CONSUMPTION SMOOTHING AFTER THE FINAL MORTGAGE PAYMENT: TESTING THE MAGNITUDE HYPOTHESIS

Transcription

1 FORTHCOMING: REVIEW OF ECONOMICS AND STATISTICS NOTE CONSUMPTION SMOOTHING AFTER THE FINAL MORTGAGE PAYMENT: TESTING THE MAGNITUDE HYPOTHESIS Barry Scholnick School of Business, University of Alberta, Edmonton, Alberta, Canada, T6G 2R6 April 2012 Abstract We examine whether the magnitude of an anticipated income change impacts consumption smoothing (i.e. the magnitude hypothesis). Even though this hypothesis has been discussed for 50 years, we are one of the first to provide formal statistical evidence to support it. We consider the natural experiment of an individual s final mortgage payment, which is an anticipated income change, and examine how it impacts credit card expenditure. We can identify causality because the dates of final mortgage payments across individuals are uncorrelated with unobserved determinants of consumption. Using an event study methodology, we provide evidence to support the magnitude hypothesis. Funding for this project was provided by the Social Sciences and Humanities Research Council of Canada (SSHRC) and the Federal Deposit Insurance Corporation (FDIC). We thank Mark Watson (the Editor) and three anonymous referees. Jason Allen, Christopher Carroll, Dimitris Georgarakos, Rasmus Fatum, Haifang Huang, Andrew Leach, Nadia Massoud, Slava Mikhed, Antony Saunders, and seminar participants at the European Central Bank Payments Conference, the Bank of Canada, the CFS Athens Conference on Household Finance and the Granada University/Chicago Fed Payments Conference provided valuable comments. 0

2 1. INTRODUCTION Consumption smoothing predicts that consumption should be smoothed if future income changes are anticipated. This prediction is of central importance in many branches of economics, and has generated an extremely large empirical literature, as surveyed in Browning & Crossley, (2001), Meghir & Pistaferri, (2010) and Jappelli & Pistaferri, (2010) who conclude that there is now considerable evidence that consumption appears to respond to anticipated income increases, over and above what is implied by standard models of consumption smoothing (p. 47). They argue, however, that the reasons for this failure of the theory are not yet well understood (p. 48). One potential reason for this failure is that the magnitude of the anticipated income change may impact the response of consumption. We call this argument the magnitude hypothesis. Discussions of the magnitude hypothesis date back more than 50 years to Kreinin (1961). A variety of explanations have been proposed for why magnitudes should matter, the most common of which is bounded rationality. Browning and Collado (2001), argue that because of bounded rationality, individuals do smooth (consumption) if there are large and predictable income changes (p. 682) but that they will not bother to adjust optimally to small income changes since the utility cost of doing so is small (p. 690). Similarly, Hsieh (2003) summarizes the bounded rationality argument by noting that there may be costs associated with the mental processing of these (small) forecastable income changes (p. 404). Much of the existing evidence supporting the magnitude hypothesis (e.g. the literature review of Browning and Crossley (2001)) is derived from comparing findings across different studies, where the sizes of the income changes differ between the studies. On the one hand, Browning and Crossley (2001) point to the finding of Parker (1999) that consumption overreacts in response to small anticipated income change (US household stop paying the 7% social security deduction at the end of September). On the other hand, Browning and Crossley (2001) point to evidence of consumption smoothing following the medium sized anticipated income change examined by Hsieh (2003) (anticipated payments from the Alaska Permanent Fund), and large income changes examined by 1

3 Browning and Collado (2001) (predictable bonuses paid to Spanish workers, which are equivalent to two months salary) 1. While this cross study evidence seems consistent with the magnitude hypothesis, a more powerful test is to examine a single type of anticipated income change that varies in magnitude across individuals. The aim of this paper is to provide such a test of the magnitude hypothesis, using a unique database and natural experiment design. Our data consists of a large sample of individual level, monthly, mortgage and credit card accounts, provided to us by a Canadian bank. The natural experiment concerns the very final monthly mortgage payment, which pays off the mortgage in full. Following Coulibaly and Li (2006) and Stephens (2008), we argue that the final payment of a long term debt contract (such as a mortgage) can be considered as an anticipated change in financial resources available. This is because the individual knows in advance the date of their final mortgage payment. Once the mortgage is paid in full, and the stream of monthly mortgage payments stops, the individual will have an anticipated increase in financial resources (which is analogous to an anticipated increase in disposable income). Consumption smoothing thus predicts that there should be no impact on consumption on the date of the final mortgage payment. We test this hypothesis by comparing the credit card expenditure of individuals who make a final mortgage payment (the treatment group), with those who continue making monthly mortgage payments (the control group). The key element of the identification strategy is that the date of each individual s final mortgage payment is independent of the date of every other individual s final mortgage payment. We argue that because the timing of this anticipated change in resources is uncorrelated with unobserved determinants of spending, we can identify a causal relationship from the timing of the final mortgage payment, to the individual s credit card expenditure response 2. 1 Browning and Crossley (2001) suggest that the welfare costs of ignoring these anticipated income changes can be used as a metric for determining if the anticipated income receipts are large or small. They argue that the welfare costs of ignoring the (large) Spanish bonuses is equivalent to an annual loss of a month s consumption; the welfare costs of ignoring the (medium sized) Alaska payments is equivalent to an annual loss of a week s consumption; and the welfare cost of ignoring the (small) September change in Social Security payments is equivalent to an annual loss of an afternoon s consumption. 2 As highlighted by Agarwal, Liu and Souleles (2007) this may not be the case for anticipated income changes that are clustered for all individuals in the same month(s) of the year (e.g. from anticipated 2

