Is There a VA Advantage - Station-Specific OLS Estimates

Figure A.8: Station-Specific OLS Estimates of VA Advantage 0 5 Density 10 15 20 A: Stations with at Least 5,000 Rides −.06 −.04 −.02 Station−specific OLS effect 0 .02 0 10 Density 20 30 40 B: All Stations, Shrunken Estimates −.05 −.04 −.03 −.02 Station−specific OLS effect −.01 0 Note: Panel A of this figure shows the kernel density distribution of station-specific OLS estimates of the ℓ VA advantage, or 𝛽ˆ𝑂𝐿𝑆 estimated from Equation (A.26) for rides corresponding to each station. We include estimates from 32 stations with at least 5,000 rides, comprising a sample of 276, 483 rides. Panel B of this figure shows the kernel density distribution of empirical Bayes posteriors of the station-specific OLS estimates ℓ of the VA advantage. These posteriors are given by 𝛽˜𝑂𝐿𝑆 in Equation (A.27). The figure displays posteriors from all 94 stations in our baseline sample in Appendix Table A.1, comprising 401, 319 rides. A.22 40 Percent of veterans 20 30 10 1 3 Number of hospitals 2 A: Number of Hospitals 4+ 0 30 modal non-VA, to non-modal non-VA modal non-VA, to modal non-VA modal non-VA, to VA 20 Percent of veterans modal VA, to non-VA modal VA, to VA 10 B: Systems of Care and Ambulance Transport Figure A.9: Sources of Prior Utilization 40 Note: This figure shows patterns of prior utilization and ambulance transport among a sample of patients who have some prior utilization either at the VA or affiliated with a non-VA hospital. Panel A shows the percentage of patients in this sample who utilize care associated with different numbers of hospitals. Panel B shows ambulance transport patterns to either the VA or a non-VA hospital depending on whether a patient’s modal hospital in prior utilization was associated with the VA or with a non-VA hospital. If the patient’s modal hospital utilization was at a non-VA hospital, the figure also shows the percentage of patients who were transported to their modal non-VA hospital or another non-VA hospital. The sample selection for this group of patients is given in Appendix Table A.13. 0 A.23 Figure A.10: Modal Hospital First Stage, Balance, and Reduced Form .6 Modal hospital .7 .8 .9 A: First Stage .5 coeff. = 0.767 (0.006) N = 1,414,217, J = 5,716 .2 .3 .4 .5 Leave−out ambulance Pr(modal hospital) .6 (Predicted) 28−day mortality .105 .11 .115 .12 B: Balance and Reduced Form .1 red. form coeff. = −0.005 (0.003) balance coeff. = 0.000 (0.001) N = 1,414,217, J = 5,716 .2 .3 .4 .5 Leave−out ambulance Pr(modal hospital) Mortality .6 Predicted mortality Note: Panel A shows a binned scatterplot of arrival at the veteran’s modal hospital against the ambulance leave-out propensity to arrive at that hospital on the x-axis. The figure is a graphical representation of the firststage regression in Equation (10). Panel B shows binned scatterplots of 28-day mortality and predicted 28-day mortality on the y-axis against the ambulance leave-out propensity to arrive at the veteran’s modal hospital on the x-axis. Mortality bin means are shown in solid circles, while predicted mortality bin means are shown in hollow circles. The figure represents the reduced-form regression in Equation (11) and the corresponding balance regression replacing mortality with predicted mortality. The sample includes 1, 414, 217 ambulance rides and 5, 716 combinations of ambulance company identifiers and Dartmouth Atlas Hospital Referral Regions (HRRs). The sample includes patients who have some utilization affiliated with a non-VA hospital and no utilization at the VA in the prior year. The selection details of this sample is given in Appendix Table A.13. Controls include patient zip code dummies, ALS/BLS dummies, source of the ambulance ride, time categories, and patient prior utilization, which are detailed in Appendix Table A.2. A.24 .105 28−day mortality .11 .115 .12 Figure A.11: Modal Hospital