrowing limit equals the natural limit -the present value of the sum of endowment sequence. See Bewley (1977). 23

22 Test 2: Simple Heterogeneity: Spender-saver model If the aggregate consumption comoves with aggregate credit, deviating from the permanent income benchmark, one simple way to rationalize this is to consider a world populated with two types of consumers: some following the permanent income hypothesis with C0 = CPI 0 and a small fraction whom are hand-to-mouth and consume all disposable resources, C0 HtM = A 0 + Y The latter group of consumers may either be myopic 17 (e.g. either impatient or following a simple rule-of-thumb), or forward looking with a currently binding credit constraint. The spender-saver model predicts a mass of consumers at the constraint for whom the marginal utility of consuming today exceeds marginal value of saving and consuming tomorrow. For these consumers, changes in credit availability translates one-for-one to changes in consumption, their MPC is 1. The magnitude of the average response then depends on the fraction of rule-of-thumb consumers. In this model with simple exogenous heterogeneity, distributional dynamics are muted, and credit shocks are born only by a small set of constrained consumers. Therefore testing whether credit availability affects behavior of unconstrained consumers constitutes a test of the spender-saver model, H0 SS : DC 0 DL A 0 > 0 = 0. Test 3: Buffer-stock model Models featuring precautionary savings are qualitatively consistent with the finding that all consumers, not only those with binding constraints, respond as if they experienced a DL increase in wealth. For example, in the buffer-stock model 18 consumption will be depressed relative to the permanent income benchmark, as uninsurable income uncertainty raises the expected value of marginal utility on the right-hand-side of the Euler equation (3), which increases precautionary savings and leads to a drop in consumption -even if the credit constraint is not strictly binding. The precautionary drop is larger at low holdings of liquid assets, leading to a consumption rule C that is concave in assets. When credit limits expand, a looser credit constraint is less likely to bind and makes consumption less responsive 16 See Hall and Mishkin (1982) and Campbell and Mankiw (1989), recently used by Eggertsson and Krugman (2012) and Korinek and Simsek (2015) to study household leveraging dynamics. 17 With discount rates significantly higher than the interest rate, as in Krusell and Smith Jr (1998). 18 Which corresponds to the case where u c (C) is convex, b < 1 R and L < LNatural, as in Deaton (1991), and Carroll (1997). 24

23 to income shocks, therefore decreases the volatility of future consumption and expected marginal utility. Consumer then relaxes the precautionary buffer, and the consumption rule expands towards C PI, even for unconstrained consumers. See Figure 5. Theories of precautionary savings, despite being able to account for the basic qualitative patterns, are unable to account for magnitudes. In Section E, I provide a parametric, quantitative test of the buffer-stock model, and show that any MPC-C in the order of 10% per quarter for unconstrained consumers appear clearly inconsistent with the benchmark buffer-stock model, using parameters in a conventional range. 19 H0 BS : DC 0 DL < 10%. Test 4: Impatient model A common way to deliver quantitatively large MPCs in incomplete market models is to calibrate a low discount factor that makes the consumer highly impatient. The consumption rule of such myopic models lie in the inner envelope of rule-of-thumb and buffer-stock behavior, with the discount factor determining the level of consumption between the two extremes. A final set of predictions then can be derived using the longitudinal features of the data, to discipline the discount factor and rule out impatient models. Identification of the discount factor is due to the tension between a high MPC and the dynamics of the response: if consumers are impatient in consuming out of credit, the same discount factor would imply that they should not appear very patient when saving their way out of debt. As an example, for a completely myopic agent with b = 0, the MPC-C is 1, and savings are zero, therefore these consumers are expected to remain at the constraint. Only if consumers are sufficiently patient to save their way out of debt, then both the level of debt, and the MPC can be mean-reverting. This tension points to the limitations of a range of myopic models, in simultaneously accounting for the magnitude of the MPC-C and the longitudinal features of credit line utilization. In Section E, I provide simulation evidence that, under reasonably calibrated parameters, there exists no impatient calibration of the standard incomplete markets model that can rationalize both the observed levels of MPC- C and mean-reversion of credit card debt. If the large consumption response to credit expansions are due to impatience, then Spending of durables is investment in the model: 19 Foreshadowing the empirical findings, the lower bound of the 95% confidence interval for the estimate of MPC-C based on the findings is strictly above 10%. Comparing this lower bound, one could reject any parameter combination with b 2 [0.85, 1], R 2 [1.01, 1.3], g 2 [0.5, 10]. 25

24 Moreover, the spending response to a credit shock bounds below the spending response to a wealth shock. For example, Guerrieri and Lorenzoni (2015) shows that MPC = MPC-C + R 1 R MPC-P, where MPC-P is the MPC out of a permanent income shock. For example, under the permanent income model, MPC-C=0, MPC-P=1 and MPC= R R 1, the annuity factor. For the buffer-stock model, Carroll (2009) presents simulation evidence that MPC- P is assuming R=1.04 (quarterly), a rough approximation of quarterly MPC will be 13%+0.85* =17%. 6 Implications and Conclusions Understanding economic downturns preceded by credit driven expansions requires an understanding of the expansion itself. In this paper, I have estimated the consumer spending and debt response to a credit shock, and used this information to test competing predictions of commonly used models dynamic consumption models. My emphasis has been on the strengths of sharp research design coming from a randomized trial -hopefully, this paper will serve as an example of the usefulness of test-tube experiments in answering traditional macroeconomic questions. Findings also shed light on the lackluster performance of consumer spending during the recovery -if the big decline in household spending was due to liquidity problems, then why has it taken so long for consumption to jump back up again, given that interest rates are at all-time lows and banks are lending again? Why is the recovery not unleashing pent-up demand? This study points to two possible answers. First, the findings indicate that a significant portion of consumer-spending sensitivity to credit is driven by purchases of durable goods such as furniture, home improvement and appliances, which are purchased infrequently, and only if they anticipate enough future income to justify the purchase. If consumers binged on durables before the recession, then there would be a hangover effect, with consumers waiting before spending again. Second, the research speaks to the issue of precaution: the primary factor affecting the behavior of a prudent consumer is uncertainty about the future. Uncertainty with respect to either can have a very large effect on consumer spending and thereby exacerbate the economy s downward plunge. I conclude by discussing the implications of experimental findings -in particular the magnitude, heterogeneity and composition of the MPC- for the standard toolkit of macroeconomic modeling and formulating policy. 26