Welfare Wednesday and the Distribution of Hospital Discharges

There are two types of overdose cases that we do not observe in our data: a fatal over- doses, and b nonfatal overdoses that do not result in a hospital admission. 15 One of the current views—with considerable empirical support—in the literature on drug overdoses is that most fatal drug overdoses could be prevented because fellow users are unlikely to call for an ambulance Warner-Smith et al. 2001. For the current analysis, unobserved overdoses are problematic if the distribution of overdoses that result in a hospital admis- sion differs from the distribution of unobserved overdoses. Overall, we believe this is unlikely; however, one possibility is that the spike in overdose cases in Figure 1 is due to a greater police presence in the Downtown Eastside during the welfare week and thus a higher probability of the police finding an individual who has overdosed and calling paramedics resulting in a hospital admission that would not normally occur. However, the argument could go the other way—with a greater police presence in what is a very small geographic area, users may be even less likely to call for an ambulance. There appears to be no previous evidence on the link between welfare day and either drug overdose hospital admissions or fatal drug overdoses. Phillips, Christenfeld, and Ryan 1999 examine U.S. death certificates from 1973 to 1988 and find an overall increase of 1 percent in the number of deaths in the first week of the month the U.S. wel- fare week relative to the last week of the previous month, and a 14 percent increase in substance abuse-related deaths which includes a wide variety of causes of death. From Canada, there is only evidence from 1993 for British Columbia where Verheul, Singer, and Christenson 1997 find a 50 percent increase in coroner-reported deaths on welfare day. The authors also find increases in detox center admissions and 911 calls on welfare day.