Visual IV .1 coeff. = −0.006 (0.004) N = 1,414,217, J = 5,716 .1 .2 .3 Probability of modal hospital .4 .5 Note: This figure shows the visual IV plot corresponding to the IV regression of the effect of arrival at a patient’s modal hospital on 28-day mortality. For each bin of the instrument, which is the ambulance leave-out propensity to arrive at the patient’s modal hospital, we plot the mean 28-day mortality on the y-axis and the probability that the index patient arrives at his modal hospital on the x-axis. Modal hospital arrival predictions correspond to a first-stage regression in Equation (10), and mortality predictions correspond to a reducedform regression in Equation (11). The best-fit line in the visual IV plot replicates the IV estimate of the effect of arrival at a patient’s modal hospital on 28-day mortality, which we perform to obtain the standard error (in parentheses). This IV regression uses 1, 414, 217 observations and 5, 716 combinations of ambulance company identifiers and Dartmouth Atlas Hospital Referral Regions (HRRs). We use the sample of non-VAonly utilizers, given in Appendix Table A.13. Controls include patient zip code dummies, ALS/BLS dummies, source of the ambulance ride, time categories, and patient prior utilization, which are detailed in Appendix Table A.2. A.25 Effect of modal hospital −.03 −.025 −.02 −.015 −.01 −.005 0 .005 Figure A.12: Modal Hospital OLS and IV Specifications baseline controls 1 2 3 hold−out controls 4 5 6 7 Specification OLS 8 9 10 11 12 IV Note: This figure shows the effect of arrival at a patient’s modal hospital on 28-day mortality estimated from OLS and IV specifications, with progressive sets of controls. Numbered incremental controls correspond to categories or subcategories of variables that are presented in order in Appendix Tables A.2 and A.3. Estimates are shown along solid lines, while 95% confidence intervals are shown in dashed lines. All specification control for hospital identities and use the sample of non-VA-only utilizers, given in Appendix Table A.13. A.26 Effect of modal hospital −.03 −.025 −.02 −.015 −.01 −.005 0 .005 Figure A.13: Modal Hospital Combinations of Controls baseline controls 1 2 3 hold−out controls 4 5 6 7 8 Number of controls Mean estimate 9 10 11 12 Min/max estimates Note: This figure shows IV estimates of the effect of arrival at a patient’s modal hospital on mortality on the y-axis, with first-stage and reduced-form Equations (10) and (11), varying the number of controls included in the IV regression. Control variables are detailed in Appendix Tables
A.2 and A.3. All specifications include the five baseline controls. Therefore, the figure represents 5 + 27 − 1 = 132 specifications. For each number of controls 𝑛 for 𝑛 > 5, we consider “7 choose 𝑛 − 5” specifications. The mean IV estimate is shown with a dashed line; the minimum and maximum IV estimates are shown with a short dashed line. We use the sample of non-VA-only utilizers, given in Appendix Table A.13. A.27 A.28 Require ED visit within 24 hours after ambulance ride, non-missing demographic data, enrollment in Medicare Parts A and B for at least one year, date of death (if non-missing) weakly after the ambulance ride. Drop rides linked to more than one ED visit (i.e., visits in different hospitals), with patients younger than 20 years or older than 99 years, with missing Health Referral Region, with missing ambulance diagnosis code, or from VA New Orleans (destroyed in 2005 due to Hurricane Katrina). Drop patients who do not live within 20 miles of a VA hospital and within 20 miles of a non-VA hospital. Drop rides to a hospital over 50 miles from the patient’s home. Drop rides by an ambulance company with fewer than 20 patients in a given zip code. Drop rides by an ambulance company with less than 5% of rides in a given zip code to a VA hospital. Drop rides from zip codes with only one remaining ambulance company. Drop rides for patients with no VA utilization (inpatient, ED, or primary care) in the prior year. Drop rides for patients with a prior ride within 30 days. 1. Build initial sample of ambulance rides to EDs from January 1, 2001, to December 31, 2014. 