The effect of a public transportation expansion on no‐show appointments

Abstract

Objective

To test whether there were fewer missed medical appointments (“no‐shows”) for patients and clinics affected by a significant public transportation expansion.

Study setting

A new light rail line was opened in a major metropolitan area in June 2014. We obtained electronic health records data from an integrated health delivery system in the area with over three million appointments at 97 clinics between 2013 and 2016.

Study design

We used a difference‐in‐differences research design to compare whether no‐show appointment rates differentially changed among patients and clinics located near versus far from the new light rail line after it opened. Models included fixed effects to account for underlying differences across clinics, patient zip codes, and time.

Data extraction methods

We obtained data from an electronic health records system representing all appointments scheduled at 97 outpatient clinics in this system. We excluded same‐day, urgent care, and canceled appointments.

Principal findings

The probability of no‐show visits differentially declined by 0.5 percentage points (95% confidence interval [CI]: −0.9 to −0.1), or 4.5% relative to baseline, for patients living near the new light rail compared to those living far from it, after the light rail opened. The effects were stronger among patients covered by Medicaid (−1.6 percentage points [95% CI: −2.4 to −0.8] or 9.5% relative to baseline).

Conclusions

Improvements to public transit may improve access to health care, especially for people with low incomes.

What is known on this topic

• Missed medical appointments (i.e., “no‐shows”) are harmful to patients and create inefficiencies in health care delivery.

• Transportation barriers are frequently cited as a reason for no‐shows, but little research has examined the causal effect of improved transportation options on no‐shows.

• Existing studies have focused on the effects of ridesharing interventions and have been limited to small patient samples.

What this study adds

• Using data reflecting over three million scheduled outpatient appointments, we found that the opening of a new light rail transit line led to relative reductions in rates of no‐shows among patients living near it compared to those not living near it.

• We found the effects of the new transit line were larger among patients covered by Medicaid.

• Our findings suggest that improvements to public transit may improve access to care for patients.

1. INTRODUCTION

No‐show medical appointments—when patients do not show up for their scheduled appointments—are common. Providers and delivery systems use many strategies to encourage patients to attend their appointments, but despite these efforts, recent studies find that anywhere from 6 to 33% of scheduled appointments are no‐shows. ¹

No‐show appointments are harmful to patients and create inefficiencies for providers and health systems. For patients, no‐shows can delay screenings, treatments, or medication refills; exacerbate chronic disease; and increase future use of emergency department and hospital services. ¹ , ² , ³ , ⁴ For providers, no‐shows disrupt workflow and complicate treatment plans. For health systems, no‐shows mean lost revenue and unintended downtime for the clinical workforce. ⁵

Many prior observational studies have identified factors associated with no‐shows, such as patient age, socioeconomic status, appointment wait time, and health. ⁵ , ⁶ , ⁷ , ⁸ , ⁹ Transportation barriers are also associated with increased no‐shows, especially among people with low incomes living in urban areas. ⁴ , ⁹ , ¹⁰ , ¹¹

To address transportation barriers, state Medicaid programs have historically been required to provide nonemergency medical transportation (NEMT) benefits for Medicaid enrollees. NEMT benefits are meant to cover trips to and from medical appointments via taxi, van, personal vehicles, and public transportation. However, the delivery and administration of NEMT vary considerably across and within states, and even with NEMT benefits, Medicaid enrollees frequently report missing or delaying care because they lack transportation. ¹² , ¹³

Most existing research on the relationship between transportation barriers and no‐shows has relied on cross‐sectional research designs and small sample sizes. ⁴ There is limited research on the effectiveness of NEMT benefits in Medicaid, ¹² , ¹⁴ , ¹⁵ and only two studies have examined the causal effect of transportation availability on no‐shows. An analysis of administrative data from an inner‐city safety net provider found no change in the rate of no‐show appointments during a three‐week mass transportation strike, compared to before the strike. However, the study's research design lacked a control group, and the study is more than 20 years old. ¹⁶ A more recent randomized controlled trial found no evidence that offering complimentary ridesharing services to Medicaid patients was effective in reducing no‐shows. The study included a sample of fewer than 1500 patients covered by Medicaid with appointments at two internal medicine practices, and it did not consider the role of public transportation. ¹⁷

