First. since the regression results include time fixed effects. such autocorrelations could not ccur from events that happen at regular intervals during the year -e.g, from summer vacations. We separately verify that, in any case, fee payments are not seasonal(although spending certainly is). Second, the presence of highly negative autocorrelations at a monthly level would rule out events that last more than one month. For example, a personal crisis that raised the opportunity cost of time for two months would create a positive autocorrelation in time needs and fee payments over the two months, not a negative one. Third, the time-varying needs would have to produce a higher than average fee payment in one month followed by a lower than average fee payment in the following month. This would rule out episodes of high opportunity cost of time for one month followed by a return to the status quo For most plausible processes, needs are likely to be positively autocorrelated. For exam- pIle, the available evidence implies that income processes are positively autocorrelated(e. g Guvenen(2007). While we cannot rule out the"negatively autocorrelated needs"story, existing microeconomic evidence suggests it is highly unlikely to be the right explanation for the empirical patterns that we observe. We conclude that the finding of 0 <0 in(6) is most plausibly explained by a recency effect -consumers become temporarily vigilant about fe voidance immediately after paying a fee 3.6 Medical expenses Negative autocorrelation in fee payments could also be induced by a one-time medical emergency that is not repeated in subsequent months. In our sample, less than 3% of account-holder spending is in medical-related categories. Moreover, spending in such cate- gories does not increase during periods in which a fee is paid 4 Extensions In this section, we present ancillary evidence that extends our analysis of learning and backsliding. Teasing out some of the determinants of learning is challenging, since we do not bserve many of the underlying factors that influence learning dynamics. For example, we tre positively autocorrelated
First, since the regression results include time fixed effects, such autocorrelations could not occur from events that happen at regular intervals during the year – e.g., from summer vacations. We separately verify that, in any case, fee payments are not seasonal (although spending certainly is). Second, the presence of highly negative autocorrelations at a monthly level would rule out events that last more than one month. For example, a personal crisis that raised the opportunity cost of time for two months would create a positive autocorrelation in time needs and fee payments over the two months, not a negative one. Third, the time-varying needs would have to produce a higher than average fee payment in one month followed by a lower than average fee payment in the following month. This would rule out episodes of high opportunity cost of time for one month followed by a return to the status quo. For most plausible processes, needs are likely to be positively autocorrelated. For example, the available evidence implies that income processes are positively autocorrelated ( e.g. Guvenen (2007)). While we cannot rule out the “negatively autocorrelated needs” story, existing microeconomic evidence suggests it is highly unlikely to be the right explanation for the empirical patterns that we observe. We conclude that the finding of 0 in (6) is most plausibly explained by a recency effect — consumers become temporarily vigilant about fee avoidance immediately after paying a fee. 3.6 Medical expenses Negative autocorrelation in fee payments could also be induced by a one-time medical emergency that is not repeated in subsequent months. In our sample, less than 3% of account-holder spending is in medical-related categories. Moreover, spending in such categories does not increase during periods in which a fee is paid. 4 Extensions In this section, we present ancillary evidence that extends our analysis of learning and backsliding. Teasing out some of the determinants of learning is challenging, since we do not observe many of the underlying factors that influence learning dynamics. For example, we are positively autocorrelated. 16
do not see the printed format of the bill that was used during our sample period, and can not tell how salient the fees were(although we suspect that the credit card company likely did not go out of its way to call attention to them Below, we evaluate several factors that we think might influence the rate of learning and backsliding in fee payment, including differences across demographic groups 4.1 Late payments and timeliness of subsequent payments One possible learning mechanism is variation in salience: paying a fee brings the existence of that fee to the account holder's attention, leading her to change subsequent behavior to reduce the frequency of fee payment. We do not directly observe many of the factors that contribute to fee salience-for example, how many times or how carefully the account holder reviews her monthly statement. However, one consequence of a shift in salience of the late payment fee is that the account holder will begin to pay on an earlier day in the billing cycle Moreover, we hypothesize that the shift to earlier payments will be particularly pronounced in the cycle immediately after a late fee payment has been incurred Figure 5 presents a histogram of the days before(negative numbers) and days after (positive numbers) the due date the bill is paid for consumers with either 1 month or 36 months of account tenure. As expected, the distribution shifts to the left over account tenure-reflecting the net reduction in late fee payment observed over tenure. The mass of the distribution largely lies within a two-week period before the due date and a one-week period afterwards Figure 6 presents histograms of the days before (negative numbers) or after(positive numbers)the due date the bill is paid for consumers with account tenures of 1-12 months or 25-36 months. for the billing cycle following the incursion of a late fee. The distribution of dates for high-tenure account holders has shifted to the left -about 42 percent of high-tenure account holders pay their bills more than two weeks early, while only about 27 percent of low-tenure account holders do so-a 15 percentage point increase. The daily frequency of fee payment within two weeks of either side of the due date do not show similarly-sized changes. For both account tenures, the frequency of payment more than two weeks early is much higher than the frequency of payment for that range for all account holders of those
do not see the printed format of the bill that was used during our sample period, and can not tell how salient the fees were (although we suspect that the credit card company likely did not go out of its way to call attention to them). Below, we evaluate several factors that we think might influence the rate of learning and backsliding in fee payment, including differences across demographic groups. 4.1 Late payments and timeliness of subsequent payments One possible learning mechanism is variation in salience: paying a fee brings the existence of that fee to the account holder’s attention, leading her to change subsequent behavior to reduce the frequency of fee payment. We do not directly observe many of the factors that contribute to fee salience—for example, how many times or how carefully the account holder reviews her monthly statement. However, one consequence of a shift in salience of the late payment fee is that the account holder will begin to pay on an earlier day in the billing cycle. Moreover, we hypothesize that the shift to earlier payments will be particularly pronounced in the cycle immediately after a late fee payment has been incurred. Figure 5 presents a histogram of the days before (negative numbers) and days after (positive numbers) the due date the bill is paid for consumers with either 1 month or 36 months of account tenure. As expected, the distribution shifts to the left over account tenure—reflecting the net reduction in late fee payment observed over tenure. The mass of the distribution largely lies within a two-week period before the due date and a one-week period afterwards. Figure 6 presents histograms of the days before (negative numbers) or after (positive numbers) the due date the bill is paid for consumers with account tenures of 1-12 months or 25-36 months. for the billing cycle following the incursion of a late fee. The distribution of dates for high-tenure account holders has shifted to the left—about 42 percent of high-tenure account holders pay their bills more than two weeks early, while only about 27 percent of low-tenure account holders do so—a 15 percentage point increase. The daily frequency of fee payment within two weeks of either side of the due date do not show similarly-sized changes. For both account tenures, the frequency of payment more than two weeks early is much higher than the frequency of payment for that range for all account holders of those 17