The effect of the stimulus payments on total visits

utes a disproportionate share of measurement error, since patients with missing SSN s are either not affected by the stimulus payments for instance, if they are undocu- mented immigrants or are randomly assigned to other mailing dates. We thus drop that SSN group from the analysis. 8 The remaining rows of Table 2 demonstrate that the other SSN groups have similar characteristics. That comparison is reassuring, because the SSN groups are randomly assigned.

III. The Effect of the Stimulus Payments on Hospital Utilization

This section presents our main empirical results. We fi rst demonstrate the effect of the stimulus payments on ED visits and hospitalizations. We then investi- gate the mechanisms involved. To do so, we decompose visits by characteristics of the patients and their medical conditions.

A. The effect of the stimulus payments on total visits

In principle, the impact of the stimulus payments could arise and fade at any delay after payment receipt. We seek to be fully agnostic about these dynamics, and we use mul- tiple approaches to study the effects. First, we run a standard difference- in- difference specifi cation to measure the average effect of the stimulus payments over the followup observation period. We aggregate the data to counts of visits by SSN group and week, Y gt , and estimate logY gt = β 1 · I{CheckSent} gt + β 2 · I{DirectDepositSent} gt + α t + γ g + ε gt This regression includes a fi xed effect for each week, α t , and a fi xed effect for each SSN group, γ g . The indicator functions, I{CheckSent} gt and I{DirectDepositSent} gt , indicate, respectively, whether checks were mailed and whether the direct deposits were made to group g by time t. We thus interpret the point estimates as the percentage change in utilization for groups that have received their stimulus payments relative to groups that have not yet received their payments. The fi xed effects control for season- ality in hospital utilization and variation driven by differences in the size of the groups. Table 3 reports estimates of this specifi cation. Each cell of the table presents an estimate of β 1 when the logarithm of ED visits, inpatient visits, or all visits is the dependent variable. 9 The fi rst column demonstrates that after the stimulus payments are mailed, total ED visits increase by 1.1 percent p- value of 0.036, over a baseline average of 95,076 visits per week. Inpatient visits increase by less than 1 percent, a change that is not statistically signifi cant at the 5 percent level. ED and inpatient visits combined increase by 0.9 percent, over a baseline of 141,982 visits per week, implying an increase of nearly 1,200 visits per week on average for 23 weeks after the checks 8. Online Appendix Table 3 presents our main results with this fi rst SSN group added back in. The results are generally not sensitive to the exclusion of this group. 9. In all specifi cations throughout the paper, estimates of β 2 are statistically insignifi cant so we do not report them. There may be less statistical power to detect an effect of the direct deposit payments because the deposits were made over only three weeks and they may have been less salient to recipients. were sent. The remaining columns of Table 3 present estimates of β 1 separately for visits by adult men and adult women. Both genders experienced a roughly 1 percent increase in ED visits; both estimates are statistically signifi cant at the 10 percent level. Younger patients are not matched to the SSN group of their parents. Reassuringly, we fi nd no statistically signifi cant change in visits for children p- value of 0.41. For all remaining estimates, we focus solely on visits by adults. These difference- in- difference estimates assume that the stimulus payments have a constant, persistent effect on hospital visits. The treatment effect, however, may not be constant, for instance decaying as time passes. Our second empirical approach attempts to measure the dynamics of the response to the stimulus payments. We estimate distributed- lag specifi cations, by replacing I {CheckSent} gt in the regression equation above with a series of indicator variables that are equal to one if the hospital visit occurred 1–2 weeks before rebate receipt, the week of rebate receipt or 1 week after, 2–3 weeks after rebate receipt, and so on. 10 10. All main text fi gures report results when estimating two leads and six lags. Robustness is demonstrated in Online Appendix Figures 1–2. Table 3 The Effect of the Stimulus Payments on Hospital Visits All Adult Visits 1 Men 2 Women 3 A. Dependent Variable: Logarithm of ED Visits After Check Sent 0.011 0.011 0.010 0.004 0.005 0.005 [0.036] [0.073] [0.093] Average visits week 95,076 40,822 54,250 B. Dependent Variable: Logarithm of Inpatient Visits After Check Sent 0.006 0.005 0.006 0.004 0.009 0.006 [0.183] [0.636] [0.294] Average visits week 46,906 18,669 28,236 C. Dependent Variable: Logarithm of All Visits After Check Sent 0.009 0.009 0.009 0.003 0.004 0.005 [0.024] [0.072] [0.119] Average visits week 141,982 59,491 82,486 Notes: This table reports estimates from difference- in- difference regressions. In each case the sample con- sists of counts of California hospital visits by SSN group and week, covering 19 weeks before and 23 weeks after the rebates were sent. Full sets of SSN group fi xed effects, week fi xed effects, and an indicator for whether direct deposits have been made are also included in the regressions. N = 9 × 1 + 19 + 23 = 387. The standard errors in parentheses adjust for correlation between observations from the same SSN group. Associated p- values in brackets. Figure 1 presents the point estimates from this regression, for all visits and sepa- rately by gender. In each panel, the solid line plots the point estimates, whereas the dashed lines plot 95- percent confi dence intervals. The omitted lag in each regression is the period immediately prior to the week in which