To our knowledge, no prior study has directly evaluated the effect of improvements or expansions of public transit systems on no‐shows. In this article, we used the opening of a major light rail line in Minneapolis–St. Paul in 2014 as a natural experiment to investigate the relationship between public transit availability and the probability of no‐shows at outpatient primary care and specialty clinics. We contribute to existing literature on no‐shows and the role of transportation in health care accessibility in three key ways. First, we used a quasi‐experimental research design that leveraged variation in which patients and clinics benefitted from a public transportation expansion to estimate the causal effect of improved public transit on no‐shows. Second, we used data that represented over 3 million appointments at 97 clinics over a four‐year timespan, far exceeding the sample size of many past studies. ² , ⁶ , ¹⁷ , ¹⁸ Third, we described heterogeneous treatment effects, focusing on Medicaid patients for whom additional low‐cost transportation options might be most effective in lowering no‐show rates.

2. METHODS

2.1. Study context

On June 14, 2014, the METRO Green Line (“Green Line”) opened to the public in the Twin Cities of Minneapolis and St. Paul, Minnesota, constituting a significant improvement to the local public transit system. ¹⁹ The 11‐mile, 23‐station Green Line light rail operated 24 h a day, 7 days a week, and connected downtown Minneapolis to downtown St. Paul via rail for the first time in over 60 years. ²⁰ , ²¹ The line generated 3.4 million additional rides along its route in its first 6 months alone—more than one‐third of existing annual light rail ridership. ²² , ²³

The Green Line was the second major light rail line to open in the Twin Cities, following the METRO Blue Line, which had opened 10 years prior and ran primarily through Minneapolis. ²⁴ Initial planning for the project began as early as 1981, with a real focus on improving transit along the popular route—known as the Central Corridor—starting in 2001. Despite considerable political resistance, the project was ultimately approved by the Federal Transit Administration in May of 2010, and construction began shortly thereafter. ²⁵

As a rail system, the Green Line provided a reliable and frequent public transit option compared to the existing bus lines along its route. The residential neighborhoods surrounding the Green Line experienced high rates of poverty and were home to some of the most racially and ethnically diverse populations in the Twin Cities. ²⁶ The rail service was impervious to traffic patterns, including congestion due to rush hour and traffic accidents, and was more resilient to inclement or winter weather. The Green Line also ran more frequently than buses—peaking at every 10 min during the day. ²⁷ For patients traveling to medical appointments, the frequency and reliability of the Green Line likely made it an appealing mode of transportation. ¹⁰ , ²⁸ This may have been particularly true for patients working inflexible or hourly paying jobs where long or unpredictable travel times posed high opportunity costs.

2.2. Data

We used a database of Electronic Health Records (EHR) and scheduling information from 97 primary care and specialty clinics operated by M Health Fairview, a large, integrated health care system in the Twin Cities region. The data reflected all outpatient appointments scheduled to occur with M Health Fairview providers from 2013 through 2016. For each scheduled appointment, we observed the appointment status (completed, canceled, or no‐show). We also observed patient demographics, provider type, department location and specialty, and visit‐level characteristics such as the time of day that the appointment was scheduled to occur. We grouped departments into “clinics” according to street address. We obtained neighborhood socio‐economic characteristics (zip code level) from the American Community Survey.

2.3. Study sample

Our study sample included completed and no‐show appointments scheduled to occur between January 1, 2013 (approximately 18 months before the Green Line opened) and December 31, 2016 (approximately 24 months after the Green Line opened). We excluded appointments canceled one or more days in advance, since we did not know why the cancellations occurred (i.e., whether the cancellation was initiated by the patient or provider), nor if the appointments were rescheduled for a different date. We also excluded all urgent care and same‐day appointments (appointments added to the schedule the same day they occurred) because these appointments were highly unlikely to be no‐shows by definition. We excluded appointments scheduled at departments that opened for the first time later than June 2014 or closed prior to June 2014 (and therefore did not exist in both the “pre” and “post” periods), and we excluded the small number of appointments (<2%) missing key variables (Table S1).

