Association between Medicare eligibility at age 65 years and in-hospital treatment patterns and health outcomes for patients with trauma: regression discontinuity approach

Abstract

Objective

To determine whether health systems in the United States modify treatment or discharge decisions for otherwise similar patients based on health insurance coverage.

Design

Regression discontinuity approach.

Setting

American College of Surgeons’ National Trauma Data Bank, 2007-17.

Participants

Adults aged between 50 and 79 years with a total of 1 586 577 trauma encounters at level I and level II trauma centers in the US.

Interventions

Eligibility for Medicare at age 65 years.

Main outcome measures

The main outcome measure was change in health insurance coverage, complications, in-hospital mortality, processes of care in the trauma bay, treatment patterns during hospital admission, and discharge locations at age 65 years.

Results

1 586 577 trauma encounters were included. At age 65, a discontinuous increase of 9.6 percentage points (95% confidence interval 9.1 to 10.1) was observed in the share of patients with health insurance coverage through Medicare at age 65 years. Entry to Medicare at age 65 was also associated with a decrease in length of hospital stay for each encounter, of 0.33 days (95% confidence interval −0.42 to −0.24 days), or nearly 5%), which coincided with an increase in discharges to nursing homes (1.56 percentage points, 95% confidence interval 0.94 to 2.16 percentage points) and transfers to other inpatient facilities (0.57 percentage points, 0.33 to 0.80 percentage points), and a large decrease in discharges to home (1.99 percentage points, −2.73 to −1.27 percentage points). Relatively small (or no) changes were observed in treatment patterns during the patients’ hospital admission, including no changes in potentially life saving treatments (eg, blood transfusions) or mortality.

Conclusions

The findings suggest that differences in treatment for otherwise similar patients with trauma with different forms of insurance coverage arose during the discharge planning process, with little evidence that health systems modified treatment decisions based on patients’ coverage.

Introduction

Understanding the unique consequences of discrepant health insurance systems in the United States is a pressing policy concern. Recent policy proposals, such as Medicare for all, aim to provide uniform health coverage for all Americans, citing inequalities in access and administrative inefficiency associated with the fragmented status quo. For decades studies have found differences in health service utilization and health outcomes between people with different types of health insurance coverage or between those with and without any coverage.^{1 2 3 4 5 6 7 8} Selection bias—wherein people with distinct healthcare needs are classed according to different health insurance products or between being insured and not insured—has, however, limited the understanding of what drives differences in health service use associated with insurance coverage.

In the US, people younger than 65 years may be covered by private insurers or by the Medicaid programme (or other public insurance), or not insured. In contrast, most Americans qualify for Medicare at age 65, resulting in a more uniform source of primary health insurance coverage among older people. The Medicare programme covers hospital inpatients, skilled nursing facilities, home healthcare, hospice services, outpatient and professional services, durable medical equipment, (such as blood sugar meters) and prescription drugs. Within the Medicare programme, beneficiaries decide whether they want their benefits administered directly by the government or through private, Medicare Advantage plans.

These different forms of health insurance coverage may shape the financial incentives of providers (and patients) and impact decisions about treatment. For example, health systems receive lower (or no) reimbursement for treating uninsured or underinsured patients, and even when health systems are reimbursed, payment rates vary widely between payers.^{9 10 11} Compared with private insurers, for example, Medicare has been found to reimburse physicians and hospitals at substantially lower rates.¹² Treatment incentives may also vary based on the type of contracts providers have with payers; fee-for-service and value based contracts, for example, incentivize particular levels of treatment intensity, potentially leading to different treatment paths for otherwise similar patients.^{13 14 15} For institutional care, such as hospital and nursing home care, Medicare relies heavily on prospective payment systems and has introduced new payment models in recent years, including the accountable care organization model and bundled payments. Given that private health insurers are less likely to use prospective payment,¹⁶ the financial incentives around treatment decisions for patients may change as their primary source of coverage changes. Finally, differences in coverage exist between insurers—including Medicare’s “three day rule” in which patients must have a medically necessary inpatient hospital stay of three consecutive days to qualify for additional care in a skilled nursing facility, may also impact how patients are guided through the patchwork US health system.