2. Clean sample 3. Distance restrictions 4. Ambulance restrictions 5. Prior utilization restriction 6. Prior ride restriction 401,319 491,193 1,051,093 3,465,588 8,828,997 8,952,884 Rides 188,299 188,299 365,163 1,118,302 2,862,557 2,898,667 Patients 1,217 1,217 1,217 14,662 180,320 Ambulance companies 183,693 99 99 100 118 125 127 VA 1,386 1,404 1,577 3,071 7,744 7,816 Non-VA Hospitals Note: This table details selection steps to create the baseline sample. At each step, the table lists the number of ambulance rides, patients, ambulance companies, and VA and non-VA hospitals. Table 1 shows average patient characteristics among observations at each sample step. Description Sample step Table A.1: Selection of Baseline Sample Table A.2: Baseline Control Variables Category Location (1,633 indicators) Subcategory Zip code (1,630 indicators) Variables Zip code indicators (1,630 indicators) Pickup source (3 indicators) Indicators for whether pickup is from residence, residential (including domiciliary, custodial facility), skilled nursing facility, or scene of accident (omitted category) Ambulance service (3 indicators) Indicators for whether ambulance is ALS special (CPT codes A0427, A0330, A0370), ALS non-special (CPT codes Q3019, A0368, A0328), ALS level 2 (CPT code A0433), or BLS (omitted category; CPT codes A0429, A0362, A0322) Time categories (173 indicators) Day of the week (6 indicators); Month-year interactions (167 indicators) Prior utilization (6 indicators) Indicators for utilization in prior year of Medicare primary care, VA primary care utilization, Medicare ED, VA ED, Medicare inpatient, and VA inpatient services
Note: This table describes baseline controls variables, denoted as 𝑧 (𝑖) , X0𝑖 in Condition 1 and throughout the text. We consider our quasi-experiment to be conditional on these variables, and we include these variables as controls in all of our analyses. Numbers of non-collinear indicators are given in parentheses. A.29 Table A.3: Hold-Out Control Variables Category Patient background (60 variables) Subcategory Demographics (30 indicators) Variables Age: 5-year age bins from 20-64 years, 2-year age bins from 65-100 years (26 indicators); Male gender; Race: indicators for white, Black, Hispanic, and Asian/other (omitted category) Socioeconomic status, combat history, and eligibility (22 indicators) Terciles of income and net worth (4 indicators); Period of combat: WWII, Korean, Vietnam, other (omitted category) (3 indicators); Indicator for aid and attendance for in-home care; Priority group indicators (7 indicators); Service connection: service connected, not service connected, or non-veteran/other (omitted category) (2 indicators); 6 missing indicators for each of the above characteristics Extended prior utilization (8 variables) Counts of VA primary care visits, outpatient visits, ED visits, and inpatient visits in prior year; Analogous counts of Medicare visits in prior year Prior diagnoses (93 indicators) 31 Elixhauser indicators (dividing hypertension indicator into 2 indicators for complicated and uncomplicated hypertension), in four categories: present in VA data only, present in Medicare data only, and present in both VA and Medicare data (31 × 3 = 93 indicators) 3-digit ambulance diagnosis codes (641 indicators) 3-digit ambulance diagnosis (ICD-9) codes (641 indicators) Co-rider characteristics (33 variables) Co-rider baseline controls (12 variables) Co-rider pickup source proportions (3 variables); Co-rider ambulance service proportions (3 variables); Co-rider prior utilization proportions (6 variables) Co-rider hold-out controls (21 variables) Co-rider