Article Text

Transportation characteristics associated with non-arrivals to paediatric clinic appointments: a retrospective analysis of 51 580 scheduled visits
  1. David J Wallace1,
  2. Kristin N Ray2,
  3. Abbye Degan2,
  4. Kristen Kurland3,
  5. Derek C Angus1,
  6. Ana Malinow2
  1. 1Critical Care Medicine, University of Pittsburgh School of Medicine, Pittsburgh, Pennsylvania, USA
  2. 2Pediatrics, University of Pittsburgh School of Medicine, Pittsburgh, Pennsylvania, USA
  3. 3School of Architecture, Carnegie Mellon University, Pittsburgh, Pennsylvania, USA
  1. Correspondence to Dr David J Wallace, Critical Care Medicine, University of Pittsburgh School of Medicine, Pittsburgh, PA, USA; wallacedj{at}


Background Prior work has not studied the effects of transportation accessibility and patient factors on clinic non-arrival.

Objectives Our objectives were: (1) to evaluate transportation characteristics and patient factors associated with clinic non-arrival, (2) to evaluate the comparability of bus and car drive time estimates, and (3) to evaluate the combined effects of transportation accessibility and income on scheduled appointment non-arrival.

Methods We queried electronic administrative records at an urban general pediatrics clinic. We compared patient and transportation characteristics between arrivals and non-arrivals for scheduled appointments using multivariable modeling.

Results There were 15 346 (29.8%) clinic non-arrivals. In separate car and bus multivariable models that controlled for patient and transit characteristics, we identified significant interactions between income and drive time, and clinic non-arrival. Patients in the lowest quartile of income who were also in the longest quartile of travel time by bus had an increased OR of clinic non-arrival compared with patients in the lowest quartile of income and shortest quartile of travel time by bus (1.55; P<0.01). Similarly, patients in the lowest quartile of income who were also in the longest quartile of travel time by car had an increased OR of clinic non-arrival compared with patients in the lowest quartile of income and shortest quartile of travel time by car (1.21, respectively; P<0.01).

Conclusions Clinic non-arrival is associated with the interaction of longer travel time and lower income.

  • ambulatory care
  • health services research
  • primary care

Statistics from


Clinic non-attendance is a significant problem in the USA, with average no-show rates near 30% and varying from 3% to 80%.1 Missed paediatric appointments have an impact not only through undelivered acute healthcare and preventative services,2 but additionally create financial strains on clinics,3 4 and may be associated with increased use of emergency medical services.5–7 Self-reported factors related to paediatric clinic non-attendance include older patient age, well-child appointment status, parent forgetfulness8 and insurance status.1 Although transportation in general is cited as a reason for missing appointments, a more detailed understanding of how transit affects attendance is important for addressing this barrier.

Geographic information systems analysis provides a means of evaluating access to healthcare resources that incorporates driving time into access calculations.9 10 These analyses are capable of showing the predicted population around a particular clinic or hospital that is capable of reaching the facility in a specified amount of travel time. An important and common assumption of these models, however, is that a patient’s mode of transportation is by car and that the path is the most direct. This approach may result in erroneous estimates of actual access for patients who use public transportation to travel to clinic.

Our analysis had three objectives: evaluate transportation characteristics and patient factors associated with clinic non-arrival, compare bus and car drive time estimates from home to clinic, and develop a multivariable model to understand the relationships between various factors and patient non-arrival. We hypothesised that longer transport times would be associated with patient non-arrival, that bus and car drive time estimates would not be directly or proportionately equivalent measures, and that the effect of transport time on non-arrival would be modified by patient income level.


We obtained electronic records for patients aged less than 21 years old who were scheduled to attend an academic general paediatrics clinic in Pittsburgh, Pennsylvania, for the years 2014 and 2015. The clinic provides a range of services, including well-child exams, vaccinations, preventative care, acute care, newborn care and lactation counselling, and basic dentistry. Free parking is provided for patients or their driver. The clinic schedules approximately 30 000 visits each year and reports an approximately 30% non-arrival rate. Pittsburgh is located in Allegheny County, the 31st most populous county in the USA. In 2010, the Allegheny County population was 81.5% white, 13.2% black, 2.8% Asian and 2.5% Other, compared with the entire US population, which was 72.4% white, 12.6% black, 5.6% Asian and 9.4% Other.11

In 2015, the Port Authority of Allegheny County reported 54 843 567 bus trips,12 and in 2013 the city of Pittsburgh ranked 27th out of 290 cities in the USA according to the annual number of public transportation trips per city resident (37.8).13 During the years of the study, bus fare was based on a three-zone system and rider age, with longer one-way adult trips costing $3.75, moderate one-way adult trips costing $2.50 and trips inside the downtown area being free. Children under 6 years old rode for free and children 6–11 years old rode at half-price. A maximum of four children age 5 and under rode for free with one fare-paying passenger—additional children under 6 years of age rode for half-fare. The paediatrics clinic was not located in the downtown area; patients and accompanying family members incurred charges for using bus transportation to clinic appointments. The payment system was simplified in 2016.

