Determinants of Economic Costs Following Road Traffic Injuries in Canada: A Quantile Regression Forests Machine Learning Approach

Abstract

Introduction

This study aimed to identify major determinants of the cost of road traffic (RT) injuries, rank their importance, and assess their effects on different quantiles of cost distribution.

Methods

This study analyzed data collected from 1372 Canadian RT survivors from July 2018 to March 2020. Costs, including healthcare and lost productivity costs over a year following RT injury, were estimated for each participant in 2023 Canadian dollars. Productivity loss was measured using the Institute for Medical Technology Assessment Productivity Cost Questionnaire. We considered 24 potential determinants of costs, which were grouped into five domains: sociodemographic, psychological, health, crash, and injury factors assessed during baseline interview. We employed a quantile regression forests machine learning approach alongside classical quantile regression to analyze costs. These methods were selected to capture heterogeneous effects across cost distribution, which are overlooked by traditional mean-based models, and to inform policy decisions targeting high-cost subgroup.

Results

The results showed that the 10th, 50th, and 90th quantiles of costs were $1,141.9, $7,403.1, and $49,537.5, respectively. ISS, GCS, and age were the top three influential variables among low-cost, medium-cost, and high-cost patients. ISS, GCS, age, sex, employment status, and living situation were common major determinants at all quantiles. Ethnicity was selected as an important determinant at the 50th and 90th quantiles. Education level, years lived in Canada, somatic symptoms severity, psychological distress, HRQoL, road user type, and head, torso, spine/back, and lower extremity injuries were selected only for high-cost patients (90th quantile). Classical quantile regression showed that selected major predictors disproportionately affected low-cost, middle-cost and high-cost patients.

Conclusion

High-cost patients were more likely to be older, retired, less educated, and have worse clinical and psychological indicators. These insights can guide targeted prevention and resource allocation strategies to reduce the economic burden of RT injuries.

Introduction

Road traffic (RT) is a serious problem that accounts for 150,000 injuries per year in Canada, including over 1,900 deaths and 9,400 serious injuries.¹ The annual cost of RT injuries in Canada was estimated at $3.6 billion in 2018.² The reduction in non-fatal traffic injuries has been less significant than the reduction in fatal injuries in recent years, indicating a need for greater attention to non-fatal injury prevention.^3–5 Non-fatal injuries have a significant impact on economic outcomes, such as long-term disability, early retirement, property loss, and the utilization of healthcare resources.^6,7

The cost of RT injuries is a key metric for measuring the societal burden of RT injuries. The economic costs of non-fatal injuries are significant, encompassing both healthcare costs and lost productivity. Healthcare costs are determined by factors such as emergency department (ED) visits, hospital admissions, medications, and rehabilitation costs. The costs of lost productivity are determined by absenteeism (time off work) and presenteeism (reduced productivity at work) in both paid and unpaid work.^8,9 Currently, there are limited studies on the costs of RT injuries,^7,10–21 and insufficient information regarding the long-term economic costs associated with RT injuries.^7,11–13 Furthermore, many studies concentrate solely on healthcare costs while overlooking the expenses associated with lost productivity.^10,11,14,15

The economic cost data are typically distributed with a high level of skewness;²² see Figure 1 for economic cost distribution in our data. The magnitude or direction of a factor’s effect on economic costs could be different in low-cost patients than in high-cost patients. Therefore, the assumption of a uniform relationship between covariates and cost across different percentiles of the cost distribution does not always hold. Also, identifying determinants of economic costs at higher quantiles is more interesting for policymakers and insurance plans as high-cost patients can then be targeted for cost containment interventions that yield greater potential savings. Quantile regression is a relatively new statistical model in the field of healthcare cost analysis. It requires no assumptions about the distribution of outcome variables and is robust to outliers. Quantile regression is able to estimate how specified quantiles of the distribution of outcomes vary with covariates. This model can also discover differences in the upper or lower tails of the distribution as coefficient estimates are generated by weighting different portions of the sample.^23,24

Figure 1.

Histogram of economic costs of participants.