4 Our paper contributes to the literature in a variety of ways. In terms of our natural experiment methodology (i.e. using the final payment on a long term debt as a measure of an anticipated income change) we follow the approach initiated by Coulibaly and Li (2006) and Stephens (2008). However, while both of these papers examine consumption smoothing following the final payment of a long term debt, neither examines the magnitude hypothesis. Our paper is thus the first to utilize the final debt payment natural experiment in the specific context of testing the magnitude hypothesis. Our paper is also part of the growing literature which utilizes the exactly measured data available in monthly credit card bank accounts (e.g. Gross and Souleles (2002) and Agarwal et al (2007) among many others). In this respect, our paper is somewhat similar to the test of anticipated income changes of Agarwal et al (2007) 3. Unlike Agarwal et al (2007), however, our monthly credit card data is matched to monthly mortgage account data, thus allowing us to use the final mortgage payment as a natural experiment. This particular natural experiment allows us to specifically test the magnitude hypothesis because we can measure the exact date and magnitude of the anticipated income change, where there is a significant variance in the magnitude of income changes across individuals. Our paper is thus the first to use monthly credit card and mortgage account data to test the magnitude hypothesis. We know of only two papers (Kreinin, (1961) and Souleles (1999)) that have formally tested the magnitude hypothesis by examining a single kind of income change which varies in magnitude across individuals. Neither of these papers, however, finds strong support for the magnitude hypothesis 4. Our paper is similar to these papers in that we distinguish between large and small income changes by squaring the anticipated income term in a regression on consumption. The central advantage of our natural government payments, or from anticipated salary bonus payments). If the timing of anticipated income receipts is clustered in time, it is difficult to disentangle whether each individual s consumption is responding to the specific income change or to other macroeconomic factors at that time. 3 Agarwal et al (2007) examine the impact of an anticipated income change (tax rebates) on monthly credit card consumption. Because they measure the income change from tax rebates using a dummy variable, their income change data does not have large variance. They do attempt to examine the magnitude hypothesis by using family status data as a proxy for magnitude i.e. married couples receive double the tax rebate as single individuals. 4 Kreinin (1961) examines Israeli reparations payments but his results reject the magnitude hypothesis, while Souleles (1999) examines anticipated tax refunds in the US, but he concludes that it is not possible to reach a firm conclusion about the magnitude hypothesis. 3

5 experiment design over these papers, however, is our ability to identify causation because the timing of final mortgage payments across individuals can be assumed to be uncorrelated with unobserved determinants of spending. Furthermore, our monthly credit card and mortgage account data provide significant advantages compared to the survey type data used in these previous studies, in terms of measurement accuracy, data frequency and sample selection DATA - TREATMENT AND CONTROL GROUPS Our main database consists of individual level monthly credit card and mortgage accounts provided to us confidentially by an individual Canadian bank (summary statistics in Table 1). We use credit card statement data to measure credit card expenditure, and monthly mortgage statements to measure anticipated income changes. The period of our data runs from December 2004 to June The bank that provided us with their credit card and mortgage data is a full service retail bank that provides a full set of financial services to its clients, including investments, mortgages, credit cards and deposit and checking accounts. We examine the behavior of a treatment group (i.e. those who make a final mortgage payment) relative to a control group (i.e. those who continue to make mortgage payments) in an approach following Agarwal et al (2007), Coulibaly and Li (2006) and Stephens (2008). Our aim is to construct the control group so that it is as similar in as many respects as possible to the treatment group, with the one exception that individuals in the control group continue to make mortgage payments whereas those in the treatment group make their final mortgage payment. Our data allows us to measure the total mortgage balance outstanding for each individual for each month, thus we define the 5 In addition to these contributions our paper is also one of the first to examine anticipated income changes using individual level Canadian data. DeJuan, Seater, and Wirjanto (2010) examine this issue using Canadian Provincial time series data. 6 A related literature has examined the Retirement Consumption Puzzle, which examines whether consumption responds on the date of retirement, even though the date of retirement is predictable in advance (e.g. Banks et al (1998), Aguila et al (2011) and many others). Our data does not include the age of individual mortgage holders, thus it is not possible to directly verify if retirement date is interacting with the date of the final mortgage payment. However, Masier and Villanueva (2011), using detailed Spanish mortgage level data, show that banks are very reluctant to provide mortgages when the final scheduled payment date occurs when the mortgage holder is 65 years or older, because of the payment risk resulting from lower income and decreased health. 4