To evaluate patient factors associated with clinic non-arrival, we obtained patient age, sex, race and ethnicity directly from the electronic record. We classified patient age into four categories: less than 1 year, 1–5 years, 6–12 years, and 13 years or older. We combined self-reported patient race into four categories: black, white, Asian and Other. Ethnicity was reported in the electronic record in three categories: non-Hispanic or Latino, Hispanic or Latino, or not specified or declined. We obtained the median income for each patient ZIP code using 2014 US Census Bureau estimates,11 and classified ZIP codes in quartiles of the overall distribution, merging the intermediate two quartiles to simplify interpretations (less than $26 625, $32 118–$42 604 and $45 738–$51 994). We classified appointment month into four seasons: summer (June, July or August), fall (September, October or November), winter (December, January or February) and spring (March, April or May). We classified appointments into well-child and not well-child using administrative codes. We excluded unscheduled walk-in visits using codes from the administrative record.

To evaluate transportation characteristics associated with clinic non-arrival, we performed geospatial analyses using an offline ArcGIS package and Google’s travel calculator ( We first calculated the geographical centre of each patient’s home ZIP code in ArcGIS and entered this location and the street address of the paediatrics clinic into Google’s travel calculator to obtain both car driving and public transportation transit times. We recorded the cumulative number of miles of walking required to use the bus for transportation (including walking to the bus depot, between bus depots if applicable and from the final bus depot to the paediatrics clinic) and the number of transfers between buses. We performed travel calculations on a weekday at 09:00 to capture regular work hour traffic patterns. We used the geographical centre of the ZIP code for privacy concerns (Google stores online travel requests) as well as practical reasons (characterising transport variables for the geographical centre of each ZIP code required fewer searches than individual patient home addresses). We created categorical variables based on these calculations using quartile distributions, merging the intermediate two quartiles to simplify interpretation. We limited the analysis to ZIP codes with more than 400 scheduled appointments each year. We performed this restriction to avoid false inference resulting from small area variation effects.

Pursuant to our first objective, we performed univariate comparisons of patient factors and transportation characteristics for appointment arrivals and non-arrivals using Χ2 statistics. We performed this analysis to evaluate individual associations with non-arrival status and for subsequent comparison to results from our multivariable model.

For our second objective, to examine the agreement between car and bus travel time estimates across a range of values, we generated a Bland-Altman plot for the estimated car and estimate bus travel time from each ZIP code to the paediatrics clinic. This test has advantages over the simple correlation coefficient, as it can identify if and where car and bus estimates could be used in substitution of each other. We additionally directly compared agreement between transportation estimates at each ZIP code location for each transportation method.

For our third objective, to examine the impact of travel time accounting for other variables, we used generalised estimating equations to evaluate the association between clinic non-arrival and predictor variables. We analysed the entire cohort separately using two assumptions: (1) all transportations would occur by bus and (2) all transportations would occur by car. We used these two approaches as we did not know the individual transportation method used (or intended to be used) for each scheduled appointment. We used generalised estimating equations specifically because this approach allowed us to account for clustering of values at the patient level, as many patients had more than one scheduled clinic appointment. We used an exchangeable correlation matrix, binomial family distribution and robust SEs. We performed testing for first-order interactions between variables and testing for violations of regression assumptions. We planned a priori to include significant transportation characteristics in the model and control for demographic characteristics.

As a sensitivity analysis, we additionally created a simplified version of our primary analysis for bus transportation characteristics that did not include cumulative walk distance or the presence of a bus transfer. We performed this analysis to determine the strength of association between appointment non-arrival and longer bus transits, without requiring model components that may be more difficult to obtain. This would allow clinic administrators to potentially more directly apply the results.

