The effect of tax rebates on consumption expenditures: evidence from state tax rebates.

In most of the states, rebate amounts were functions of either adjusted gross income or taxable income in previous years. Thus, imputing rebate amounts requires a measure of income earned for those individuals who lived in that state one or two years prior to the rebate year. Unfortunately, the Consumer Expenditure Survey does not contain data on income or location that far back, so some assumptions must be made in order to make this imputation. I first assume that all individuals who report living in one of the four states in the interview year have not moved in the past two years, and are, thus, eligible to receive a rebate check. Second, since the CEX contains data on income in the previous 12 months as of the second and fifth interviews, I create a measure of income in the year relevant to the rebate calculation by taking the earliest report of income and discounting it using the CPI-U price index. (18) The Minnesota and Oregon rebates were based on the amount of taxable income in a previous year. To translate the income variables reported in the CEX to taxable income, I assume that all married couples file jointly, all single individuals file single, and all individuals claim the standard deduction and exemptions available in that particular state and year. In Wisconsin, the rebate amounts were based on adjusted gross income, so I assume that respondents in the CEX reported their AGI as their income before taxes. Finally, using the state rebate formulae, rebate amounts were calculated for each household from these states.

In Connecticut, on the other hand, rebates in 1998 were distributed to anyone who had filed an income tax return or had paid property taxes on a residence or motor vehicle, and the amount of the rebate was fixed subject only to the requirement that the rebate check could not exceed the individuals' tax liability after taking the property tax credit in the previous year. To impute these rebate amounts, I use data from the respondent's base interview on home and car ownership, and assume that only households that reported ownership of either of these received a rebate check. In addition, due to the lack of information on tax liability in the previous year, I assume the maximum amount was received. (19)

Since some consumer units may contain more than one tax unit, I use the CEX member--characteristics files to calculate rebate amounts separately for the head (and wife if applicable) of the consumer unit, and for all other individuals residing in the consumer unit. (20) I then sum these amounts within the consumer unit to arrive at a total amount of imputed rebates received.

The benefit of using these imputed rebate amounts in the regressions is that they account for the different rebate amounts that households received, both across and within states. However, although these imputations should work well for those individuals with steady income and whose geographic location is steady, they will probably perform poorly for those observations with highly variable income or those who have moved.

Although the CEX cannot be used to gauge the appropriateness of these assumptions, other datasets can shed some light on this issue. In Table 3, using data from the Panel Study of Income Dynamics (PSID), I examine the extent to which residents of the states and years under analysis exhibited steady state of residence and income amounts. (21)

As can be seen in the left panel of this table, among those who were in one of the rebate--receiving states in a rebate-receiving year, over 90 percent of respondents were in that state in the prior year on which the rebate is based. In the right panel, the correlation between discounted current income and income in the year on which the rebate was based is also generally high, exceeding 80 percent in all but three of the rebate state--year pairs.

Thus, it appears that the assumptions made in the calculation of the rebate amounts are not bad approximations to reality. Nevertheless, if they are wrong, then the rebate amounts may suffer from nonclassical measurement error, with an unknown bias resulting. To provide a check on the robustness of the results using imputed rebate amounts, I also perform the regressions using dummy variables for rebate receipt as regressors. This variable likely suffers from less measurement error, as the imputation of this variable depends primarily on the location of the respondent. (22) However, using dummy variables to characterize rebate receipt has a downside in that rebate checks differed greatly in amount both across states and within states. (23) As a result, treating such disparate policies identically in the estimation equations could result in magnified standard errors on the rebate dummy variable coefficients.

Finally, to examine further who responds to the receipt of rebate checks, I run specifications that include only those who might be credit constrained. To do this, I include individuals with low asset/ income ratios by dividing a respondent's earliest observation of the total value of the balance in their savings and checking accounts, U.S. savings bonds, and the value of all stocks, bonds, mutual funds, and other securities by their earliest report of income, and cut the sample according to the magnitude of this variable. I also run regression separately by the marital status of the head of the consumer unit.

Summary statistics are presented in Table 4. The base sample of individuals from the four rebate-receiving states includes 6,316 respondent-quarter pairs. On average, the quarterly change in all measures of expenditures is approximately zero, as one would expect. However, there is substantial variation around this mean. A rebate check was received in 8.2 percent of the observations, and the mean rebate check amount among those who received one was $372. In Table 5, the characteristics of these rebates are presented by state and year of receipt. Overall, the sample contains 518 households that received a rebate check in some quarter.

Estimation Method

If the rebate policies implemented in these four states induced their residents to increase their expenditures in the quarter in which the rebate check was received, then their expenditures should increase more (or decrease less) relative to the previous quarter than those of similar individuals in states that did not receive the rebate check.

One way to identify the effects of the rebate program, then, would be to compare the individuals in the four rebate-receiving states with individuals in other states. In essence, one could do a difference in differences estimation, letting individuals in the four states be the treatment group and individuals in other states be the control group. In order for this to be valid, however, one would have to make the assumption that, absent the tax rebate, the change in consumption in the rebate-receiving states would have been the same as that in the non-receiving states. This assumption would be problematic if the reason that these states issued rebate checks was that they had received some positive income shock, which would presumably also affect consumption directly. Since this story seems plausible, individuals in other states might not be a good control group for those in rebate-receiving states.

Instead of this identification strategy, I exploit variation in the timing of rebates across rebate-receiving states to identify the effects of these rebates on expenditures. Essentially, for a rebate in a given state-quarter, individuals in the other states that at some point received a rebate, but did not receive one in that particular quarter, act as a control group for those who did receive the rebate in that quarter. For example, a rebate was distributed in Oregon during November of 1995. In the estimation, then, respondents from quarters containing November of 1995 living in Connecticut, Minnesota, or Wisconsin (who did not receive a rebate in that quarter) serve as a control group for those in Oregon (who did receive a rebate).

In order for this identification strategy to be valid, I have to assume that the change in consumption in a rebate-receiving state-quarter, absent the rebate, would have been the same as in the other rebate receiving states. This assumption seems much more plausible than assuming that consumption would be the same as in the non-'rebate states, but might still be a problem if rebates were passed quickly in response to positive shocks in income. In this case, the timing of the rebate would be correlated with the timing of the positive income shock, and so it may be the income shock to which the individuals are responding instead of the rebate. However, legislative delays diminish this concern, since the initial discussions of tax rebates usually occurred several months before any law was passed or any check was sent out. Nevertheless, as a robustness check, I also perform regressions using individuals in all other states as controls for the individuals in rebate-receiving states.

In the base specification, similar to Souleles (1999, 2002) and Johnson et al. (2004), I estimate equations of the form

[1] [[DELTA]C.sub.i,t] = [[alpha].sub.0] + [[alpha].sub.r][rebate.sub.ist] + [summation over (t)][[alpha].sub.t][d.sub.t] + [summation over(s)][[alpha].sub.s][d.sub.s] + [a.sub.z][Z.sub.i] + [[epsilon].sub.ist],