6 control group as the group of individuals who have a remaining mortgage balance of only $ or less. On average this implies mortgage payers with only slightly more than one year s worth of mortgage payments remaining 7. Based on this criterion there are approximately 4000 individuals in our control group. Our treatment group consists of those individuals for whom we can observe that they have paid off their mortgage in full. Because our data consists of the mortgage balance outstanding each month we can identify an individual as making a final mortgage payment by observing a monthly series of lower and lower outstanding mortgage balances, which ultimately reaches a balance of zero. An important issue when testing consumption smoothing, is that the timing of the anticipated income change needs to be exogenous or predetermined in advance (i.e. a predetermined date for the final payment on a debt contract). If, however, the individual is able to determine the timing of when they receive the income change (for example by paying off the mortgage in full as a lump sum on a date determined by the individual), then the income change is endogenous, and the test of consumption smoothing is not appropriate 8. Our data allows us to observe predetermined final mortgage payments, as distinct from final lump sum payments (full details of which are provided in the Web Appendix). Based on these data characteristics we are able to identify 147 individuals who can be included in our treatment group EMPIRICAL SPECIFICATION The structure and interpretation of our empirical specification closely follows that used by Gross and Souleles (2002a) and Agarwal, Liu and Souleles (2007), whose data has a very similar structure to ours (i.e. monthly individual credit card accounts). Following Gross and Souleles (2002a) and Agarwal et al (2007), our dependent variables is monthly credit card expenditure (CCExpenditure), for individual i at time t, and the key independent variable (anticipated income changes) is the magnitude of the final mortgage payment made at time t, which we designate FINAL. 7 This approach to defining our treatment and control groups is very similar to Stephens (2008) whose treatment group consists of individuals making a final payment on a car loan, and whose control group consists of those with one year or less remaining on their car loans. 8 This issue is also addressed by Stephens (2008), who excludes individuals who prepay their car loans in advance, from his treatment group. 9 As a comparison, Coulibaly and Li (2006) identify 286 individuals who have paid off their mortgages, out of their total sample of almost mortgage holders. 5

7 ( ) To test the magnitude hypothesis we follow Kreinin (1961) and Souleles (1999) by including the square of the anticipated income term (FINAL_SQ) and examine if the quadratic term has a negative coefficient. A negative coefficient on the quadratic term implies a hump shaped response of expenditure as the magnitude of the anticipated income change (FINAL) increases. Equation (1) also includes a number of other control variables (Z) which we describe below, as well as month fixed effects (time) and individual fixed effects (CustID). Following (Petersen, 2008) all our panel data results use clustered robust standard errors, with clustering at the level of the individual over multiple months. Consumption smoothing implies that the coefficients on FINAL and FINAL_SQ in (1) are insignificant, because the anticipated income change (the final mortgage payment) should not have a significant impact on monthly expenditure. Following closely from Gross and Souleles (2002a) and Agarwal et al (2007), we interpret this equation as an event study. The event of interest is the final mortgage payment, which is normalized across individuals in the treatment group so that the final payment event occurs at the event date t=0 for each treated individual. Thus, for the treatment group, if a particular mortgage payment is the final mortgage payment (i.e. the event date at t=0) then FINAL takes the dollar magnitude of that payment. Conversely, if a mortgage payment is not the final payment (i.e. a date other than the event date), then FINAL is coded as zero. FINAL is coded as zero for all individuals in the control group over all periods. The distributed lag term on FINAL can be interpreted as the event window (n, m), which runs from t=n to t=m. We have the flexibility to examine event windows over different periods both after and, as a placebo test, before the final mortgage payment. In all of the models, three separate control variables from our monthly bank account data are added to control for any systematic differences between the treatment 6

8 and control groups (Z in eq (1)). These variables are (1) credit utilization rate i.e. the ratio of the individual s credit card debt outstanding relative to their credit card credit limit for each month, (2) the individuals credit card credit limit, which is set by the bank for each individual, and changes periodically and (3) the FICO score for each individual, which is an external measure provided by a credit rating agency and captures past credit behavior by the individual across all credit products. 4. RESULTS 4.1. Main Result Results for tests of equation (1), for event windows after the event, (i.e. (0,0), (0,2), (0,4) and (0,6)) are reported in Table 2. F tests indicate that the FINAL and FINAL_SQ terms are (with one exception) jointly significant in the different models, indicating rejection of consumption smoothing. An important summary statistic used by Gross and Souleles (2002) and Agarwal et al (2007) is the cumulative response of expenditure to FINAL and FINAL_SQ. This is measured by adding the monthly lag coefficients over all months within each event window, and captures the total impact on expenditure over those months. The key test of the magnitude hypothesis is that the cumulative response to FINAL_SQ is negative and significant, indicating a hump shaped response in expenditure as FINAL increases in magnitude. Table 3, Panel A indicates that the cumulative response to FINAL_SQ is negative and significant for all windows, indicating that the magnitude hypothesis cannot be rejected Extensions Our data allows us to examine two extensions to equation (1). First, we examine whether the magnitude hypothesis still holds when we restrict the sample, for both the treatment and control groups, to only non-credit constrained individuals. (It is a standard argument that consumption smoothing may not hold if individuals are credit constrained). We define credit constrained individuals as being credit card revolvers (i.e. those who pay any positive interest on credit card debt) and not credit card transactors (i.e. those 7