It is not, however, clear that system level financial incentives to customize treatment are transmitted to individual providers (eg, physicians) who make decisions. Professional ethics and coordination costs may render the differential treatment of patients (within a health system) based on health insurance unlikely in practice, particularly for the existing types of potentially life saving treatments.^{17 18} Yet, previous studies have found increases in access to care and health, reductions in disparities, and even improvements in consumer financial health when adults become eligible for Medicare at age 65 years.^{19 20 21 22 23 24 25 26 27 28} However, a full understanding of whether, and where, in the continuum of care, health systems tailor treatments on the basis of coverage is poorly understood. Hence, the clinical consequences of the disparate US health insurance systems remain unclear, potentially stifling efforts for reform.

Identifying the causal effect of insurance type on health service use and outcomes presents key empirical challenges. People in different insurance products may have different personal profiles (eg, age) or unobserved differences (eg, adherence) that drive their use of services. To account for these differences, we leveraged a natural experiment, wherein nearly all Americans become eligible for Medicare at age 65, to evaluate the association between entry to Medicare and treatment patterns and outcomes for patients with trauma. For several reasons we also limited our sample to people needing care after a trauma. First, previous observational studies found differences in inpatient outcomes by insurance coverage status in the trauma setting.^{29 30 31 32} Second, trauma is often an isolated and acute health event, and the robust system of prehospital emergency medical services and state trauma systems ensure more uniform access for people who have experienced major injury, regardless of their insurance status. As a result, the prevalence and severity of trauma encounters is likely to be similar in those older or younger than 65 years. In this study, we exploited this feature using a regression discontinuity design, comparing patients with trauma who were just older than and just younger than age 65 to balance the observable and unobservable clinical characteristics of participants. Third, there is likely to be less heterogeneity in patient adherence to recommended treatment after trauma, limiting the impact of unobserved differences. Assuming that Medicare eligibility is the only thing changing discontinuously at age 65, sharp changes in treatment of patients with trauma (or their health outcomes) would suggest health systems modify treatments based on coverage for otherwise similar patients.

Methods

Study design and population

We analyzed 2007-17 data from the American College of Surgeons’ (ACS) National Trauma Data Bank (NTDB) for patients at ACS certified trauma centers. NTDB is the world’s largest repeated cross section of trauma data linked to clinical information, containing data from more than 900 trauma centers in the United States, receiving data from nearly every level I and level II trauma center.³³ NTDB collects data submitted voluntarily by hospitals, with each record corresponding to a single trauma encounter. Given concerns about the quality and representativeness of data from lower level trauma centers,³⁴ we restricted our sample to patients at level I or level II trauma centers. NTDB contains individual level personal data, including patient age, gender, race or ethnicity, and insurance type. We limited our primary sample to trauma encounters for patients aged 50 to 79 from 2007 to 2017. The supplementary eMethods section provides additional details of study exclusion criteria.

Study variables

We assessed three primary domains of outcomes: in-hospital treatment, length of stay and discharge disposition, and health outcomes. These measures were chosen to encompass clinically relevant outcomes across the continuum of care—that is, trauma bays, intensive care units (ICU), and hospital wards. To assess in-hospital treatment, we examined the receipt of services in the trauma bay, and, separately, after admission to hospital, as we hypothesized that levels of discretion about treatments might differ between these settings. In the trauma bay we assessed the number of minutes spent in the emergency department, whether patients received transfusions in the first four hours, and whether operative intervention was immediate. We assessed treatment intensity after admission to hospital as measured by the number of procedures performed. Using ICD-9 (international classification of diseases, ninth revision) procedure codes, we grouped procedures into nine categories relevant to the care of patients with acute injuries: diagnostic imaging, blood transfusions, orthopedic surgical intervention, other surgical interventions, bedside procedures (eg, placement of a chest tube), routine medical care (eg, administered antibiotics), invasive medical care (eg, hemodialysis), use of mechanical ventilation, and consultations for physical and occupational therapy. In primary analyses, we present results for overall procedure counts and the three most frequent procedure categories: imaging, orthopedic procedures, and routine medical care. Finally, we examined whether patients were admitted to the ICU (and ICU length of stay) and whether patients were mechanically ventilated (and ventilator days). We included potentially life saving treatments (eg, immediate operative intervention) as outcomes—even though these services should be provided irrespective of insurance status—given observational evidence of insurance status as a risk factor for trauma mortality^{29 31} and the finding that uninsured patients with trauma were less likely to have an operation.³⁵

Second, we examined length of stay and discharge disposition. For length of stay, we assessed how the mean and composition of length of stay changed at age 65 by constructing indicator variables for each discrete length of stay (ie, four days) for each encounter with a length of stay less than 15 days. Encounters with a length of stay of 15 days or longer were grouped together. We also assessed whether discharge location was a nursing home, another inpatient setting, or home. See supplementary eMethods for additional details.