average continuous age; Co-rider proportion male gender; Co-rider race proportions (3 variables); Co-rider 1-digit ambulance code proportions (15 variables); Co-rider average predicted mortality Note: This table describes hold-out control variables. These variables are used to test robustness of our findings, particularly in Figures 2, A.2, A.13, and A.12. Numbers of non-collinear indicators or variables are given in parentheses. A.30 A.31 Yes No No No No -2,245 (899) 12,173 401,319 -0.053 (0.019) 0.097 401,319 No Yes No No No No No Yes No No (3) A: Dependent variable: -0.042 -0.045 (0.018) (0.018) 0.097 0.097 401,319 401,319 B: Dependent variable: -2,583 -2,748 (832) (880) 12,173 12,173 401,319 401,319 (2) No No No Yes No No No No No Yes (4) (5) 28-day mortality -0.045 -0.045 (0.018) (0.018) 0.097 0.097 401,319 401,319 28-day spending -2,501 -2,569 (816) (822) 12,173 12,173 401,319 401,319 Yes Yes Yes Yes Yes -2,559 (992) 12,173 401,319 -0.049 (0.020) 0.097 401,319 (6) Note: This table presents IV estimates of the effect of the VA on 28-day mortality (Panel A) and on 28-day spending (Panel B). In each panel, each column involves including a set of controls for ambulance actions on the specific ride (flexible functions of the charges incurred by the ambulance company, flexible functions of the mileage driven by the ambulance company) and for non-VA hospitals chosen by the ambulance company (non-VA hospitals chosen by the ambulance company in rides originating from the same zip code and “out-of-sample” averages of non-VA hospital mortality and spending). “Out-of-sample” refers to averages that are computed using patients outside of the main analytical sample (Appendix Table A.1) because they have no VA utilization in the prior year; specifically, they are computed using patients with only non-VA utilization in the prior year (Panel B of Appendix Table A.13). Regressions are run on the main analytical sample. Further details are given in Appendix A.1.1. Ambulance charges splines Non-VA hospitals in chosen set Mileage splines Out-of-sample non-VA mortality Out-of-sample non-VA spending Outcome mean Observations VA hospital Outcome mean Observations VA hospital (1) Table A.4: Robustness of Exclusion Restriction Table A.5: Monotonicity Tests First stage sample Observations VA share Age ≤ 80 239,611 0.347 Age > 80 161,707 0.305 White 314,064 0.304 Non-white 87,176 0.426 Comorbidity count (high) 167,332 0.292 Comorbidity count (low) 233,987 0.358 Mental illness or substance abuse 188,961 0.354 No mental illness or substance abuse 212,358 0.309 VA visits in prior year (high) 183,087 0.508 VA visits in prior year (low) 218,232 0.181 Advanced Life Support 274,690 0.301 No Advanced Life Support 126,616 0.393 Predicted VA user (high) 200,659 0.543 Predicted VA user (low) 200,660 0.117 Predicted mortality (high) 200,659 0.328 Predicted mortality (low) 200,660 0.333 Instrument ReverseBaseline sample 0.931 0.497 (0.038) (0.022) 0.789 0.456 (0.041) (0.022) 0.821 0.221 (0.037) (0.016) 0.992 0.596 (0.068) (0.041) 0.758 0.427 (0.041) (0.019) 0.938 0.553 (0.039) (0.027) 0.931 0.514 (0.040) (0.024) 0.815 0.456 (0.037) (0.020) 1.038 0.710 (0.050) (0.035) 0.718 0.284 (0.031) (0.014) 0.836 0.249 (0.036) (0.018) 0.840 0.301 (0.048) (0.032) 1.113 0.865 (0.054) (0.055) 0.559 0.218 (0.030) (0.011) 0.835 0.368 (0.036) (0.019) 0.898 0.502 (0.046) (0.024) Dual eligibles Instrument sample Analytical sample Note: This table presents first-stage coefficients on different subsamples of patients. For each subsample, we present results for two different instruments: (i) the baseline leave-out instrument, 𝑍𝑖 , given in Equation (1) and calculated from observations among dually eligible veterans (Step 1 of Appendix Table A.1), and (ii) a reverse-sample instrument, 𝑍˜ 𝑖−𝑚 , given in Equation (A.2) and calculated from observations in the analytical sample (Step 6 of Appendix