9 who pay zero interest in any given month) 10. We can measure this because our credit card data measures the total interest paid each month by each individual (summary statistics for the group of transactors only are in Table 1). Second, we examine whether the magnitude hypothesis still holds when we divide the magnitudes of FINAL by income in equation (1) (i.e. relative as opposed to absolute magnitudes). Our data does not include measures of current individual income (because banks do not typically collect this data). However, our data does include the Canadian post code of each individual, thus we are able to measure post code level estimates of individual income (full details of these data are provided in the Web Appendix). Table 3 (Panels B to D) provides results for the cumulative responses (as in Gross and Souleles (2002) and Agarwal et al (2007)) of FINAL and FINAL_SQ for these alternative specifications. 11 Table 3 (Panels B to D) indicates that, across these specifications, the cumulative FINAL_SQ responses are negative and (with one exception) significant for the (0, 0) and (0, 2) event windows. This implies that we cannot reject the magnitude hypothesis for these alternative specifications for these shorter event windows Marginal Effects In order to examine the economic magnitudes of these hump shaped relationships, Table 4 reports on the marginal effect of FINAL and FINAL_SQ on credit card expenditure, when FINAL is measured at its 25 th, 50 th or 75 th percentile levels in equation (1). All these estimated marginal effects, across all models are significantly different from zero, which is inconsistent with consumption smoothing. The final two columns of Table 4 show the change in predicted credit card expenditure when FINAL moves from its 25 th to 50 th percentile, and from its 50 th to 75 th percentile. In all cases (with one exception) the predicted change in expenditure when FINAL moves from its 25 th to 50 th 10 Since credit card debt is generally one of the more expensive forms of consumer debt, it can be argued that any utilization of expensive credit card debt (i.e. any positive monthly interest charges on the credit card) could imply some level of credit constraints. We also replicate the procedures used by Agarwal et al (2007) and Gross and Souleles (2002) (i.e. defining credit constraints as a utilization rate above 75%) and find that the magnitude hypothesis still holds. 11 Full regression results for all these subsidiary models are available on request. 8

10 percentile is greater than when it moves from its 50 th to 75 th percentile, which is consistent with the hump shaped relationship as FINAL gets larger Placebo Tests In order to examine the robustness of our results, as a placebo test we examine the cumulative response of expenditure to FINAL and FINAL_SQ for different event windows that all end in the month before the event month (i.e. (-1,-1), (-2,-1), (-4,-1) and (-6-1)). Table 5 shows that cumulative responses of credit card expenditure to the FINAL_SQ terms are either insignificant or in one case marginally positive (indicating a U shape). These results show that there is no evidence of the hump shaped response in the months before the event month. This confirms that the magnitude of FINAL is indeed a causal factor in the hump shaped expenditure response we observe after the event. 5. CONCLUSIONS Even though the magnitude hypothesis has been discussed for 50 years (at least since Kreinin, 1961), most of the discussions in favor of the hypothesis have been informal. These include the comparison of results across studies (e.g. Browning and Crossley (2001)), or discussions in the concluding sections of papers regarding the possible interpretation of results (e.g. Browning and Collado, (2001), Hsieh, (2003), Coulibaly and Li, (2006))) However, those few formal statistical tests of the magnitude hypothesis, which examine a single type of anticipated income change that varies in magnitude across individuals (e.g. Kreinin, 1961, Souleles, 1999), have generally not provided strong support for the hypothesis. The contribution of our paper is to provide a new test of the magnitude hypothesis, which shows formally that magnitudes do matter. An important issue for future research is to understand in more detail exactly which cognitive or behavioral processes could be underlying the magnitude hypothesis. While most authors discussing the magnitude hypothesis have invoked bounded rationality (e.g. Kreinin, 1961, Browning and Collado (2001) and Hsieh (2003)), other authors have suggested that magnitudes could matter because of related behavioral hypotheses such as mental accounting (e.g Souleles (1999)) or inattention (e.g. Coulibaly and Li (2006)). Given that these behavioral explanations are somewhat related but not 9

11 exactly the same, future research to distinguish between these explanations, possibly using laboratory type behavioral experiments, will increase our understanding of exactly which cognitive processes underlie the magnitude hypothesis. 10