Third, we assessed changes in two sets of health outcomes: inpatient complications (the five most prevalent being urinary tract infections, pneumonia, acute respiratory distress syndrome, drug or alcohol withdrawal, and deep vein thrombosis or pulmonary embolism), and inpatient mortality, measured as a discharge disposition of death or hospice care.

We also assessed each patient’s primary source of health insurance coverage, categorized as Medicare, private or commercial insurance, non-Medicare public insurance, other insurance, or not insured (including self-pay) (see supplementary eMethods).²⁷ To examine the comparability of encounters before and after age 65, we also assessed personal characteristics, including patient age, race or ethnicity, sex, mechanism of injury, and clinical characteristics present on admission (eg, injury severity score).

Statistical analysis

We use a regression discontinuity design that exploits the natural experiment age 65 when nearly all Americans become eligible for Medicare. To isolate the effect of Medicare, we compared outcomes such as coverage, treatment, and health outcomes for individuals who were just younger than and just older than the age 65 eligibility threshold. The key assumption for our method to be valid—which we assess in the study—is that other determinants of coverage, treatment, and health outcomes do not change at age 65, and hence they are similar for individuals with ages above and below the threshold.^{36 37 38} Previous studies have shown that in the US a substantial discontinuous change at age 65 no longer exists in retirement,³⁹ a common concern. Therefore, we used data from NTDB to assess whether any discontinuous changes exist in the personal characteristics of patients with trauma or the clinical characteristics of trauma such as mechanism of injury at age 65 years. To estimate the discontinuity at age 65, we adjusted for age trends using local linear regression with a uniform kernel, allowing for different age trends above and below the discontinuity. The difference between the fitted values of the local linear regressions at age 65, with and without Medicare, measures the adjusted discontinuity as a result of the Medicare eligibility age. The regression discontinuity approach is less susceptible to bias than an unadjusted comparison of outcomes between those with and those without Medicare where there may be confounding due to, for example, patients with trauma without Medicare being on average younger than those with Medicare, and many of our outcomes, such as in-hospital mortality, differ by age. We used the RDHonest R package, an approach developed specifically to analyze regression discontinuity designs when the running variable is discrete (that is, age measured in years in our study) to select the analytic bandwidth (the range of ages around the cut-off) with a data driven method that has a trade-off between bias and variance, and accounts for discreteness in the running variable.^{40 41} The RDHonest approach has become the standard for evaluating discrete regression discontinuity designs in clinical,^{42 43} policy,⁴⁴ and economics research.^{28 45} We report bias adjusted confidence intervals that account for the additional uncertainty in our estimates owing to extrapolation in discrete regression discontinuity designs.^{40 46}

We conducted several sensitivity analyses. First, we assessed the sensitivity of our results to how we model the running variable (linear or quadratic age trends), to the use of a triangular kernel that places more weight on observations closer to the discontinuity, to varying the bandwidth around age 65, to the sensitivity of the approach to extrapolation owing to our discrete variables,⁴⁷ and to including additional covariates. We avoided using higher order polynomials to model the association between our outcomes and age because of evidence that this can lead to overfitting and poor coverage of confidence intervals in regression discontinuity.⁴⁸ Second, we assessed the smoothness of covariates around age 65, common in regression discontinuity, to ensure results were not driven by changes in patient level or facility level characteristics. Third, we conducted placebo tests at ages other than 65 years. Fourth, to assess the stability of our findings we performed sub analyses focused on the period after the implementation of the major coverage provisions in the Affordable Care Act.

Statistical analyses were performed using Stata (version:16.0) and R Studio (version 1.3.1056) statistical software, with primary analyses using the RDHonest R package.⁴⁰ We used two tailed tests of statistical significance with α set at 0.05. To control the false discovery rate of independent hypotheses within families, we used the Benjamini-Hochberg procedure to adjust P values. The supplementary eMethods section provides additional details on sensitivity analyses.