Table A.1) that are outside of the regression subsample. Each regression uses baseline controls defined in Appendix Table A.2. Further details are given in Appendix A.1.2. A.32 Table A.6: Monotonicity Tests (Continued) First stage sample Age ≤ 80 Observations 239,611 VA share 0.347 Age > 80 161,707 0.305 White 314,064 0.304 Non-white 87,176 0.426 Comorbidity count (high) 167,332 0.292 Comorbidity count (low) 233,987 0.358 Mental illness or substance abuse 188,961 0.354 No mental illness or substance abuse 212,358 0.309 VA visits in prior year (high) 183,087 0.508 VA visits in prior year (low) 218,232 0.181 Advanced Life Support 274,690 0.301 No Advanced Life Support 126,616 0.393 Predicted VA user (high) 200,659 0.543 Predicted VA user (low) 200,660 0.117 Predicted mortality (high) 200,659 0.328 Predicted mortality (low) 200,660 0.333 Instrument Baseline In-sample 0.586 0.525 (0.021) (0.020) 0.494 0.394 (0.023) (0.022) 0.504 0.513 (0.019) (0.020) 0.676 0.440 (0.032) (0.033) 0.493 0.438 (0.020) (0.021) 0.583 0.504 (0.022) (0.020) 0.592 0.518 (0.021) (0.020) 0.501 0.426 (0.020) (0.020) 0.691 0.572 (0.026) (0.021) 0.421 0.445 (0.018) (0.021) 0.523 0.518 (0.020) (0.021) 0.531 0.433 (0.025) (0.024) 0.743 0.619 (0.028) (0.021) 0.331 0.423 (0.016) (0.027) 0.513 0.458 (0.020) (0.019) 0.570 0.479 (0.023) (0.021) Analytical sample Instrument sample Analytical sample Note: This table presents first-stage coefficients on different subsamples of patients. For each subsample, we present results for two different instruments: (i) the baseline leave-out instrument, 𝑍˜ 𝑖 , given in Equation (1), and (ii) a in-sample instrument, 𝑍˜ 𝑖𝑚 , given in Equation (A.2) and calculated from leave-out observations in the same regression subsample. Both instruments are calculated using observations in the analytical sample (Step 6 of Appendix Table A.1). Each regression uses baseline controls defined in Appendix Table A.2. Further details are given in Appendix A.1.2. A.33 Table A.7: Always-Taker and Never-Taker Characteristics Male Age Black Income Rural zip code Residential source Comorbidity count Mental illness Substance abuse Prior VA ED visit Prior Medicare ED visit Ambulance rides in prior year Advanced Life Support Predicted VA user Predicted mortality Always takers Mean Ratio 0.961 1.00 (0.002) [0.99 - 1.00] 75.6 0.99 (0.158) [0.99 - 1.00] 0.222 1.14 (0.012) [1.02 - 1.26] $18,039 0.86 ($200) [0.84 - 0.88] 0.064 1.27 (0.015) [0.67 - 1.87] 0.685 0.97 (0.011) [0.94 - 1.00] 5.85 0.95 (0.046) [0.94 - 0.97] 0.469 1.10 (0.006) [1.07 - 1.13] 0.150 1.04 (0.005) [0.97 - 1.10] 0.823 1.56 (0.004) [1.54 - 1.57] 0.262 0.54 (0.006) [0.52 - 0.57] 2.212 1.03 (0.030) [1.00 - 1.05] 0.576 0.84 (0.013) [0.81 - 0.88] 0.969 1.14 (0.001) [1.14 - 1.15] 0.103 1.07 (0.002) [1.03 - 1.10] Never takers Mean Ratio 0.965 1.00 (0.001) [1.00 - 1.00] 76.3 1.00 (0.153) [1.00 - 1.01] 0.184 0.95 (0.010) [0.85 - 1.05] $22,397 1.07 ($232) [1.05 - 1.09] 0.053 1.04 (0.011) [0.62 - 1.46] 0.667 0.95 (0.009) [0.92 - 0.97] 6.44 1.05 (0.032) [1.04 - 1.06] 0.420 0.98 (0.004) [0.97 - 1.00] 0.137 0.95 (0.004) [0.90 - 1.00] 0.383 0.72 (0.006) [0.70 - 0.75] 0.613 1.27 (0.004) [1.26 - 1.29] 2.210 1.03 (0.025) [1.00 - 1.05] 0.707 1.03 (0.010) [1.01 - 1.06] 0.778 0.92 (0.002) [0.91 - 0.92] 0.100 1.03 (0.001) [1.01 - 1.05] Note: This table presents average characteristics for always takers and never takers. Always takers are defined as patients who present to the VA even when they receive a residualized instrument below the 20th percentile; never takers are defined as patients who present to a non-VA hospital even when they receive a residualized instrument above the 80th percentile. To form these residualized instruments, we residualize the baseline instrument, 𝑍𝑖 , given in Equation (1), by baseline controls, described in Appendix Table A.2. Observations are drawn from the baseline sample described in Appendix Table