12 TABLE 1: Descriptive Statistics: Individual Level Monthly Bank Account Data Obs Median Mean Std Dev 1: All Credit Card Accounts CONTROL GROUP (All Months, Mortgage Balance Outstanding < $10 000) Credit Card Expenditure ($ /month) Card Utilization - Balance/Limit (%) Credit Card Credit Limit ($) FICO Score TREATMENT GROUP (Data for Month of FINAL) FINAL (Final Monthly Mortgage Payment) ($) Credit Card Expenditure ($ /month) Card Utilization - Balance/Limit (%) Credit Card Credit Limit ($) FICO Score : Transactors Only (Monthly Credit Card Interest Due = $0) CONTROL GROUP (All Months, Mortgage Balance Outstanding < $10 000) Credit Card Expenditure ($ /month) Card Utilization - Balance/Limit (%) Credit Card Credit Limit ($) FICO Score TREATMENT GROUP (Data for Month of FINAL) FINAL (Final Monthly Mortgage Payment) ($) Credit Card Expenditure ($ /month) Card Utilization - Balance/Limit (%) Credit Card Credit Limit ($) FICO Score

13 TABLE 2: Impact of Final Mortgage Payment on Credit Card Expenditure Main Tests of Eqn (1) as discussed in Section 4.1. Event Window (0,0) (0,2) (0,4) (0,6) Variable (lag) Final(0) * * (0.269) (0.260) (0.305) (0.397) Final(1) (0.216) (0.231) (0.245) Final(2) * (0.304) (0.302) (0.352) Final(3) (0.214) (0.262) Final(4) * (0.345) (0.330) Final(5) (0.394) Final(6) 0.721** (0.298) Final_Sq(0) ** *** ** ** ( ) ( ) ( ) ( ) Final_Sq(1) * ( ) ( ) ( ) Final_Sq(2) ( ) ( ) ( ) Final_Sq(3) ( ) ( ) Final_Sq(4) ( ) ( ) Final_Sq(5) ( ) Final_Sq(6) * ( ) Card Utilization (%) 7.088** 7.733** 13.03*** 13.50*** (3.225) (3.205) (1.095) (1.182) Card Limit (ln) 398.1*** 371.0*** 487.1*** 550.6*** (81.97) (96.44) (88.79) (109.2) FICO Score 0.548** 0.587** 0.777*** 0.879*** (0.240) (0.245) (0.218) (0.270) Constant -3,155*** -2,964*** -4,296*** -4,937*** (842.7) (992.9) (826.9) (1,010) F test: All Final_Sq 6.91*** 2.77*** 1.82* 2.05** + All Final terms =0 F test: All Final_Sq 5.67** 3.49** * terms = 0 Observations 35,425 33,506 31,283 28,286 R-squared

14 TABLE 3: Cumulative Response of Expenditure to FINAL and FINAL_SQ Main Specification of Equation (1) described in Section 4.1 (Panel A) Alternative Specifications of Equation (1) described in Section 4.2 (Panels B to D) Cumulative Response is Sum of Coefficients in Event Window, i.e. Final_Sq and Final Income in Panels C and D Measured at Post Code Level (Annual Average Individual Income). Data Details in Web Appendix Event Window (0,0) (0,2) (0,4) (0,6) Panel A: Absolute Magnitudes Final_Sq Coefficients ** *** * * ( ) ( ) ( ) ( ) Final Coefficients ** * ** Panel B: Absolute Magnitudes Only Transactors ( ) ( ) ( ) ( ) Final_Sq Coefficients ** *** ( ) ( ) ( ) ( ) Final Coefficients * ** Panel C: Relative Magnitudes (Final/Income) ( ) ( ) ( ) ( ) Final_Sq Coefficients * ( ) ( ) ( ) ( ) Final Coefficients * * ( ) ( ) ( ) ( ) Panel D: Relative Magnitudes (Final/Income) Only Transactors Final_Sq Coefficients ** ** ( ) (177786) ( ) ( ) Final Coefficients * * ( ) ( ) ( ) ( ) 13

15 TABLE 4: Marginal Estimates of Expenditure at 25th, 50th and 75th Percentiles of FINAL or FINAL/Income Marginal Estimated described in Section 4.3 (Monthly value of FINAL Mortgage Payment) Panels C and D: Income measured using post code level data. (Annual Average Individual Income). Full Data Details in Web Appendix. Percentile 25th 50th 75th Difference in Percentile Panel A: Absolute Magnitudes Value of FINAL Window Marginal std e Marginal std e Marginal std e th 25 th 75 th - 50 th (0,0) *** *** *** (0,2) *** *** *** (0,4) *** *** *** (0,6) *** *** *** Panel B: Absolute Magnitudes Only Transactors Value of FINAL Window (0,0) *** *** *** (0,2) *** *** *** (0,4) *** *** *** (0,6) *** *** Panel C: Relative Magnitudes (Final/Income) Value of FINAL/Income Window (0,0) *** *** *** (0,2) *** *** *** (0,4) *** *** *** (0,6) *** *** *** Panel D: Relative Magnitudes (Final/Income) Only Transactors Value of FINAL/Income Window (0,0) *** *** *** (0,2) *** *** *** (0,4) *** *** *** (0,6) *** *** ***