Patient and public involvement

No patients or members of the public were involved in setting the research question or the outcome measures, nor were they involved in the design, implementation, or writing of the study. Although we strongly believe in incorporating patient and public input, we did not have adequate funding for such involvement.

Results

Study population

The study sample included 1 586 577 patients with encounters in level I and level II trauma centers in the US from 2007 to 2017 (table 1 and supplementary eFigure 1). The personal and clinical characteristics of the participants, characteristics of the treating hospitals, and number of trauma encounters were similar above and below the Medicare eligibility age of 65 years (table 1, supplementary eTable1, and supplementary eFigure 2). The share of patients with trauma and any form of health insurance coverage increased by 9.6 percentage points (95% confidence interval 9.1 to 10.1 percentage points) after entry to Medicare at age 65 (table 2 and fig 1), coinciding with a 71.99 percentage point increase in the share of patients covered by Medicare (71.48 to 72.49 percentage points).

Table 1.

Characteristics of study population

	Sample means*			Change at age 65 years
	Age 50-64 years (n=899 454)	Age 65-79 years (n=687 123)		Expected mean†	Adjusted discontinuity (95% CI)‡
Personal characteristics
Sex:
Female	308 341 (34)	345 621 (50)		41.83	0.48 (−0.16 to 1.13)
Race or ethnicity:
Black	119 735 (13)	46 525 (7)		9.33	−0.13 (−0.51 to 0.23)
White	673 737 (75)	576 119 (84)		80.50	0.43 (−0.07 to 0.80)
Hispanic	66 589 (7)	35 249 (5)		5.93	−0.18 (−0.49 to 0.15)
Mechanism of injury:
Motor vehicle crashes	293 866 (33)	141 903 (21)		26.91	0.09 (−0.47 to 0.66)
Falls	377 178 (42)	469 255 (68)		56.67	−0.21 (−0.81 to 0.38)
All other injuries	228 410 (25)	75 965(11)		16.14	0.33 (0.02 to 0.63)
Head injury	780 782 (87)	620 515 (90)		88.64	0.01 (−0.35 to 0.38)
Injury type:
Blunt	712 216 (90)	568 097 (96)		94.16	0.18 (−0.16 to 0.53)
Penetration	44 878 (6)	11 183 (2)		2.99	−0.06 (−0.32 to 0.20)
Clinical characteristics
Glasgow coma score§:
3-8	56 714 (6)	30 946 (5)		5.18	−0.07 (−0.28 to 0.15)
9-12	20 609 (2)	14 018 (2)		2.04	−0.11 (−0.21 to −0.01)
13-15	754 026 (84)	575 311 (84)		83.80	0.03 (−0.39 to 0.32)
Injury severity score¶:
1-8	287 566 (32)	189 068 (28)		29.55	0.21 (−0.19 to 0.61)
9-15	233 572 (26)	204 971 (30)		28.68	−0.46 (−0.71 to −0.20)
16-24	113 405 (13)	91 598 (13)		12.98	0.21 (−0.02 to 0.44)
≥25	264 911 (29)	201 486 (29)		28.79	−0.10 (−0.49 to 0.29)
Hypotensive in trauma bay	29 047 (3)	16 954 (2)		2.94	0.11 (−0.09 to 0.32)
Hospital characteristics
Teaching hospital	456 760 (51)	307 633 (45)		48.08	−0.14 (−0.62 to 0.33)
Level I trauma center	566 385 (63)	398 294 (58)		60.69	−0.09 (−0.49 to 0.31)

^*Values are raw counts of trauma episodes (percentages).

^†Predicted based on local linear association between age and each outcome. The column shows the expected value of the outcome at age 65 years in the absence of Medicare eligibility.

^‡Estimates in percentage points.

^§Measures decrease in consciousness—the lower the score, the more severe the injury.

^¶Measures trauma severity for patients with multiple injuries, with a score of 1 representing a minor injury and a score of 75 representing a fatal injury.

Table 2.