The construction of the Green Line prompted considerable residential and commercial development, which may have attracted new residents with different underlying propensities for no‐shows. ²⁹ To account for the possibility that the opening of the Green Line spurred compositional changes in the population of patients in its proximity or the population of patients scheduling appointments at clinics near the Green Line, we limited our main analytic sample to a fixed cohort of patient‐clinic combinations. Specifically, in the “post” period, we only included patient–clinic combinations with at least one appointment scheduled in the “pre” period. In sensitivity analyses, we did not restrict the sample to a fixed cohort of patients, and the results were similar (Table S2). Our final analytic sample included 3,572,963 appointments for 371,597 unique patients. Table S1 summarizes the creation of our study sample.

2.4. Key variables

The key outcome variable was a dichotomous indicator equal to one when an appointment was recorded as a no‐show in the scheduling data. We created two key “treatment” variables: one reflecting patient proximity to the Green Line and one reflecting clinic proximity to the Green Line. Patients may travel to appointments from their place of work rather than from their home address, meaning a clinic's proximity to the Green Line could be relevant in addition to a patient's home address proximity. We therefore considered both of these perspectives since medical appointments are typically scheduled to occur during weekday, standard business hours.

The patient proximity treatment variable was a dichotomous indicator equal to one if a patient's residential zip code—the most granular information we had for patient residential address—had a Green Line station located within it (i.e., if the patient lived “near the Green Line”). Each patient was attributed to the most recent residential zip code associated with their patient record, as this is the only zip code we could observe. The zip‐code centroids with a Green Line station were an average walking distance of 0.7 miles (about 14 min) from the nearest station, according to Google Maps. The clinic proximity treatment variable was a dichotomous indicator equal to one if a clinic was located within a one‐mile walk (about 20 min) of a Green Line station (i.e., if the clinic was located “near the Green Line”), based on the street address of the clinic and walking distances estimated by Google Maps. In sensitivity analyses, we changed this one‐mile threshold to a half‐mile.

Figure 1 depicts the geographic location of the M Health Fairview primary care clinics and their proximity to the Green Line (as of 2016). Fourteen of the 97 clinics (14.4%) were located within a one‐mile walking distance of the Green Line, and 8 (8.2%) were located within a half‐mile.

FIGURE 1.

Location of light rail stations and M Health Fairview clinics in the Twin Cities, c. 2016. Bottom panel shows magnified view of outlined rectangle in top panel [Color figure can be viewed at wileyonlinelibrary.com]

2.5. Study design

We used a difference‐in‐differences (DID) research design to compare no‐show rates over time for patients living near the new light rail to a comparison group of patients not living near it. We also compared no‐show rates at clinics located near the new light rail to a comparison group of clinics not near the new light rail. In the first set of models, we considered patients to be “treated” if they lived near the Green Line (i.e., within the same zip code as a station) and to be “comparison” patients if they did not live near the Green Line. In the second set of models, we considered clinics to be “treated” if they were located near the Green Line (i.e., within a one‐mile walking distance of a station) and to be “comparison” clinics if they were not located near the Green Line.

We repeated both analyses, focusing on patients covered by Medicaid. We defined Medicaid patients as patients who had Medicaid coverage at any time during 2013–2016. Medicaid patients were of particular interest to us for three reasons. First, they had high rates of no‐shows. Second, a disproportionate share of this group lived in close proximity to the new light rail line, potentially making it a more salient expansion of their transportation options. Third, we expected an increase in public transit availability to be particularly influential for this group since it is a low‐cost mode of transportation, and Medicaid patients are disproportionately likely to experience transportation barriers. ⁴ , ⁶ , ¹⁷ , ³⁰

Since the Green Line opened to the public on June 14, 2014, in all DID models, we classified appointments scheduled to occur between January 1, 2013, and June 13, 2014, as belonging to the “pre” period. We chose January 1, 2013, as the first day of the preperiod because heavy construction on the line was 84 percent complete by the end of 2012, and 2013 was largely spent installing communications and electrical systems and testing trains, meaning major traffic disruptions near the line were less likely during 2013. ²⁵ , ³¹ We classified appointments scheduled to occur between June 14, 2014, and December 31, 2016, as belonging to the “post” period. Therefore, the DID models estimated the change in the expected probability of no‐shows after the Green Line opened for patients living near the Green Line (clinics located near the Green Line) relative to patients not living near it (clinics not located near it), after differencing‐out unobservable characteristics of patient zip codes (clinics) and underlying time‐specific patterns in no‐shows. The estimating equations were of the following general form:

Models where patients were treated or not

$$\mathrm{No}\_{\text{Show}}_{\mathit{\text{izjt}}}=\beta \text{Near}\_{\text{Green}}_{\mathit{zt}}+{\lambda }_z+{\tau }_t+{\delta }_j+{\epsilon }_{\mathit{\text{izjt}}}$$ (1)

Models where clinics were treated or not

$$\mathrm{No}\_{\text{Show}}_{\mathit{\text{izjt}}}=\beta \text{Near}\_{\text{Green}}_{\mathit{jt}}+{\lambda }_z+{\tau }_t+{\delta }_j+{\epsilon }_{\mathit{\text{izjt}}}$$ (2)

In both specifications, $\mathrm{No}\_{\text{Show}}_{\mathit{\text{izjt}}}$ was equal to one when patient $i$ from zip code $z$ was a no‐show at clinic $j$ in year‐month $t$, and equaled zero otherwise. In Equation (1), $\text{Near}\_{\text{Green}}_{\mathit{zt}}$ was equal to one when the patient's zip code $z$ was near a Green Line station and year‐month t followed the opening of the Green Line (i.e., June 14, 2014), and equaled zero otherwise. In Equation (2), $\text{Near}\_{\text{Green}}_{\mathit{jt}}$ was equal to one when clinic $j$ was near a Green Line station and year‐month t followed the opening of the Green Line and equaled zero otherwise. Zip code level–fixed effects (${\lambda }_z$) accounted for underlying differences across patient zip codes that may have affected no‐shows (e.g., geographic location within the metropolitan area) and year‐month fixed effects (τ _t ) accounted for time‐specific changes in the rates of no shows across all clinics (e.g., weather‐related incidents). Clinic‐fixed effects (δ _j ) accounted for underlying differences across clinics. In sensitivity analyses, we included patient‐, visit‐, and area‐level control variables in the specifications (Table S3).

We used ordinary least squares to estimate Equation (1), with standard errors clustered at the patient zip code level. The fixed effects allowed us to interpret $\widehat{\beta }$ as the average within‐zip code change in the probability of a no‐show for patients living near the Green Line after it opened, relative to patients not living near the Green Line. To estimate Equation (2), we used ordinary least squares with standard errors clustered at the clinic level. $\widehat{\beta }$ represents the average within‐clinic change in the probability of a no‐show at clinics located near the Green Line after it opened, relative to clinics not located near the Green Line.

In addition to the DID analyses, we estimated event study models. In the event study specifications, we replaced the single $\text{Near}\_\text{Green}$ variables with interaction terms for an indicator of patient/clinic proximity to the Green Line and each quarter relative to the Green Line opening.

2.6. Assumptions for causal inference

In order to interpret our DID estimates as the causal effect of the Green Line opening on no‐shows, we had to assume that trends in no‐shows would have been similar among patients living near the Green Line (or clinics located near the Green Line) compared to patients not living near the Green Line (or clinics not located near the Green Line) after June 14, 2014, had the Green Line not opened. We had to assume that there were no outside factors simultaneously correlated with the opening of the Green Line and with the probability of no‐shows among patients (or clinics) near the Green Line.

A possible threat to this assumption would arise if significant changes to neighborhood characteristics coincided with the Green Line opening. Specifically, an influx of new real estate and retail development along the line could attract residents with different underlying propensities for no‐shows compared with residents who lived near the Green Line before its opening. We observe only one zip code per patient and therefore cannot see if patients moved during the study period. However, by limiting our analytic sample to a fixed cohort of patient–clinic combinations with at least one visit in the pre‐period, we minimize the bias that would arise from a changing population of patients scheduling appointments at any given clinic.

Additionally, we do not have reason to believe that significant shifts in neighborhood characteristics coincided exactly with the light rail's opening date. Development along the line occurred gradually over the course of several years; it did not emerge immediately when the line opened for use. ³² Specifically, there is evidence that development activity began 3 years prior to the Green Line opening, when a full funding grant announcement for the project increased media coverage and planning activities. ²⁹ Increases in housing value along the Green Line also began in 2011. ³³ To further account for changing local demographic and economic conditions throughout the study period, we included zip code‐year level measures of unemployment and poverty in our regression models (Table S3).