16 TABLE 5: Placebo Tests: Cumulative Responses to FINAL and FINAL_SQ in Windows Before Event Described in Section 4.4 Cumulative Response is sum of coefficients in event window i.e Final_Sq and Final Event Window (-1,-1) (-2,-1) (-4,-1) (-6,-1) Panel A: Full Sample Final_Sq Coefficients ( ) ( ) ( ) ( ) Final Coefficients * ( ) ( ) ( ) ( ) Panel B: Transactors Only Final_Sq Coefficients * ( ) ( ) ( ) ( ) Final Coefficients ** ( ) ( ) ( ) ( ) 15

17 References Aguila, E., Attanasio, O., & Meghir, C. (2011) Changes in Consumption at Retirement: Evidence from Panel Data. Review of Economics and Statistics Agarwal, S., Liu, C., & Souleles, N. S. (2007). The reaction of consumer spending and debt to tax rebates--evidence from consumer credit data. Journal of Political Economy, 115(6), Banks, J., Blundel, R. & Tanner, S. (1998); Is There a Retirement-Savings Puzzle?, American Economic Review 88(4), Browning, M., & Collado, M. D. (2001). The response of expenditures to anticipated income changes: Panel data estimates. American Economic Review, 91(3), Browning, M., & Crossley, T. F. (2001). The life-cycle model of consumption and saving. Journal of Economic Perspectives, 15(3), Coulibaly, B., & Li, G. (2006). Do homeowners increase consumption after the last mortgage payment? An alternative test of the permanent income hypothesis. Review of Economics and Statistics, 88(1), DeJuan, J. P., J. J. Seater, and T. S. Wirjanto (2010). Testing the Stochastic Implications of Permanent Income Hypothesis Using Canadian Provincial Data. Oxford Bulletin of Economics and Statistics, 72,

18 Gross, D. B., & Souleles, N. S. (2002). Do liquidity constraints and interest rates matter for consumer behavior? Evidence from credit card data. Quarterly Journal of Economics, 117(1), Hsieh, C. (2003). Do consumers react to anticipated income changes? Evidence from the Alaska permanent fund. American Economic Review, 93(1), Jappelli, T., & Pistaferri, L. (2010). The consumption responce to income changes. Annual Review of Economics, Johnson, D. S., Parker, J. A., & Souleles, N. S. (2006). Household expenditure and the income tax rebates of American Economic Review, 96(5), Kreinin, M. E. (1961). Windfall income and consumption: Additional evidence. American Economic Review, 51, Masier, G. and Villanueva. E. (2011) Consumption and initial mortgage conditions - evidence from survey data Working Paper 1297, European Central Bank. Meghir, C., & Pistaferri, L. (2010). Earnings, consumption and lifecycle choices. Handbook of Labor Economics Parker, J. A. (1999). The reaction of household consumption to predictable changes in social security taxes. American Economic Review, 89(4), Petersen, M. A. (2008). Estimating standard errors in finance panel data sets: Comparing approaches. Review of Financial Studies, 17

19 Souleles, N. S. (1999). The response of household consumption to income tax refunds. American Economic Review, 89(4), Stephens, M.,Jr. (2008). The consumption response to predictable changes in discretionary income: Evidence from the repayment of vehicle loans. Review of Economics and Statistics, 90(2),

20 FORTHCOMING: REVIEW OF ECONOMICS AND STATISTICS WEB APPENDIX FOR: NOTE CONSUMPTION SMOOTHING AFTER THE FINAL MORTGAGE PAYMENT: TESTING THE MAGNITUDE HYPOTHESIS Barry Scholnick School of Business, University of Alberta, Edmonton, Alberta, Canada, T6G 2R6 April

21 WEB APPENDIX: DATA APPENDIX In this appendix we provide additional details on various data issues. A1. Identifying Predetermined Dates of Final Mortgage Payments An important element in defining our treatment group is to ensure that we only include final mortgage payments in the treatment group that are predetermined in advance rather than lump sum final payments (which could be considered endogenous). To do this we exploit the fact that many mortgage holders make arrangements with the bank to make a steady stream of equal mortgage payments which continues until the final payment at a known future date. The date of the final mortgage payment can be considered predetermined, if this stream of equal payments remains until the final mortgage payment. Our monthly mortgage balance outstanding data allows us to observe whether individuals make their final series of payments as a steady stream of equal sized monthly mortgage payments, or alternatively whether the size of the final mortgage payment is of significantly different size compared to the previous stream of payments (which can be considered as a final lump sum payment on a date determined by the individual). Specifically, the criteria we use to include an individual in our treatment group is that the outstanding mortgage balance must decline in equal monthly increments (which we define as within 10% of any other payment) over time until it reaches zero. Furthermore, we only include an individual in our treatment group if we can observe equal monthly payments over at least the final four months before the mortgage balance outstanding reaches zero. By constraining our treatment group to those where we can actually observe that at least the final four monthly payments are of equal size, implies that we can ensure that, at a minimum, the date of the anticipated income change is known to the individual at least five months in advance. In practice, this implies that individuals with a final mortgage payment in the first 4 months of our sample cannot be included in our treatment group because we cannot observe 4 consecutive equal monthly payments before the balance reaches zero. A2. Credit Constraints: Revolvers and Transactors As an extension to our main results, we re-estimate our results to ensure that our results are not being driven by credit constraints. An important element of the consumption smoothing literature is the argument that if credit constraints exist then consumption smoothing is predicted not to occur (see e.g. Browning & Crossley, (2001), Meghir & Pistaferri, (2010) and Jappelli & Pistaferri, (2010). Consumption smoothing implies that an individual is able to increase consumption by borrowing in advance of the expected income increase, and then pay back that loan after the receipt of the income increase. If the individual is credit constrained however, then she may not be able to borrow in advance of the anticipated income increase, and thus she will not be able to smooth her consumption. For this reason, it is important that we determine whether our results are being driven by credit constraints. 20