A.1. For each row corresponding to a characteristic, the table presents average characteristics and the ratio between this average and the overall sample average. Overall sample means are given in Table 4. Standard errors are calculated by bootstrap, blocking observations by zip codes, and are shown in parentheses. Corresponding 95% confidence intervals of the ratio are presented in brackets. Further details are given in Appendix A.3. A.34 A.35 0.097 401,319 Linear (1) -0.043 (0.017) 0.003 (0.001) Cubic 3 0.097 401,319 Cubic 5 0.097 401,319 Gaussian basis 3 0.097 401,319 Gaussian basis 5 0.097 401,319 (5) -0.033 (0.005) 0.012 (0.011) Note: This table presents estimates of the average treatment effect (ATE) from the selection model in Equation (A.21), under different specifications. Column 1 presents results from a control function that is linear in the first-stage residual, corresponding to the regression in Equation (A.23). Columns 2 to 5 present results from semiparametric control functions, corresponding to regressions of the form in Equation (A.25). The columns vary in whether the spline functions are cubic functions or Gaussian basis functions and in the number of knots. In addition to the ATE, each column presents the difference between the ATE and the local average treatment effect (LATE). The LATE is estimated from Equation (A.24) and is numerically equivalent to the LATE from our benchmark analysis in Section 3. We compute standard errors (shown in parentheses) for the ATE and the ATE-LATE difference by bootstrap, blocking by zip codes. Appendix A.4 provides further details. Control function Knots Outcome mean Observations ATE-LATE difference ATE Dependent variable: 28-day mortality (2) (3) (4) -0.033 -0.033 -0.033 (0.006) (0.006) (0.005) 0.013 0.012 0.012 (0.011) (0.011) (0.011) Table A.8: Treatment Effects from Selection Model Table A.9: Heterogeneity by Non-VA Hospital Characteristics Regression estimates VA VA × 𝐼˜𝑥,𝑖 Volume, Size, and Capabilities ED visits Admissions Total staffed beds Teaching hospital Trauma center Advanced cardiac care Stroke center Staffing ED staff per 1,000 ED visits Nurses per 1,000 patient-days Physicians per 1,000 patient-days Hospitalists per 1,000 patient-days Intensivists per 1,000 patient-days Characteristic means 𝐼 𝑥,𝑖 = 0 𝐼 𝑥,𝑖 = 1 -0.046 (0.016) -0.046 (0.017) -0.046 (0.017) -0.045 (0.017) -0.045 (0.016) -0.046 (0.017) -0.045 (0.017) -0.002 (0.002) -0.003 (0.002) -0.004 (0.002) -0.000 (0.002) 0.004 (0.002) -0.000 (0.002) 0.001 (0.002) 28,082 53,849 9,664 17,859 199 375 0.02 0.51 0.28 0.93 0.64 1.00 0.03 0.65 -0.045 (0.017) -0.046 (0.016) -0.045 (0.017) -0.045 (0.017) -0.045 (0.017) -0.001 (0.002) 0.006 (0.002) 0.002 (0.002) 0.003 (0.002) 0.003 (0.002) 0.30 0.75 4.13 6.58 4.36 10.79 0.12 0.39 0.05 0.23 Note: This table presents regression results investigating heterogeneous treatment effects along binary indicators of average non-VA hospital characteristics associated with each zip code. For each zip code, hospital characteristics are averaged with weights proportional to the number of rides going to each non-VA hospital from the zip code. We then divide observations 𝑖, based on whether their zip codes 𝑧 (𝑖) have below- vs. abovemedian averages, denoted by 𝐼 𝑥,𝑖 = 0 and 𝐼 𝑥,𝑖 = 1, respectively. Regression results correspond to Equation (8). The coefficient on VA represents the LATE of going to the VA, and the coefficient on VA × 𝐼˜𝑥,𝑖 represents the difference in the LATE between observations with 𝐼 𝑥,𝑖 = 1 and observations with 𝐼 𝑥,𝑖 = 0. Appendix Table A.10 presents results for additional characteristics. Appendix A.5 provides further details on the hospital characteristics. A.36 Table A.10: Heterogeneity by Non-VA Hospital Characteristics (Continued) Regression estimates VA VA × 𝐼˜𝑥,𝑖 Spending and Outcomes Relative spending Mortality rate Readmission rate Organization and IT Network or