In a post-hoc analysis, we evaluated the association between appointment non-arrival and transportation time, stratified within ZIP code income groups. For these multivariable analyses we used the same covariates as in the primary analyses but varied the base comparison group. We performed these tests to isolate the effect of transport time within ZIP code income groups.

We considered all statistical tests to be significant at the alpha less than 0.05 level. Analyses were performed on Stata V.13.0 and ArcGIS V.10.4 (Redlands, California, USA).


We studied a total of 51 580 scheduled patient appointments from 20 ZIP codes (table 1). Appointments were scheduled for 11 812 individual patients, with a median of 3 scheduled appointments per patient (25–75th IQR 2–6). Over the period of analysis, scheduled appointments ranged from 1 (n=2281) to 42 (n=1) per patient. There were 10 930 patients (92.5%) who had 10 or fewer scheduled clinic appointments, and 8617 patients (73.0%) had 5 or fewer scheduled clinic appointments (online supplementary figure 1). Scheduled appointments included patients across a range of ages and were balanced with regard to patient sex (48.4% female). The majority of appointments were from patients who identified as black (83.5%) and non-Hispanic (93.7%). Scheduled appointments were similar across seasons.

Supplementary file 1

Table 1

Characteristics of appointments (n=51 580)

Overall there were 15 346 (29.8%) patient non-arrivals to scheduled appointments for 4581 patients. In our first objective analysis, non-arrivals varied by patient age, with the lowest percentage of non-arrivals occurring in the age group 13 years and older (13.4%) and highest in patients 1–5 years old (41.3%, P<0.01 for Χ2; table 2). The median ZIP code income was associated with non-arrival, with a higher percentage of non-arrivals compared with arrivals in the lowest quartile of income (23.5% and 22.7%, respectively) and a higher percentage of arrivals compared with non-arrivals in the highest quartile of income (16.1% and 13.4%, respectively; P<0.01 for Χ2). Non-arrivals were more likely to be well-child appointment types (P<0.01). Patient non-arrivals were not associated with estimated travel times by car (P=0.26), and were less frequent with longer bus transit times (the longest quartile travel time was associated with lower clinic non-arrival; 18.4% compared with 19.3%, P<0.01). Longer cumulative walking distance with bus transit was associated with lower clinic non-arrival (P<0.01), and the necessity of a bus transfer was associated with a lower frequency of clinic non-arrival (31.0% compared with 32.4%, P<0.01).

Table 2

Characteristics of appointments by arrival status

Our second objective analysis produced a Bland-Altman plot of car drive times compared with bus drive times that did not show consistent bias, as longer travel times showed increasing differences between the estimates (figure 1). In addition, many estimates fell outside a 2 SD limit of agreement, indicating that even regression-corrected substitutions of car values would lack accuracy as bus estimates. In a direct comparison of transportation estimates from each ZIP code starting point, we found that bus and car estimates only categorically agreed for 10 ZIP codes (47.6%; online supplementary table 1). Additionally, estimates that were intermediate transport time by bus were equally likely to be shortest (n=2) or longest (n=2) by car, as they were to be considered intermediate by car (n=4).

Figure 1

Bland-Altman plot comparing estimated drive travel time with estimated bus travel time from the geographical centre of the ZIP code to clinic. The dashed black line shows the mean bias between the two estimates. The solid black line shows the fitted regression line for the difference and average of estimate pairs. The solid grey lines show the 2 SD limits of agreement.

For our third objective, we created multivariable models of transport characteristics and patient factors associated with clinic non-arrival. We identified significant effect modification between ZIP code income and transit time for both car (Χ2=210.2, P=0.04) and bus (Χ2=223.7, P<0.01) estimates. We therefore included an interaction term between income and travel time in both generalised estimating equation models. In our final adjusted multivariable model, patients living in the lowest income quartile ZIP codes were more likely to not arrive to clinic at the longer car transit times (OR: 1.21 for longest drive quartile, P<0.01, compared with patients in the lowest income ZIP codes and shortest car transit times group; figure 2, online supplementary table 2). This relationship was not observed in intermediate or high quartile income ZIP codes, with the exception of appointments from the highest income ZIP codes with the shortest car travel time, which had lower odds of non-arrival (OR: 0.84, P=0.049, compared with patients in the lowest income ZIP code and shortest car transit times group). All model parameters included for adjustment were statistically significant and matrix specification was good.