It is ultimately impossible to entirely rule out potential violations of the identifying assumption. However, we examined whether the comparison group trends were an appropriate counterfactual for the treatment group by examining the extent to which treatment and comparison group trends were parallel in the pre‐period. The more parallel the trends in the pre‐period, the more confident we were that they would have remained parallel in the post period had the Green Line not opened. First, we regressed the no‐show outcome variable on a series of interaction terms between the treatment indicator and indicators for each quarter prior to the Green Line opening, omitting the last quarter prior the Green Line opening (January–March 2014) as the reference quarter. We also included clinic‐fixed effects in these models. We then tested whether the interaction terms for quarters prior to the Green Line opening were jointly equal to zero. Failing to reject the hypothesis that these interaction terms were equal to zero implied, we did not have evidence to contradict that the preperiod treatment and comparison group trends were the same.

However, a lack of evidence to reject the hypothesis that the preperiod treatment and comparison group trends were the same is not equivalent to having sufficient evidence to prove they were the same. On the other hand, statistically rejecting the null hypothesis of equivalent pre‐period trends would not mean that differential pre‐period trends had a practically meaningful effect on the estimated treatment effects. To address these limitations of the test for parallel pre‐period trends, we implemented the non‐inferiority approach suggested by Bilinski and Hatfield ³⁴ (Appendix B in Data S1). Specifically, we estimated our primary equations with added parameters to allow for differential pre‐period trends. We then compared the estimated treatment effects from these equations to the main equations, to assess whether allowing for differential pre‐period trends affected the estimated treatment effects. We expected the coefficients of interest to be similar if there was no violation of the parallel trends assumption. ³⁴

3. RESULTS

3.1. Descriptive statistics

Of the 3,572,963 appointments in our analytic sample, 300,525 of them (8.4%) were no‐shows. Approximately, 2.7% of these appointments were for patients who lived near the Green Line, and 11.4% of these appointments were no‐shows. Approximately, 25.3% of the appointments were scheduled to occur at one of the 14 clinics located near the Green Line, and 9.8% of these appointments were no‐shows. Table 1 presents appointment characteristics stratified by patient and clinic proximity to Green Line.

TABLE 1.

Appointment characteristics by proximity to new transit line

	Overall	Patient proximity to Green Line		Clinic proximity to Green Line
		Near	Not near	Near	Not near
Number of clinics	97	76	97	14	83
Number of scheduled visits	3,572,963	95,808	3,477,155	903,712	2,669,251
Number of unique patients	371,597	8353	363,244	76,936	309,858
Avg. probability of no‐show (%)	8.4	11.4	8.3	9.8	8.0
Patient characteristics
Female (%)	60.2	59.6	60.2	58.2	60.9
Age (mean (SD))	45.8 (23.5)	42.9 (21.7)	45.9 (23.6)	43.6 (22.6)	46.5 (23.8)
Race and ethnicity (%)
White, non‐Hispanic/Latino	79.6	56.5	80.2	67.3	83.8
Black, non‐Hispanic/Latino	9.3	24.5	8.9	16.3	7.0
Asian, non‐Hispanic/Latino	4.4	8.7	4.3	6.0	3.9
Hispanic/Latino	1.7	1.9	1.7	2.5	1.4
Other or unknown	5.0	8.4	4.9	8.0	4.0
Non‐English speaker (%)	5.5	12.2	5.3	8.9	4.4
Non‐US born (%)	10.0	17.2	9.8	14.4	8.5
Married (%)	44.5	28.4	45.0	37.0	47.1
Charlson comorbidity index (mean (SD))	0.57 (1.26)	0.52 (1.15)	0.57 (1.26)	0.74 (1.56)	0.51 (1.13)
Visit characteristics
Insurance type (%)
Commercial	42.9	35.2	43.2	37.3	44.9
Medicare	23.2	15.8	23.4	20.1	24.3
Medicaid	24.8	40.2	24.4	34.8	21.4
Uninsured	5.3	7.3	5.2	6.2	5.0
Other	3.8	1.6	3.8	1.6	4.5
Wait time in days (mean (SD))	21.0 (34.7)	21.8 (33.0)	21.0 (34.8)	34.4 (45.4)	16.5 (28.9)
Specialty department (%)	35.8	38.2	35.7	67.1	25.2
Provider type (%)
Physician	70.9	72.9	70.8	74.0	69.8
Nurse practitioner	8.3	7.4	8.3	5.3	9.2
Physician assistant	8.3	4.0	8.4	2.7	10.2
Other	12.6	15.7	12.5	17.9	10.8