hospital system HMO or ACO Health IT Share of non-VA rides (max.) -0.045 (0.017) -0.045 (0.017) -0.045 (0.017) -0.002 (0.002) -0.003 (0.002) -0.002 (0.002) -0.045 (0.017) -0.045 (0.017) -0.046 (0.016) -0.045 (0.017) -0.002 (0.002) -0.002 (0.002) -0.002 (0.002) 0.002 (0.002) Characteristic means 𝐼 𝑥,𝑖 = 0 𝐼 𝑥,𝑖 = 1 0.97 1.04 11.62 12.89 17.30 18.90 0.65 1.00 0.00 0.47 0.00 0.80 0.42 0.73 Note: This table presents regression results investigating heterogeneous treatment effects along binary indicators based on non-VA hospital characteristics associated with each zip code. For “Share of non-VA rides (max.)”, we take the maximum non-VA hospital share of non-VA rides for each zip code as the zip code characteristic. For the remaining characteristics, hospital characteristics are averaged within each zip code with weights proportional to the number of rides going to each non-VA hospital from the zip code. We then divide observations 𝑖, based on whether their zip codes 𝑧 (𝑖) have below- vs. above-median statistics, denoted by 𝐼 𝑥,𝑖 = 0 and 𝐼 𝑥,𝑖 = 1, respectively. Regression results correspond to Equation (8). The coefficient on VA represents the LATE of going to the VA, and the coefficient on VA × 𝐼˜𝑥,𝑖 represents the difference in the LATE between observations with 𝐼 𝑥,𝑖 = 1 and observations with 𝐼 𝑥,𝑖 = 0. Appendix Table A.9 presents results for additional characteristics. Appendix A.5 provides further details on the hospital characteristics. A.37 Table A.11: Heterogeneity by VA Hospital Characteristics Regression estimates VA VA × 𝐼˜𝑥,𝑖 Volume, Size, and Capabilities ED visits Admissions Total staffed beds Teaching hospital Trauma center Advanced cardiac care Staffing ED staff per 1,000 ED visits Nurses per 1,000 patient-days Physicians per 1,000 patient-days Hospitalists per 1,000 patient-days Intensivists per 1,000 patient-days Spending and Outcomes Relative spending Mortality rate Readmission rate Characteristic means 𝐼 𝑥,𝑖 = 0 𝐼 𝑥,𝑖 = 1 -0.045 (0.017) -0.044 (0.016) -0.044 (0.017) -0.045 (0.017) -0.052 (0.018) -0.051 (0.018) -0.001 (0.002) -0.003 (0.002) -0.007 (0.002) -0.003 (0.002) 0.006 (0.004) -0.004 (0.002) 8,625 23,111 3,247 8,148 139 463 0.00 0.93 0.00 1.00 0.00 1.00 -0.050 (0.022) -0.045 (0.017) -0.050 (0.022) -0.051 (0.022) -0.050 (0.022) -0.001 (0.003) 0.003 (0.002) -0.000 (0.003) 0.006 (0.003) 0.001 (0.003) 0.19 1.21 3.80 8.60 1.12 7.95 0.03 0.30 0.00 0.15 -0.045 (0.016) -0.045 (0.017) -0.045 (0.017) -0.002 (0.002) 0.005 (0.003) -0.003 (0.002) 0.95 1.22 7.11 7.98 11.70 12.70 Note: This table presents regression results investigating heterogeneous treatment effects along characteristics of the VA hospital associated with each zip code. For each VA hospital characteristic 𝑥, we divide observations 𝑖, based on whether 𝑥 is below vs. above the median, denoted by 𝐼 𝑥,𝑖 = 0 and 𝐼 𝑥,𝑖 = 1, respectively. Regression results correspond to Equation (8). The coefficient on VA represents the LATE of going to the VA, and the coefficient on VA × 𝐼˜𝑥,𝑖 represents the difference in the LATE between observations with 𝐼 𝑥,𝑖 = 1 and observations with 𝐼 𝑥,𝑖 = 0. Appendix A.5 provides further details on the hospital characteristics. A.38 Table A.12: Heterogeneity by Patient Characteristics Older than 80 Black Hispanic Income Comorbidity count Mental illness or substance abuse VA visits in prior year Ambulance rides in prior year Advanced Life Support Predicted VA user Predicted mortality Regression estimates VA VA × 𝐼˜𝑥,𝑖 -0.047 0.004 (0.017) (0.003) -0.043 -0.002 (0.017) (0.003) -0.045 -0.008 (0.017) (0.008) -0.044 0.003 (0.017) (0.002) -0.044 -0.014 (0.016) (0.002) -0.045 -0.005 (0.017) (0.002) -0.044 -0.004 (0.017) (0.002) -0.043 -0.008 (0.017) (0.002) -0.046 -0.013 (0.017) (0.002) -0.044 -0.005 (0.017) (0.003) -0.045 -0.018 (0.016) (0.002) Characteristic means 𝐼 𝑥,𝑖 = 0 0.00 𝐼 𝑥,𝑖 = 1 1.00 0.00 1.00 0.00 1.00 $10,651 $31,159 3.90 9.28 0.00 1.00 2.15 