Changes in insurance coverage, complications, and mortality after age 65 years, 2007-17

	Sample means*			Change at age 65 years
	Age 50-64 years (n=899 454)	Age 65-79 years (n=687 123)		Expected mean†	Adjusted discontinuity (95% CI)‡	P value§
Health insurance status
Medicare	115 148 (12.8)	651 023 (94.8)		20.92	71.99 (71.48 to 72.49)	<0.001 (<0.001)
Any insurance	721 314 (80.2)	666 399 (96.9)		86.09	9.60 (9.05 to 10.13)	<0.001 (<0.001)
Select complication rates
Any complication	257 393 (29.8)	190 355 (28.7)		28.35	0.21 (−0.34 to 0.76)	0.47 (0.75)
Urinary tract infection	7783 (0.9)	9994 (1.5)		1.15	0.02 (−0.06 to 0.10)	0.63 (0.75)
Pneumonia	21 333 (2.4)	15 540 (2.3)		2.39	−0.11 (−0.24 to 0.02)	0.09 (0.28)
Drug or alcohol withdrawal	15 494 (1.7)	5064 (0.7)		1.40	−0.17 (−0.32 to −0.02)	0.02 (0.11)
DVT or PE	13 677 (1.5)	9921 (1.4)		1.56	−0.009 (−0.11 to 0.09)	0.87 (0.87)
Acute respiratory distress	10 558 (1.2)	7676 (1.1)		1.13	0.02 (−0.04 to 0.09)	0.52 (0.75)
Mortality	27 946 (3.1)	35 470 (5.2)		3.94	−0.14 (−0.35 to 0.07)	0.21

DVT=deep vein thrombosis; PE=pulmonary embolism.

^*Values are raw counts of trauma episodes (percentages).

^†Predicted based on local linear association between age and each outcome. The column shows the expected value of the outcome at age 65 years in the absence of Medicare eligibility.

^‡Estimates in percentage points.

^§To control for the false discovery rate of independent hypotheses within families, the Benjamini-Hochberg procedure was used to adjust P values, presented in parentheses (see supplementary eMethods).

Fig 1.

Changes in health insurance, treatment intensity, and health outcomes after age 65 years. The mean of outcomes is plotted by age (in years) for the study period. For illustrative purposes, a line of best fit is plotted based on a local linear regression, with the optimal bandwidth selected by the study’s regression discontinuity model. Dashed line denotes Medicare eligibility threshold, and open circle represents the expected mean of the outcome at age 65 in the absence of Medicare eligibility

Changes in processes of care at age 65 years

We found little evidence of statistically significant discontinuities at age 65 in treatments received in the trauma bay, including the number of minutes in the emergency department (decrease of 4.3 minutes, 95% confidence interval −12.8 to 4.1 minutes), blood transfusion volume in four hours (decrease of 0.4 mL, 95% confidence interval −1.3 to 0.5 mL), plasma transfusion volume in four hours (increase of 1.05 mL, −4.4 to 6.5 mL), or platelet transfusion volume in four hours (increase of 0.6 mL, −0.9 to 2.1 mL). A small reduction was found in the likelihood of an immediate operative intervention (decrease of 0.3 percentage points, 95% confidence interval −0.5 to −0.1 percentage points), but it was not robust to sensitivity analyses (see supplementary eTable 2 and eTable 3).

We also found little evidence of changes in treatment intensity during the broader trauma encounter. The only exception was a statistically significant, albeit small, increase of 0.04 percentage points (95% confidence interval 0.00 to 0.08 percentage points), or nearly 2%, in the number of diagnostic imaging procedures after age 65, but this change was not statistically significant after adjustment for multiple hypothesis testing. We did not detect a change in any other procedure category at age 65, nor in the overall number of procedures (table 3, supplementary eTable 4, and supplementary eFigure 3 and eFigure 4). Nor did we detect statistically significant differences after age 65 in ICU admission, number of ICU days, the likelihood of requiring ventilatory support, number of ventilator days, or number of minutes to venous thromboembolism prophylaxis.

Table 3.