Figure 2

OR of not arriving to scheduled clinic appointment by quartile of median ZIP code income and direct drive time. Model controls for season, patient race, patient ethnicity, patient age and appointment type. Significant comparisons with P<0.05 are indicated with an asterisk. The reference category, lowest quartile income and lowest quartile drive time, is indicated with ‘ref’.

We identified similar associations for bus transit, with increased odds of clinic non-arrival for those in the lowest quartile income ZIP codes with intermediate and longest travel times (OR: 1.17 and 1.54, respectively, P=0.01 and P<0.01; figure 3, online supplementary table 3). Increased travel time was less strongly associated with non-arrival for intermediate income ZIP codes; in highest income ZIP codes, shortest travel time was associated with lower odds of clinic non-arrival (OR: 0.84, P=0.049). This model additionally contained terms for cumulative walk distance and bus transfer. All remaining parameters were statistically significant and matrix specification was good.

Figure 3

OR of not arriving to scheduled clinic appointment by quartile of median ZIP code income and bus travel time. Model controls for season, patient race, patient ethnicity, patient age, appointment type, cumulative walking distance in bus transit (ie, walking distance to depot, any walking between transfers and walking from final depot to clinic) and presence of a bus transfer. Significant comparisons with P<0.05 are indicated with an asterisk. The reference category, lowest quartile income and lowest quartile drive time, is indicated with ‘ref’.

The results of our sensitivity analysis for bus transportation characteristics were similar to the primary bus analysis (online supplementary table 4). Increased bus travel time was associated with appointment non-arrival in lowest income ZIP codes (OR: 1.23, P<0.01), but not in intermediate or highest income ZIP codes (ORs: 0.99 and 0.92, P=0.86 and 0.16).

The results of our post-hoc analysis within ZIP code income groups were similar to the primary analysis (online supplementary tables 5 and 6). Increased bus travel time was associated with appointment non-arrival in the lowest income ZIP group. In intermediate-income and high-income ZIP groups, intermediate transport time was associated with increased ORs for appointment non-arrival but not for longest transit times (each compared with the shortest transit time within the income strata). For car estimates, the association between increased OR of appointment non-arrival and increased transport time was only observed in the lowest income ZIP codes group.


In a large cohort of scheduled appointments at an academic general paediatrics clinic over a 2-year period, we found that the combination of longer travel distance and lower income was significantly associated with an increased odds of clinic non-arrival in a multivariable model that controlled for patient and appointment characteristics. Further, we demonstrated that car and bus travel estimates are not sufficiently correlated across their range of values to allow substitution. Finally, we did not find consistent associations between clinic non-arrival in univariate and multivariable modelling, indicating that travel time estimates alone are insufficient for accurate non-arrival risk assessment. Together these findings have important implications for efforts to study and improve clinic scheduling and attendance.

We first demonstrated that car travel time estimates are frequently not substitutable for bus travel time estimates. Bus time estimates were often longer than car estimates, but they did not have a consistent bias (ie, a simple multiplier or addition could not convert car estimates to bus estimates, or vice versa). This makes sense, as bus routes have variability, both in the number of stops for passengers and in the concordance between the bus route and the most direct path to clinic by car. As such, accessibility that is based on car routing from home to the site of medical care will often underestimate access for patients who use public transportation, and will underestimate access to different degrees for different patients14 (based on their actual mode of transportation). In a recent review of 61 studies on transportation barriers to healthcare access, distance to clinic showed mixed evidence on impact to access, likely due to the differences in travel burden to patients.15 Greater distance alone, therefore, does not necessarily translate to greater burden if the patient has access to a car. Finally, in cities that have dedicated bus lanes, the difference between car and bus transit estimates could err in favour of faster bus transit times. Although Pittsburgh has a rapid bus-only connector to the downtown area, this path would logistically not service patients coming to clinic appointments, and therefore did not factor into the analysis.

The results of the univariate analysis differed in important ways from our final fully adjusted model. In univariate analysis, longer car drive time was not associated with clinic non-arrival, and longer bus transit time was associated with lower frequency of non-arrival. Considering either factor alone will therefore paint an inaccurate picture of risk for appointment non-arrival. We suspect this reflects the complexity of factors that go into missing a scheduled appointment, which is more than simply the travel time. Many patients had both arrivals and non-arrivals, reinforcing this theory. Therefore, researchers and clinicians should not use travel time alone to identify appointments at high risk of non-arrival.