VI. Welfare Wednesday and the Distribution of Hospital Discharges

In this section, we examine whether Welfare Wednesday induces IDUs to leave the hospital against medical advice—interrupting their treatment. Figure 2 shows the distribution of discharges AMA discharges relative to planned discharges, which together equal total discharges over the sample period. A clear spike is observed on the day welfare checks are released. We see an increase on the Tuesday and Thursday as well, but clearly the Wednesday effect dominates. The distribution of planned discharges is very different on the weekends. In fact, the AMA rate is as high on the Welfare Saturday and Welfare Sunday as it is on Wednesdays, but this is entirely due to substantially fewer planned discharges on the weekend. The fall in planned dis- charges is due to two factors: a there are fewer staff working on the weekend who are involved in discharge planning for this population such as social workers and com- munity liaison nurses, and b IDUs require considerable community support home- care nursing, pharmacies for methadone that may be unavailable on the weekend. 15. The extent of these unobserved overdoses may be very large, but it is very difficult to estimate the number of nonfatal, nonhospital admission overdoses per year. In 1998, there were 300 fatal overdoses in British Columbia. In our data there were 65 overdose admissions in 1998. This is only overdose admissions for St. Paul’s Hospital, but, as noted, the latter treats a majority of the drug users in the province. Based on the overdose liter- ature see Section III, in a typical cross-section of IDUs about 20 percent report overdosing nonfatally in the preceding 12 months. Recall that it is believed that there are approximately 10,000 IDUs in Vancouver alone. Riddell and Riddell 149 The Journal of Human Resources 150 20 40 60 80 100 120 140 160 M onday before Tuesday before Wednesday before Thursday before F riday before Saturday before Sunday before Welfare M onday Welfare Tuesday Welfare Wednesday Welfare Thursday Welfare F riday Welfare Saturday Welfare Sunday M onday after Tuesday after Wednesday after Thursday after F riday after Saturday after Sunday after P la n n ed A M A s Figur e 2 Distrib ution of Disc har g es, FY1996–2000 We now further examine the pattern in Figure 2, and exploit the homeless variable to test for social assistance receipt. To do so, we estimate: 2 AMA it = f β + β 1 WW it + β 2 WW it HOMELESS it + β 3 X it + β 4 HOSPITAL it + β 5 DAY it + β 6 WEEK it + β 7 MONTH it + β 8 YEAR it + ε it where AMA equals one if the admission involved the patient leaving AMA, zero if a planned discharge; WW is a dummy variable indicating a check arrival day, zero if any other day; HOMELESS equals one if the patient reported no fixed address; X is the same vector of demographics used in Equation 1; HOSPITAL is a set of medical- related controls including the hospital ward where the individual was treated, and a set of dummies for the nature of the illness associated with the admission 16 ; DAY, WEEK, MONTH, YEAR are the same sets of day-of-the-week, week-of-the-month, calendar-month and fiscal year dummies as in Equation 1; and ε is an error. 17 We also estimate Equation 2 using the “welfare week” instead of the Wednesday the Monday through Sunday of the welfare day. We might expect check day to induce some people to leave on the Tuesday or the Thursday through Sunday of welfare week, although Figure 2 shows that virtually all of the action is on the Wednesday. Table 3 presents the results. With the day-of-the-week dummies, the results in the first column indicate that there is nearly a 16 percentage point increase in the likeli- hood of an AMA on Welfare Wednesday relative to any other Wednesday, and a six percentage point increase 0.16 −0.10 relative to Sunday the omitted day-of-the- week. When estimated using the welfare week, the estimated marginal effect on the welfare week dummy is 0.07. Our AMA check effect results are qualitatively similar to the findings of Anis et al. 2002. The authors estimate a regression similar to Equation 2 using similar data to us from the same hospital; however, they only have access to HIV positive patients IDUs and nondrug users. 18 Leaving AMA and income assistance receipt are rare among the non-IDU HIV-positive population, and so their choice of sample is 16. The diagnosis variables listed in Table 1 indicate the principal diagnosis and are used, along with the length of stay variable, to control for the severity of the illness. The diagnoses are categorized using the World Heath Organization”s International Classification of Diseases 9th Revision, Clinical Modification ICD-9, which are used to classify illnesses for morbidity and mortality data, indexing medical records, med- ical care review and so forth. The codes were aggregated in appropriate cases into higher order diagnoses; in particular, the diagnoses selected account for the most responsible diagnoses for this patient population. The omitted category is “other” diagnosis. 17. As with Equation 1, we use a simple probit model to estimate Equation 2. While there may be unob- served factors such as drug type cocaine users are more likely to leave AMA since no analogous medica- tion to methadone exists for cocaine, but users combine drugs frequently so its importance is unclear that are in the error term we can think of no convincing story for an unobserved factor that also is correlated with the explanatory variables. In any event, for robustness purposes, we estimate Equation 2 using random and fixed-effects estimators as well. The disadvantage of the latter is that we lose observations for those who only had one admission as well as those that had no variation in the AMA variable over time. For brevity, the ran- dom and fixed-effects estimates are not presented, but qualitatively—and quantitatively in cases where we can compare marginal effects such as with the random effects model—the results are unchanged. 18. The authors also restrict the sample to the unit of analysis being the individual that is, using the 1997–99 period, they choose the first admission seen in the data for a given individual, and so end up with only 448 IDU admissions, and only 36 Welfare Wednesday admissions and do not exploit the variation of where check day falls within the month. Riddell and Riddell 151 Table 3 Estimates of the Change in Probability of Leaving the Hospital Against Medical Advice Variable Specification [1] [2] Welfare Wednesday 0.157 — 0.045 Welfare Wednesday homeless −0.122 — 0.059 Welfare week — 0.069 0.024 Welfare week homeless — −0.056 0.037 Homeless 0.105 0.115 0.024 0.026 Downtown Eastside postal code 0.093 0.093 0.022 0.022 Other downtown Vancouver postal code 0.044 0.045 0.022 0.022 Female 0.015 0.016 0.014 0.014 HIV Positive 0.032 0.034 0.015 0.015 Monday −0.072 −0.074 0.024 0.024 Tuesday −0.088 −0.089 0.024 0.024 Wednesday −0.105 −0.077 0.024 0.024 Thursday −0.062 −0.063 0.025 0.025 Friday −0.124 −0.126 0.022 0.022 Saturday 0.043 0.041 0.035 0.034 Last week of the month 0.008 −0.016 0.018 0.023 Third week of the month −0.044 −0.052 0.018 0.018 Second week of the month −0.038 −0.028 0.016 0.021 Log likelihood −2567.2 −2557.3 χ 2 365.3 385.2 Number of observations 4,760 Notes: Huber-White standard errors are in parentheses. Statistical significance is denoted by for 1 per- cent level, for 5 percent level, and for 10 percent level. The dependent variable equals one if the indi- vidual left the hospital against medical advice, zero if on a planned discharge, and has a mean of 0.264. All regressions also include controls for: age and its square, primary diagnosis of illness, hospital ward treated on, length of stay in hospital and its square, 11 calendar-month, and four year dummies. All regressions are estimated by probit. All estimates are presented as marginal effects, and are evaluated at the mean of the rel- evant covariate. The source is a census of hospital admissions of injection drug-users admitted over fiscal years 1996 to 2000 at St. Paul’s Hospital in Vancouver. somewhat peculiar. 19 They find a positive correlation between welfare day and the AMA rate, but give no sense—either in levels as in Figure 2 or as a marginal effect— of the magnitude of their result. The day dummies indicate that AMAs are more likely to occur on the weekend. A hypothesis test on the equality of the six day dummies is rejected at the one percent level χ 2 = 46.9, and an equality test for the five weekday dummies can be rejected as well χ 2 = 13.1. However, as noted, the difference between the weekday dummies and the weekend there is no statistical difference between Saturday and Sunday is attributa- ble to the nature of hospital discharge planning. AMAs are evaluated relative to planned discharges, which fall on the weekend because of reduced staff and community-based clo- sures. AMAs do not respond in a similar way on the weekend since it is believed that the decision to leave AMA has very little to do with the hospital staff or community support. Ceteris paribus, the homeless are more likely to leave the hospital AMA, and the interaction term suggests that there is only a very small check effect for the homeless—as anticipated given that welfare receipt should be quite low for this group. 20 HIV positive IDUs are more likely to leave AMA. AMAs are more likely to occur in the last week of the month and the first week of the month the omitted cat- egory relative to the two middle weeks an equality test on the two middle weeks can- not be rejected; χ 2 = 0.07.

VII. Welfare Wednesday as an Environmental Cue