Changes in processes of care after age 65 years, 2007-17

	Sample means*			Change at age 65 years
	Age 50-64 years (n=899 454)	Age 65-79 years (n=687 123)		Expected mean†	Adjusted discontinuity (95% CI)‡	P value§
Trauma bay
Time in emergency department (mins)	307.31	313.99		311.82	−4.33 (−12.79 to 4.13)	0.33 (0.58)
Transfusion volume in 4 hours (mL):
Blood	4.77	2.90		4.15	−0.40 (−1.31 to 0.50)	0.40 (0.58)
Plasma	13.49	9.39		12.29	1.05 (−4.43 to 6.54)	0.72 (0.72)
Platelets	3.54	2.63		2.99	0.59 (−0.93 to 2.11)	0.46 (0.58)
Immediate operative intervention	106 764 (11.87)	50 143 (7.30)		9.63	−0.27 (−0.45 to −0.10)	0.004 (0.02)
Treatment intensity in hospital¶
Total procedure count	5.75	5.15		5.47	0.01 (−0.07 to 0.09)	0.79 (0.95)
Procedure type:
Imaging	2.29	2.14		2.18	0.041 (0.004 to 0.08)	0.03 (0.16)
Orthopedic	0.81	0.66		0.77	−0.005 (−0.025 to 0.014)	0.60 (0.95)
Routine medical care	0.52	0.59		0.56	0.002 (−0.007 to 0.011)	0.71 (0.95)
Admitted to ICU	274 459 (30.51)	224 392 (32.66)		31.74	−0.05 (−0.50 to 0.41)	0.85 (0.95)
ICU length of stay (days)	1.84	1.80		1.91	−0.001 (−0.03 to 0.03)	0.97 (0.97)
Required ventilatory support	113 630 (12.63)	75 288 (10.96)		11.74	0.10 (−0.12 to 0.33)	0.40 (0.95)
Time on ventilator (days)	0.93	0.80		0.92	−0.01 (−0.05 to 0.02)	0.54 (0.95)
Time to VTE prophylaxis (mins)	2420.82	2463.49		2490.04	−49.86 (−108.81 to 8.96)	0.10 (0.34)
Length of stay (days)	6.74	6.38		6.72	−0.33 (−0.42 to −0.24)	<0.001 (<0.001)
Discharge location:
Home	588 324 (65.41)	319 080 (46.44)		58.07	−1.99 (−2.73 to −1.27)	<0.001 (<0.001)
Nursing home**	134 499 (14.96)	224 061 (32.61)		22.25	1.56 (0.94 to 2.16)	<0.001 (<0.001)
Other inpatient††	55 201 (5.80)	63 130 (9.19)		7.74	0.57 (0.33 to 0.80)	<0.001 (<0.001)

ICU=intensive care unit; VTE=venous thromboembolism.

^*Values are raw counts of trauma episodes (percentages).

^†Predicted based on local linear association between age and each outcome. The column contains the expected value of the outcome at age 65 years in the absence of Medicare eligibility.

^‡Estimates in percentage points, when applicable.

^§To control the false discovery rate of independent hypotheses within families, the Benjamini-Hochberg procedure was used to adjust P values, presented in parentheses (see supplementary eMethods).

^¶Procedure counts rely on data from 2007 to 2016 only because the NTDB transitioned from using ICD-9 to ICD-10 (international classification of diseases, ninth and 10th revision, respectively) procedure codes in 2017.

^**Including skilled nursing facility.

^††Includes discharges or transfers to a short term general hospital for inpatient, and discharges or transfers to inpatient rehabilitation or designated unit.

Changes in length of stay and discharge location

We found evidence of substantial changes in length of stay and discharge location at age 65 years. Length of stay decreased by 0.33 days (95% confidence interval −0.42 to −0.24 days), or nearly 5% at age 65 years (table 3 and fig 1). The decrease in length of stay was driven by a reduction in trauma episodes of 15 days or longer and a concomitant increase in episodes wherein patients were discharged on exactly the fourth day after an admission (fig 2). The composition of discharge locations also changed at age 65 years. The share of individuals discharged home decreased by 1.99 percentage points (95% confidence interval −2.73 to −1.27 percentage points), whereas discharges to nursing homes (including skilled nursing facilities) and transfers to other inpatient facilities increased by 1.56 percentage points (0.94 to 2.16 percentage points) and 0.57 percentage points (0.33 to 0.80 percentage points), respectively (table 3 and fig 1).

Fig 2.

Changes in composition of length of stay after age 65 years. The plot presents local linear regression discontinuity (RD) estimates of the adjusted discontinuity between age and discrete indicators for different lengths of stay. Encounters with lengths of stay of 15 days or longer are grouped together

Changes in health outcomes at age 65 years

Of the top five most frequent complications, only the share of individuals who developed drug or alcohol withdrawal syndrome changed significantly (absolute decrease of 0.17 percentage points (−0.32 to −0.02 percentage points), or 10% (table 2 and supplementary eFigure 5), but this change was not statistically significant after adjustment for multiple hypothesis testing. A statistically insignificant 0.14 percentage point (−0.35 to 0.07 percentage points), or 4%, reduction in mortality was observed at age 65 years.