22 In this paper we operationalize credit constraints by defining credit constrained individuals as being credit card revolvers (i.e. those who pay any positive interest on credit card debt) and not credit card transactors (i.e. those who only use the credit card for transaction purposes and who pay zero interest in any given month). Our credit card account level data allows us to measure exactly the total interest paid each month by each individual, thus we can measure exactly whether an individual is a revolver or a transactor. This revolver/transactor based definition of credit constraints is similar but somewhat different to the definition of credit constraints used by Agarwal et al (2007) and Gross and Souleles (2002). These authors use the credit card utilization rate (i.e. the monthly balance outstanding/credit card limit) as a measure of credit constraints. One concern with this type of definition however, is that credit card debt is generally one of the more expensive forms of consumer debt, thus it can be argued that any utilization of expensive credit card debt (i.e. any positive monthly interest charges on the credit card) could imply some level of credit constraints. Essentially, our revolver/transactor based definition of credit constraints is a significantly stronger definition of credit constraints compared to the credit card utilization definition, in that even an individual with a relatively low utilization rate would still typically be making some use of expensive credit card debt, which could imply some degree of credit constraints. We rerun our tests of equation (1) with the exception that both our treatment as well as control groups are now restricted to being only transactors. In addition, we replicate the procedures used by Agarwal et al (2007) and Gross and Souleles (2002) (i.e. defining credit constraints as a utilization rate above 75%) and find that the magnitude hypothesis still holds. A3. Matching Post Code Level Census Data While our credit card and mortgage account level data has significant advantages in terms of the exactness with which we can measure monthly credit card expenditure as well as the final mortgage payment income change, one disadvantage to our bank account data is that banks do not generally collect updated data on the current income of their customers. Thus we are not able to use the bank data to test whether FINAL/Income (rather than simply FINAL in (1) above) impacts consumption smoothing. Similar constraints are faced by Agarwal et al (2007) and Gross and Souleles (2002) in their use of credit card account level data to test the consumption response to income changes. However, one important advantage of our data, relative to Agarwal et al (2007) and Gross and Souleles (2002) is that we know the Canadian postal code of each individual in our data base. This means we can utilize post code level Canadian Census data to measure the post code level of income for each individual, which allows us to test the impact on expenditure of FINAL/Income, with income measured at the post code level. We argue that the post code level measure of income provides a reasonable proxy for individual income or wealth, based on the assumption that an individual s income or wealth will be reflected in the average income in the specific neighborhood (i.e. post code) in which an individual lives. Statistics Canada and Canada Post provide a standard technique for matching individual six digit post codes to the smallest geographic sized for which census data is available (called the Dissemination Area or DA). It is well known in the literature (e.g. 21

23 Barrington-Leigh and Helliwell, 2008) that there are orders of magnitude fewer individuals in each post code (measured at DA level) in Canada compared to each zip code in the US which significantly adds to the precision with which we can measure these demographic variables. This is primarily caused by the small population size of Canadian Post Codes because Canada has a six digit post code system compared to the five digit US Zip Code system. Each DA has an average of 400 to 700 people, compared to more than people in each geographic area based on US census and zip code data. In addition to providing post code level data on income, we can also use post code level census data to compare the characteristics of our treatment and control groups. An important objective in the construction of treatment and control groups is that the characteristics of the individuals in the two groups be as similar as possible (with the exception of the treatment variable, i.e. final vs. continuing mortgage payers). Appendix Table 1 in the Web Appendix (below) makes comparisons between the treatment and control groups across a large number of post code level demographic variables. The key finding is that there does not appear to be systematic differences in these variables between the treatment and control groups. A4. Canadian Banking Institutions 1: Selection into Small or Large Monthly Mortgage Payments An important issue in our tests of the magnitude hypothesis is whether individuals systematically select into small or large mortgage payments for example if richer individuals systematically make larger monthly mortgage payments, and poorer individuals systematically make smaller sized monthly mortgage payments. If such biases existed than this could cause selection bias and thus impact the interpretation of our results. In this section we provide evidence based on institutional details of the Canadian mortgage system that this is not the case. Institutionally, borrowers in Canada have significant latitude in choosing the size of their monthly mortgage payments, given the total size of the mortgage. Indeed, one of the most basic financial decisions made by a mortgage payer is the tradeoff between paying down the mortgage rapidly and acquiring other financial assets (or increasing current expenditure). On the one hand, a borrower that chooses to make a series of small monthly mortgage payments over a long period can acquire relatively more non-housing financial assets, or enjoy more consumption, during the period of the mortgage. On the other hand a borrower who rapidly pays down the mortgage with large monthly mortgage payments, will pay down her mortgage sooner, but will only be able to acquire relatively more non-housing assets or consumption once the mortgage has been paid off. Evidence showing that extent to which individuals can choose the magnitude of the monthly mortgage payment is provided by Breslaw, Irvine and Rahman (1996) who study the demand for mortgages in Canada and conclude that our results indicate that borrowers leave themselves considerable latitude in the choice of the term and amortization of their mortgage. P.284. Further evidence on the wide variance of the choice of size of monthly mortgage payment across individuals is provided by the Canadian Association of Accredited Mortgage Professionals (CAAMP, 2010). They 22