There were some salient demographic differences between patients living near the Green Line and patients not living near the Green Line. Compared to patients not living near the Green Line, patients living near the Green Line were younger on average (42.9 years vs. 45.9 years); more likely to be non‐Hispanic black (24.5% vs. 8.9%) or non‐Hispanic Asian (8.7% vs. 4.3%); more likely to be non‐English speakers (12.2% vs. 5.3%) and non‐US born (17.2% vs. 9.8%) and less likely to be married (28.4% vs. 45.0%). Patients living near the Green Line were also more likely to be covered by Medicaid (40.2% vs. 24.4%) or to be uninsured (7.3% vs. 5.2%). These patterns were similar when comparing appointments at clinics near the Green Line versus not near the Green Line.

3.2. Effect of Green Line on no‐shows

Unadjusted probabilities of no‐shows by calendar quarter are depicted in Figure 2. Panel A shows no‐shows among all patients, while Panel B is limited to patients on Medicaid. In all four cases (all patients, patient proximity as treatment; all patients, clinic proximity as treatment; Medicaid patients, patient proximity as treatment; Medicaid patients, clinic proximity as treatment), the preperiod trends increased gradually but appeared parallel between the treatment and comparison groups. In the postperiod, the trends for all patients declined more for patients near the Green Line than patients not near it, and the trends for Medicaid patients declined more for patients and clinics near the Green Line compared to patients and clinics not near the Green Line. Unadjusted probabilities of no‐shows by calendar quarter for all nonurgent care appointments, including those that were canceled or excluded from the analytic sample for other reasons, reflect similar patterns (Figure S1).

FIGURE 2.

Unadjusted probability of no‐shows by calendar quarter and patient/clinic proximity to Green Line. The vertical line in each graph marks June 2014, when the Green Line opened to the public [Color figure can be viewed at wileyonlinelibrary.com]

Results from the DID models are provided in Table 2, and event study coefficients are depicted in Figure S2. The probability of no‐shows declined by a small but statistically significant amount after the Green Line opened for patients living near the Green Line relative to those not near it. Among all patients, the no‐show probability differentially declined by 0.5 percentage points (95% CI: −0.9 to −0.1) or 4.5% from the baseline probability of 11.1%. Among patients covered by Medicaid, the no‐show probability differentially declined by 1.6 percentage points (95% CI: −2.4 to −0.8) or 9.5% from the baseline probability of 16.9% (Table 2). The differential decline was concentrated among patients with appointments at a clinic near the Green Line (Table S4). There was also a small decline in no‐shows at clinics located near the Green Line relative to clinics not near it, but the change was not statistically significant; among all patients, the no‐show probability declined by 0.2 percentage points (95% CI: −1.1 to 0.6), and when limited to Medicaid patients, the no‐show probability declined by 0.9 percentage points (95% CI: −2.3 to 0.4) (Table 2). The differential decline in no‐shows at clinics located near the Green Line relative to clinics not hear it was larger and marginally significant among Medicaid patients living near the Green Line (−1.9 percentage points, 95% CI: −4.1 to 0.3) (Table S5).

TABLE 2.

Difference‐in‐differences results

Treatment	Preperiod probability of a no‐show for treatment group (%)	Differential change in probability of a no‐show (percentage points)
Panel A: All patients (N = 3,572,963)
Patient proximity to Green Line	11.1	−0.5** (−0.9, −0.1)
Clinic proximity to Green Line	9.2	−0.2 (−1.1, 0.6)
Panel B: Patients on Medicaid (N = 1,157,550)
Patient proximity to Green Line	16.9	−1.6*** (−2.4, −0.8)
Clinic proximity to Green Line	14.8	−0.9 (−2.3, 0.4)

Note: In addition to the treatment variable, defined as the interaction of the “post” period and patient or clinic proximity to the Green Line, models include clinic, patient's zip, and month fixed effects. 95% confidence intervals in parentheses.

**p ≤ 0.05; ***p ≤ 0.01.

3.3. Tests of identification assumption

The results from the F‐test for the joint significance of the pre‐period treatment and quarter interaction terms are provided in Table S6. For the models where clinic proximity to the Green Line was the treatment and the model where patient proximity was treatment and all patients were included (rows 1, 2, and 4), we did not reject the null hypothesis that the pretreatment interaction terms were jointly equal to zero, bolstering our confidence in the identifying assumption. For the model where patient proximity to the Green Line was the treatment and only Medicaid patients were included (row 3), none of the individual interaction term coefficients were significantly different from zero. The small p‐value, however, implied we could not assume that the pretreatment interaction terms were jointly equal to zero.