Additional analyses

Sensitivity analyses, including the use of alternative kernels and bandwidths, changes in the statistical model, adjustment for additional covariates, covariate smoothness tests, and restricting the period to after the Affordable Care Act was introduced supported our primary analyses (see supplementary eTables 2-7 and supplementary eFigure 6). Results did not qualitatively change after stratification by mechanism of injuries (eg, motor vehicle crashes) (see supplementary eTables 8-13), nor when recoding missing data for specific outcomes (eg, plasma and platelet transfusions) that were conditional on receiving a treatment (ie, blood transfusions) (see supplementary eTable 14). We also did not find comparable estimated effects in our primary outcomes at placebo age cut-offs other than 65 years (see supplementary eFigure 7). Finally, our primary findings were similar when we restricted analysis to a subset of patients with isolated hip fractures; we found that this result was not related to variations in time to curative intervention by presenting evidence of no change at age 65 in time to the operating room (see supplementary eTable 15).

Discussion

Entry to Medicare at age 65 years was associated with a 5% reduction in length of stay (0.33 days for each encounter with trauma) and an increase in discharges to nursing homes and nursing homes (including skilled nursing facilities) (as opposed to home). The association between entry to Medicare and in-hospital treatment patterns was more complicated: We observed relatively small (or no) changes in treatment patterns while patients were in hospital. For example, we found no changes in life saving treatments in the immediate aftermath of trauma (eg, blood or plasma transfusion), but we did detect a small increase, of 2%, in the mean number of diagnostic imaging tests after age 65 years. Although mortality at age 65 did not change substantially, there was a 10% reduction in complications due to alcohol and drug withdrawal syndrome (though no change in rates with other complications).

Comparison with other studies

Our study suggests Medicare eligibility impacts processes of care for patients with trauma, but largely by influencing how long patients remain in the hospital and discharge location, rather than the healthcare received or health outcomes (including mortality). The lack of a mortality effect is surprising, given evidence that health insurance reduces mortality in the general population,^{4 6 7 20 49} and descriptive evidence that a lack of health insurance coverage is associated with higher inpatient mortality among patients with trauma.^{29 30 31 32} Our study contributes to this literature by using a quasi-experimental design to overcome the potential for selection bias that was present in previous descriptive analyses. In addition, we focused on the trauma setting where it is well established that the key driver for inpatient survival is the acute phase of treatment (within the first few hours) of arrival in a trauma bay,^{50 51 52} differentiating our setting and population from previous quasi-experimental studies of the impact of health insurance on mortality.^{4 6 7 20 49} Although previous studies found changes in treatment intensity and mortality as a result of Medicare eligibility,^{19 39} in our study, the lack of a mortality effect was less surprising given our evidence that treatment in the trauma bay was similar for patients before and after Medicare eligibility.

Nevertheless, we found differences in length of stay and discharge location at age 65, suggesting that changes in coverage and policy impact how patients are guided through the health system, even if decisions do not directly impact access to life saving treatments. The reductions in length of stay we found were only slightly smaller in magnitude than the estimated impact of the introduction of diagnosis related groups on hospital inpatient length of stay.⁵³ In contrast with previous studies that measured the association between insurance coverage and length of stay in the trauma setting, our study was larger in scale and scope.^{35 36 37} For example, we were able to observe an increase in the share of patients discharged to a nursing home after age 65 years (and a concomitant decrease in discharges to home). Coinciding with this change in discharge location, the share of patients discharged on the fourth day of an encounter after age 65 increased. This is consistent with the Medicare three day rule—to qualify for coverage of extended care services in a skilled nursing facility, patients with Medicare must have a medically necessary inpatient hospital stay of three consecutive days (three days plus the discharge day). These findings are in line with previous studies that have shown that the decision to discharge to a skilled nursing facility, and length of stay in the hospital, is associated with the Medicare three day rule.^{54 55} Coinciding with the increase in the number of hospital admissions with lengths of stay of exactly four days (the minimum length of stay needed to qualify for the three day rule) at age 65, we find a decrease in hospital admissions with lengths of stay of 15 days or more. One possibility is that some patients younger than age 65 who are candidates for discharge to a post-acute care facility are denied access to such care owing to insufficient insurance coverage, with hospitals responding by waiting longer to discharge patients until they are safe to return to their home. In this case, reduced access to post-acute care before age 65 may indirectly increase costs for the healthcare system and for patients through longer hospital stays. Alternatively, it is possible that after age 65 hospitals discharge patients with Medicare earlier than they otherwise would, or that Medicare’s policies induce hospitals to discharge patients to post-acute care who could be discharged safely to home. While we cannot distinguish between these potential explanations, in either case distinct policies clearly incentivize distinct management of otherwise similar patients.