24 provide evidence that at the outset of new mortgage contracts, 64% had amortization periods of 25 years or less and 36% had extended amortization periods. In addition, borrowers have considerable flexibility in changing the amortization period during the course of the mortgage. CAAMP, (2010) provide evidence that in the previous year, 16% of mortgage borrowers have increased their monthly payments, 13% of mortgage borrowers made lump sum payments, 5 % of mortgage borrowers did both It is possible to provide further evidence, using our own data, on whether richer (poorer) individuals systematically select larger (smaller) monthly mortgage payments. Our full monthly bank account data provides us with the size of monthly mortgage payments across approximately individuals (of which only 4000 are in our control group). We can examine the correlation between the size of monthly mortgage payments and income levels based on post code level census data. The correlations between size of mortgage payments and various postal code level measures of income are very low (average income in postal code -0.02, average family income in postal code and employment rate in postal code -0.01). These low correlations indicate that individuals do not seem to systematically select into high or low mortgage payments, based on variables such as income. A5. Canadian Banking Institutions 2: Universal Retail Banking and Alternative Credit Cards. In common with many other papers in the literature using monthly Credit Card account data, in this paper, we can only observe data from a single bank, thus if consumers have credit cards from multiple banks, it is possible that some credit card expenditure could be happening from bank accounts that we do not observe. We argue, however, that because of the universal retail banking system in Canada, a Canadian bank customer is less likely than a US bank customer to acquire credit (credit card, mortgage etc) from multiple banks. The Canadian retail banking system is dominated by five very large banks, who act as universal banks (Ratnovski & Huang, 2009). A single bank will thus tend to provide a consumer with a large number of different financial products. Ratnovski and Huang (2009) argue that because of the dominance of the large universal banks, contestability (and thus entry of new competitors) is low and that limited external competition reduces pressures to defend or expand market share in retail financial products. Evidence on the universal nature of the Canadian retail banking can be seen, for example, from data from the Canadian Bankers Association, based on a survey by the Boston Consulting Group. This data indicates that the average Canadian household has only 2 credit cards, compared to the 6 credit cards held by the average US consumer (Boston Consulting Group, 2009). Similar data has been provided to the author from a confidential Bank of Canada survey, which shows that the average Canadian household had 1.23 bank issued credit cards and 2.51 total credit cards (including store cards etc.) The second reason why the individuals in or sample are less likely to utilize a credit card that we cannot observe in our data relates to the fact that all the individuals in our data have a mortgage at the bank whose data we can observe. Because of issues 23

25 related to relationship lending, individuals may have an incentive to use the credit card issued by their mortgage providing bank. The large literature on relationship lending has argued that the closer the relationship between a bank and a client (e.g. through the use of multiple types of financial products such as mortgages and credit cards) the greater the benefit to the client, because of the reduced information asymmetries that a closer relationship entails. APPENDIX TABLE 1: Treatment vs. Control Groups: Post Code Level Census Data: All Data taken from Post Code level (Dissemination Area) Canadian Census Data. Post Codes for individuals in Treatment and Control Groups taken from Bank Account Data More details on Data Matching in Web Appendix A3. All data measure average measure in each Postal Code Control Group Treatment Group Variable Units Obs Mean Std. Dev. Obs Mean Std. Dev. Age 0-19 % Age % Age % Age % Age % Age Over 65 % No High School % High School % Trades % College % University % Bachelor or Higher % Employment Rate % Average Income $ Median Income $ Employment Income /Total % Income Average Income Family $ Median Income Family $ Additional Web Appendix References Barrington-Leigh, C., & Helliwell, J., (2008) Empathy and Emulation: Life Satisfaction and the Urban Geography of Comparison Groups NBER Working paper No Breslaw, J., Irvine I., & Rahman, A. (1996) Instrument Choice: The Demand for Mortgages in Canada, Journal of Urban Economics, 39,