the stimulus payments were sent. In all panels, the probability of an ED visit becomes positive and statistically sig- nifi cant within fi ve weeks after the stimulus payments are sent. Some delay in the impact may be caused by the time required for households to receive and cash the stimulus checks. Alternatively, the payments may alter families’ monthly budgets, and the surplus may only become salient at the end of the month. For all visits, we observe a statistically signifi cant 2 percent increase in emergency visits in Weeks 4 and 5 after the rebates. The modest pretrends discernible here are consistent with anticipation of stimulus payments by some households. For men, the fi gures surprisingly suggest a permanent effect of the payments on ED visits. But the confi dence intervals after the fi rst month are wide. We view such long- term estimates as speculative because we possess no true control group after all groups receive their checks. 11 The estimates above rely on a proxy for when individuals would have been sent their stimulus checks if they received a payment based on their own SSN . But the actual number of individuals treated by being sent a check differs from the number identi- fi ed by our research design. The regressions above capture the intent- to- treat effect of the payments on health care consumption. If we observed actual payment receipt, we could scale the intent- to- treat effect by the share of individuals who received a payment, to estimate the treatment effect on the treated. 12 Relative to the treatment effect on the treated, the intent- to- treat effect is scaled toward zero by the probability of actual check receipt. Several considerations affect this scaling factor. First, Parker et al. 2011 report that roughly 85 percent of households received a stimulus payment and 60 percent of households received the payments via paper check. Second, in married households receiving stimulus payments, the date the check was sent was determined by the fi rst SSN listed on the joint income tax return IRS 2008. Either spouse could be listed fi rst on a joint return, and in 2008, 38 percent of returns were joint. This implies that 28 percent of payment recipients and possibly hospital patients would have received their payments based on the schedule for their spouse’s SSN group, which differs from their own 89 percent of the time. 13 If the causal effect of actual check receipt on hospital visits were identical across households, and stimulus payment amount, receipt by paper check, marital status, and ordering of spousal SSN s on the tax return are independent, then we could multiply the reduced- form intent- to- treat estimates by 1 0.85 0.6 0.75 = 2.6 to scale up to the treatment- on- the- treated effect. This scale- up would apply both for the difference- in- difference and distributed- lag specifi cations. Multiplying, our estimate of a 1.1 percent 11. Online Appendix Table 4 reports estimates of an alternative functional form, which allows for expo- nential decay of an initial effect. The point estimates in that table also show a statistically signifi cant initial increase in ED visits of 1.1 percent. The decay rate takes the wrong sign in that specifi cation but is not statisti- cally different from 0. 12. In estimating effects of the stimulus payments using the Consumer Expenditure Survey, Parker et al. 2011 run instrumental variables regressions, instrumenting for payment amount with an indicator for ran- domly assigned payment receipt. 13. In fact, men and women might not have been equally likely to be listed fi rst on their joint income tax document, in which case the reduced- form intent- to- treat coeffi cients could be scaled differently by gender. Figure 1 Distributed- Lag Estimates by Gender Notes: Each fi gure plots point estimates from a regression of log counts of visits on a set of indicators for two- week intervals. The dotted lines plot 95 percent confi dence intervals that are robust to autocorrelation between observations from the same SSN group. SSN group fi xed effects, week fi xed effects, and an indicator for whether direct deposits had been made are also included in the regressions. The omitted time period is one and two weeks before rebate checks were sent. -.01 .01 .02 .03 P oint Estimate -3 -3, -2 -2, -1 0, 1 2, 3 4, 5 6, 7 8, 9 9 Weeks Since Rebate Receipt -.02 .02 .04 P oint Estimate -3 -3, -2 -2, -1 0, 1 2, 3 4, 5 6, 7 8, 9 9 Weeks Since Rebate Receipt Dependent Variable: Logarithm of ED Visits, Men -.02 .02 .04 P oint Estimate -3 -3, -2 -2, -1 0, 1 2, 3 4, 5 6, 7 8, 9 9 Weeks Since Rebate Receipt Dependent Variable: Logarithm of ED Visits, Women effect from the difference- in- difference specifi cations implies that the effect of actu- ally receiving a stimulus payment is 2.87 percent. This number still may not equal the average treatment effect for several reasons. First, paper check recipients differ systematically from the general population. Parker et al. 2011 indicate that direct deposit recipients had higher incomes than paper check recipients, similar family sizes, and slightly larger stimulus payment amounts. Households without suffi cient qualifying income to receive a stimulus payment, and households with suffi cient income to be above the phase- out would likely have had different responses as well. Second, the reduced- form regressions are biased toward zero because of measurement error and because some check recipients may have be- gun to change consumption behavior in anticipation of their checks. It is not clear which of these reasons for differences from the average treatment effect dominate, but they go in offsetting directions. Taken together, our results provide evidence of an increase in hospital utilization caused by modest liquidity shocks. We next investigate the mechanisms for this effect by testing for variation in treatment effects by medical condition and patient charac- teristics.

B. The effect of the stimulus payments by medical condition