Our findings also highlight the set of trade-offs facing health systems such as those in the US, that have elected not to provide a universal source of coverage for all citizens. On the one hand, greater choice may allow a health system to respond to disparate preferences or needs in a population or facilitate efficiency promoting competition between payers or forms of coverage. On the other hand, we found that a patchwork system can also lead otherwise identical patients to be steered through the health system in ways that increase inequities and lead to suboptimal access to services for some citizens.

Strengths and limitations of this study

The strengths of our study include the use of NTDB, the largest trauma dataset, and the use of a regression discontinuity design, which enabled us to the measure the effect of Medicare eligibility at age 65 on coverage, treatment, and health outcomes. In addition, NTDB provides a comprehensive set of clinical outcomes to study, allows assessment of covariate balance on a wide range of conditions present on admission, and offers a unique view into how treatment and management changes at different points in the process of care for patients with trauma.

This analysis had some limitations. First, NTDB is limited to a sample of American College of Surgeons’ certified trauma centers, and we restricted our study to level I and level II trauma centers. Our results may not generalize to all trauma centers (or other facilities) in the US. Second, although NTDB has a robust set of clinical and personal covariates, it is possible that unmeasured differences in risk exist between those younger than or older than 65, potentially biasing our results. Third, the running variable for our regression discontinuity models (age) was only available in discrete years so we used the RDHonest package to account for any bias that might be introduced by misspecification in the function that relates our outcomes to age. Fourth, evidence from the trauma setting may not generalize to other care settings (eg, cancer) with greater provider discretion or where patient-provider communication impacts treatment patterns in ways that are often not possible in the trauma setting, such as the treatment of unconscious patients. However, despite these potentially unique features of the trauma setting, the literature on trauma in the US is replete with studies suggesting that health insurance coverage is independently associated with the outcomes we studied.^{29 30 31 32} Fifth, our results may not generalize to patients with trauma with missingness of data, as they were excluded from the study. The missingness of data for these patients was, however, low.

Lastly, because we did not have longitudinal data, we were unable to observe patients as they became eligible for Medicare at age 65 years. Hence, we cannot differentiate between the two different pathways from which patients gain Medicare coverage: transitioning from being uninsured to gaining Medicare at age 65 or switching from private insurance to Medicare as a primary payer at age 65 years.

Conclusions and public health implications

For patients with trauma, entry to Medicare at age 65 was immediately associated with a shorter length of stay and changes to discharge locations. However, it also showed limited impact on treatment patterns during hospital admission, with negligible detectable effects on health and no effect on mortality.

What is already known on this topic

• Studies have found differences in health service use between patients with different health insurance coverage, such as the association between being uninsured and higher mortality among patients with trauma

• It is unclear, however, whether these differences are due to patient characteristics or to differences in how patients are treated

• Some evidence shows that health insurance coverage impacts treatment patterns for otherwise similar patients, but it is unclear where in the processes of care these differences arise and whether they extend to life saving care needed to stabilize patients with trauma

What this study adds

• The findings of this study suggest that people who are eligible for Medicare at age 65 years are admitted to the hospital for less time and are more likely to be discharged to a nursing home than those just below age 65

• Differences in healthcare treatments during the hospital admission were, however, found to be minimal, and health outcomes did not differ

• The findings suggest that differences in treatment for otherwise similar patients with trauma arose during the discharge planning process, with little evidence that healthcare systems modify their treatment decisions based on patients’ coverage

Web extra.

Extra material supplied by authors

Supplementary information: eMethods, eTables 1-15, and eFigures 1-7

Ethics statements

Ethical approval

Yale University’s institutional review board deemed this study exempt from review.

Data availability statement

The dataset from this study is held securely at the Yale School of Public Health. The patient level dataset cannot be made publicly available; however, the software code used for the primary analyses will be made available upon request.