The results from the noninferiority approach are provided in Table S16. After adding a preperiod linear trend difference, the estimated treatment effects are similar (albeit larger in magnitude) compared to models that do not control for differential pre‐trends, both for all patients and among Medicaid patients only. These results further support the parallel trends assumption.

3.4. Additional analyses

Same‐day appointments, or appointments added to the schedule the same day that they occur, may be affected by public transportation availability similarly to no‐shows. Accordingly, we estimated the change in the expected probability of same‐day appointments after the Green Line opened for patients living near the Green Line (clinics located near the Green Line) relative to patients not living near it (clinics not located near it). The results show that after the Green Line opened, there was a statistically significant increase in same‐day appointments for patients living near the light rail compared to patients not living near it, and a statistically significant increase in same‐day appointments at clinics located near the light rail compared to clinics not located near it (Table S7).

In additional analyses, we tested whether the opening of the Green Line had an effect on no‐shows for patients and clinics near the Blue Line—the one other major light rail line in the Twin Cities—since there was an easy transfer in downtown Minneapolis between the two lines. Additionally, we estimated stratified models to assess the differential impact of patient proximity to the Green Line for patients with appointments at clinics near versus not near the Blue Line, and the differential impact of patient proximity to the Blue Line for patients with appointments at clinics near versus not near the Green Line. The results from these analyses suggest that the opening of the Green Line did not facilitate a reduction in no‐shows for patients near the Blue Line relative to those not near the Blue Line, even for patients traveling to appointments at clinics near the Green Line (Tables S8–S10).

In sensitivity analyses, we varied the definition used to categorize patients as Medicaid or not. Specifically, we tried requiring patients to have been stably on Medicaid for the entire period from 2013 to 2016, and we tried defining the Medicaid cohort based on patients who had Medicaid coverage during 2013. The results from these varied definitions were similar to our main results, which classified all patients with Medicaid coverage at any time during 2013–2016 as Medicaid (Table S11). Additionally, we modified the clinic proximity treatment variable to a half‐mile rather than one‐mile. The estimated effects were larger in magnitude than the primary model, as expected since the clinics closer to the light rail were “more exposed,” but they were not statistically significant (Table S12). Finally, we estimated specifications that included patient‐level‐fixed effects, and the results were directionally similar (Table S13). We also conducted heterogeneity analyses. First, we stratified the sample by primary versus specialty care and found the effect of the Green Line opening to be somewhat larger within specialty care (Table S14). Second, we stratified by morning versus afternoon appointments and found the effect to be concentrated among afternoon appointments (Table S15).

4. DISCUSSION

This study examined the relationship between improved public transportation and no‐shows, across more than 3 million appointments at 97 outpatient clinics. We found that after a large new light rail transit line opened, there was a small but statistically significant decline in no‐shows for patients living near the light rail compared to patients not living near it. The effect of the new light rail line was larger among patients covered by Medicaid—a group of patients especially likely to experience transportation barriers. ⁴ Among Medicaid patients living near the new light rail, the probability of a no‐show decreased by an additional 1.6 percentage points (or 9.5% relative to baseline) compared to Medicaid patients not living near the light rail. The effect was also concentrated among patients living near the light rail who were traveling to appointments at clinics located near the light rail, suggesting the effect of the new line was most salient for those whose route was direct. We found no significant change in the probability of no‐show appointments at clinics located near the light rail compared to clinics not near it. However, we found the rate of same‐day appointments to increase significantly after the light rail line opened for patients and clinics near the light rail compared to those not near it, suggesting the rail line promoted increased access to same‐day care in addition to reducing missed appointments.

While our estimated treatment effects were modest, even a small decrease in no‐shows benefits both patients and providers. For patients, completing appointments improves care continuity and avoids potentially harmful lapses in screening and treatment. This may be especially true for patients with chronic illness, who are also more likely to experience transportation barriers. ⁴ , ⁸ Outpatient care can help chronically ill patients access appropriate medications, achieve better disease control, and avoid future emergency department use or hospitalizations associated with their conditions. ² , ³ For providers, fewer no‐shows increases revenue and reduces scheduling inefficiencies.