11.88 1.00 3.55 0.00 1.00 0.70 1.00 0.04 0.15 Note: This table presents regression results investigating heterogeneous treatment effects along patient characteristics. For each VA hospital characteristic 𝑥, we divide observations 𝑖, based on whether 𝑥 is below vs. above the median, denoted by 𝐼 𝑥,𝑖 = 0 and 𝐼 𝑥,𝑖 = 1, respectively. Regression results correspond to Equation (8). The coefficient on VA represents the LATE of going to the VA, and the coefficient on VA × 𝐼˜𝑥,𝑖 represents the difference in the LATE between observations with 𝐼 𝑥,𝑖 = 1 and observations with 𝐼 𝑥,𝑖 = 0. A.39 A.40 Keep rides for patients with some non-VA or VA utilization (inpatient, ED, or primary care). Drop rides for patients with a prior ride within 30 days. 4. Prior utilization restriction 5. Prior ride restriction Drop rides by an ambulance company with fewer than 20 patients in a given zip code. Drop rides from zip codes with only one remaining ambulance company. Keep only rides for patients with some non-VA utilization (inpatient, ED, or primary care) but no VA utilization in the prior year. Drop rides for patients with a prior ride within 30 days. 4. Ambulance restrictions 5. Prior utilization restriction 6. Prior ride restriction 1,414,217 1,735,141 3,919,572 6,424,120 8,828,997 644,917 644,917 1,372,499 2,131,152 2,862,557 340,371 340,371 365,163 Patients 5,716 5,716 5,716 29,100 180,320 1,217 1,217 1,217 Ambulance companies 96 97 119 122 125 100 100 100 VA 3,799 3,812 3,999 5,498 7,744 1,548 1,565 1,577 Non-VA Hospitals Note: This table details selection steps to create two alternative samples. Panel A shows selection steps for the sample that we use to study descriptive utilization patterns in Appendix Figure A.9. Panel B shows selection steps for the sample of patients with only non-VA prior utilization, which we use in Section 5.5 to study the effect of receiving care at a modal non-VA hospital. At each step, the table lists the number of ambulance rides, patients, ambulance companies, and VA and non-VA hospitals. Table A.14 shows average patient characteristics among observations at each sample step. Drop rides to a hospital over 50 miles from the patient’s home. Drop zip codes without at least two non-VA hospitals within 20 miles that receive at least 5% from that zip code. See step #2 in Table A.1. 794,940 977,826 1,051,093 B: Non-VA-Only Sample 3. Distance restrictions 2. Start from clean sample See step #3 in Table A.1. 3. Start from distance restrictions in baseline sample Rides A: Sample for Descriptive Utilization Patterns Description Sample step Table A.13: Selection of Alternative Samples A.41 2,862,557 8,828,997 Number of patients Number of ambulance rides 2,131,152 6,424,120 0.898 77.12 0.124 $21,763 0.169 0.619 6.62 0.141 0.694 2.83 0.695 0.271 0.116 0.049 Add zip × hospital 1,372,499 3,919,572 644,917 1,735,141 Sample characteristics Add zip × Add ambulance non-VA-only prior utilization 0.897 0.824 77.32 77.68 0.129 0.125 $22,253 $22,800 0.125 0.120 0.657 0.614 6.60 6.96 0.138 0.000 0.687 0.797 2.82 3.13 0.699 0.676 0.271 0.270 0.113 0.117 0.049 0.002 644,917 1,414,217 0.825 78.05 0.118 $23,393 0.120 0.636 6.57 0.000 0.752 2.28 0.684 0.269 0.112 0.002 Add no ride in prior month Note: This table presents characteristics of observations remaining at each step of creating the sample of patients with only non-VA prior utilization, which we use in Section 5.5 to study the effect of receiving care at a modal non-VA hospital. Each step is detailed in Panel B of Appendix Table A.13. 0.899 77.04 0.111 $21,724 0.255 0.610 6.53 0.136 0.695 2.77 0.696 0.272 0.115 0.044 Dually eligible Male Age Black Income Rural zip code Residential source Comorbidity count Prior VA ED visit Prior Medicare ED visit Ambulance rides in prior year Advanced Life Support Weekend rate 28-day mortality Present at VA Restrictions Table A.14: Characteristics of Non-VA-Only Sample