The interaction finding from our fully adjusted model likely resulted from the relationships between neighbourhood locations and median income, with some higher income neighbourhoods being located farther from the paediatrics clinic. Parents with higher incomes have higher vehicle ownership rates,16 17 more resources to help coordinate transportation and decreased reliance on public transportation. Parents with lower incomes may have other factors, such as less predictable employment schedules or less paid sick leave, contributing to an additive burden along with longer travel times.18 A singularly univariate approach misses the interplay between these factors and would have led to false inference between travel time and clinic attendance.

Our multivariable models provide new insights into the interplay between factors associated with missed clinic appointments. Recent work has focused on reminder-based interventions to improve clinic attendance.19–21 While clinic reminders address a common barrier to attendance, they will have a more muted impact on patients who miss appointments for logistical reasons. Future work should consider expanding reminders to include questions about anticipated transportation difficulties, creating opportunities for just-in-time interventions or focused Medical Assistance Transportation Program. Improved transit time estimation can also be used to adjust appointment times based on daily traffic patterns, mitigating some personal time cost for arriving patients and families. Finally, for geographical regions where public transportation access and income predict a high risk of clinic non-attendance, healthcare systems should consider mobile clinics,22 23 school-based health centres24 25 and telemedicine clinic programmes.26 27

We identified other important factors associated with appointment non-arrival in our multivariable model. Well-child visits were associated with 56% higher odds of non-arrival, even after adjusting for other patient factors and transport characteristics. Patient age was also strongly associated with non-arrival, with all age ranges greater than 1 year of age associated with increased odds ratio of non-arrival, compared with the youngest age category. We speculate that higher clinic attendance rates in infancy are likely related to standardised immunisation schedules. In this age range, parents’ desire to keep their children up to date on immunisations, or daycare requirements for such, may be responsible for lower rates of non-arrivals.

Our analysis has limitations. First, we did not include appointments from all ZIP codes scheduled at the clinic. We performed this restriction to avoid estimates from ZIP codes with few scheduled appointments overall, as these regions would have the potential for more extreme estimates of effect from small numerator changes. However, in restricting these regions from the analysis, it is possible our findings may not generalise to areas with smaller paediatric populations. We feel this intentional trade-off improved the validity of our findings, and furthermore that issues related to access in remote or rural populations should be the focus of a dedicated analysis. Second, we did not know how patients who arrived were actually transported to clinic, or how patients who did not arrive anticipated travelling when they scheduled their appointments. Despite the absence of this information, we still identified an important signal related to clinic non-arrival. Clearly this information could be obtained through an outpatient paediatrics clinic, and we feel it should be included in efforts to individually apply our findings. We suspect that the magnitude of effect we observed would be even stronger with this information. Third, we did not use actual patient home addresses. We used the geographical centre of patient ZIP codes to approximate their home address for two reasons: practicality and privacy. Despite this limitation we were still able to observe an association between travel time, income and non-arrival to clinic. Prior work has shown that ZIP code-based geolocation performs similarly within and between states compared with census tract or block group, although it tends to detect smaller effects.28 The simplified approach we employed has the advantage of being relatively straightforward to apply, as it did not rest on a brute-force approach to travel time estimation. However, in areas with substantially heterogeneous ZIP code demographics or characteristics, analyses that use census tracts or block groups may be necessary to accurately identify smaller cohorts of at-risk populations. Finally, we analysed all scheduled appointments, which meant that there were multiple scheduled appointments for many patients. Patients with both arrivals and non-arrivals could bias associations towards the null, potentially attenuating the magnitude of true underlying associations.


Longer travel times by bus or car from lower income ZIP codes are associated with increased odds of clinic non-arrival. Efforts to improve clinic attendance should consider these geographical areas as high risk for missed appointments and target interventions to overcome travel time and resource-based barriers to care.


View Abstract


  • Handling editor Kaveh G Shojania

  • Contributors All authors contributed to study design, analysis, interpretation of findings and manuscript development.

  • Funding The article is funded by Agency for Healthcare Research and Quality (K12HS022989) and National Institutes of Health National Heart, Lung and Blood Institute (K08HL122478).

  • Competing interests None declared.

  • Ethics approval This project was approved by the human subjects research board at the University of Pittsburgh School of Medicine.

  • Provenance and peer review Not commissioned; externally peer reviewed.

  • Data sharing statement The primary data for this study were an electronic record of scheduled paediatric clinic appointments and attendance. For the years of evaluation, with the exception of specified encounter exclusions, we analysed the entire cohort. There are no unpublished data from the study.

Request permissions

If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.

Linked Articles