There are several policy implications from our findings. Public transportation availability is considered a social determinant of health and is key for promoting health equity, but limited research directly examines the causal link between public transportation and access to health care services. ⁴ , ³⁵ By documenting a decrease in no‐show appointments and an increase in same‐day appointments following a public transportation expansion—especially for people with low incomes—we provide important new evidence on the importance of adequate public transportation to achieve equity in access to care. Metropolitan planning organizations should consider this evidence to help inform cost/benefit analysis on potential transit projects, such as when crafting proposals to expand or enhance transportation systems or developing Transportation Improvement Programs. ³⁶ Our finding that Medicaid patients miss fewer appointments following a public transportation expansion suggests Medicaid patients benefit from proximity to public transportation. Providing reimbursement for public transit through NEMT programs could further enable the use of public transit by Medicaid enrollees and reduce the frequency of missed appointments. When local transit systems improve or expand, as in the case of the Green Line in the Twin Cities, NEMT programs should ensure their coverage includes all public transit options and enables enrollees to benefit from transit system improvements. ³⁷

This study had several limitations. We were unable to observe the exact proximity of patients' residences to the light rail; we were only able to observe the most recent residential zip code reflected in each patients' record. Additionally, we did not have information on the actual mode of transportation that patients used to travel to their appointments. While the data we used reflected a rich source of information about no‐shows and the associated patient and visit characteristics, they represented only one health system in one metropolitan area. Further, the improvement to public transit in this study (i.e., the opening of a major light rail line) was also a unique event, making it difficult to compare the results to other changes to public transit. Finally, while we identified a fixed cohort of patients and used a DID study design to account for underlying differences between patients and clinics near and not near the Green Line that may have been correlated with the propensity for no‐shows, we were not able to entirely rule out the possibility that changes to individuals' economic mobility spurred by the new light rail line may have driven the results. For example, the opening of the Green Line may have enabled access to higher‐paying jobs, which may have in turn affected people's ability to miss work or afford their scheduled medical appointments.

Reliable transportation is essential for health care accessibility. Barriers to transportation contribute to socioeconomic disparities in health—particularly among patients with chronic disease who require frequent medical appointments. ⁴ This study empirically documented that improving public transit availability decreased the frequency of missed appointments, especially among patients covered by Medicaid. Our findings illustrate that improvements in public sector infrastructure can improve efficiency for health systems and promote equity in access to care.

Supporting information

Table S1. Sample inclusion criteria

Table S2. Difference‐in‐differences results, not restricted to patient clinics in pre‐ and postperiods

Table S3. Difference‐in‐differences results, with control variables

Table S4. Difference‐in‐differences results, patient proximity as treatment, stratified by clinic proximity to the Green Line

Table S5. Difference‐in‐differences results, clinic proximity as treatment, stratified by patient proximity to the Green Line

Table S6. Results from Ppeperiod parallel trends test

Table S7. Difference‐in‐differences results, same‐day appointment as outcome

Table S8. Difference‐in‐differences results, proximity to Blue Line as treatment

Table S9. Difference‐in‐differences results, patient proximity to Green Line as treatment, stratified by clinic proximity to the Blue Line

Table S10. Difference‐in‐differences results, patient proximity to Blue Line as treatment, stratified by clinic proximity to the Green Line

Table S11. Difference‐in‐differences results, varying Medicaid definitions

Table S12. Difference‐in‐differences results, clinic proximity to the Green Line defined using a half‐mile radius from the stations

Table S13. Difference‐in‐differences results, patient‐fixed effects

Table S14. Difference‐in‐differences results, stratified by primary versus specialty care

Table S15. Difference‐in‐differences results, stratified by morning versus afternoon appointment start time

Table S16. Difference‐in‐differences results, from the Bilinski and Hatfield in 2019 approach

Figure S1. Unadjusted probability of no‐shows by Calendar Quarter, all appointments

Figure S2. Results from event study specifications

Smith LB, Yang Z, Golberstein E, Huckfeldt P, Mehrotra A, Neprash HT. The effect of a public transportation expansion on no‐show appointments. Health Serv Res. 2022;57(3):472-481. doi: 10.1111/1475-6773.13899