Previous studies have reported that older age, female sex, higher injury severity, and more comorbidities are associated with higher healthcare costs.^13,14 Male sex and higher injury severity are associated with higher long-term lost productivity costs.¹³ Injury severity, admission to the intensive care unit, and length of hospital stay are associated with hospitalization costs.¹⁰ The cost of motorcycle, bicycle, and pedestrian injuries is relatively higher than those of other road users.^10,11,13,25 Research has found that injuries to the spine, lower extremities or head are associated with higher healthcare costs.^14,16

Building on this body of evidence, we previously conducted a study that identified predictors of healthcare and productivity costs separately for all survivors of RT injury, including low-, medium-, and high-cost patients, using traditional statistical model.²⁶ In that work, factors such as older age, lower health-related quality of life (HRQoL), and higher injury severity were reported as significant predictors for both healthcare and productivity costs. The present study extends this earlier work in several important ways. First, instead of analyzing healthcare and productivity costs separately, we focused on total costs, encompassing both healthcare and productivity costs, to provide a more comprehensive understanding of the overall economic burden. Importantly, in contrast to our prior work, which was limited to employed survivors for calculating productivity costs, the present study estimated productivity costs for all patients. Second, we applied an advanced machine learning method—quantile regression forests (QRFs)—to capture heterogeneity in cost determinants across the distribution of total costs. Specifically, we aimed to (1) identify key predictors of economic costs across low-, medium-, and high-cost patients and (2) rank the relative importance of these predictors for each subpopulation. In a subsequent step, we complemented this with classical quantile regression to examine how the selected predictors were associated with costs at different levels of the cost distribution. Our aim is to support traffic policymakers in evaluating the effectiveness of current policies, identifying priority areas for targeted prevention strategies, improving existing interventions, and guiding the development of new policies aimed at reducing the economic burden of RT injuries.

Methods

Participants and Procedures

In this study, individuals who survived RT accidents were enrolled from three British Columbia (BC) emergency departments (ED) (Vancouver General Hospital, Vancouver; Royal Columbian Hospital, New Westminster; and Kelowna General Hospital, Kelowna) between July 2018 and March 2020. Detailed methods of the study have been previously outlined.²⁷ To be eligible for the study, RT survivors needed to i) be at least 16 years old and ii) have entered the ED within 24 hours of a collision involving a motorized vehicle. Individuals who were not residents of BC and those who died within 30 days of the accident were excluded. The present study analyzed data collected for five different road user types (cyclists, pedestrians, motorcyclists, and motor vehicle drivers and passengers) and all injury severity levels. Collisions involving lone cyclists, e-bike users without vehicle involvement, and pedestrian-only incidents were excluded from the study. Data were gathered through interviews and reviews of medical records. For individuals who did not speak English, interviews were conducted in Cantonese, French, Korean, Mandarin, Punjabi, and Vietnamese by a multilingual research assistant or through a translator. Participants provided informed consent, and the study, including the informed consent process, was approved by the University of British Columbia Clinical Research Ethics Board (H18-00284). All procedures were conducted in accordance with the Declaration of Helsinki.

Outcome

The outcome was total economic costs, including estimates of healthcare costs and productivity costs over a year following RT. Data on healthcare resource use over a year following RT were obtained from Population Data BC (PopDataBC), a data repository containing individual-level longitudinal health records for all residents of BC.²⁸ De-identified patient data are compiled from several datasets linked by the BC Personal Health Number. These include the Discharge Abstracts Database, which contains administrative and clinical data on patients admitted to hospital and also captures data on day surgery, long-term care, rehabilitation and other types of care; the National Ambulatory Care Reporting System, which contains ER visits and ambulance usage; the Medical Services Plan (MSP) Payment Information File, which contains data on medical services provided by practitioners in the province; and PharmaNet, which contains all prescriptions for drugs and medical supplies dispensed from community pharmacies in BC as well as prescriptions dispensed from hospital outpatient pharmacies for patient use at home. The Resource Intensity Weight (RIW) method was used to estimate hospital costs. RIWs are the relative case weights for Case Mix Groups (CMG) or the Day Procedure Group (DPG)/Comprehensive Ambulatory Care Classification System (CASC) used to measure the intensity of resource use associated with different diagnostic, surgical procedure, and demographic characteristics of an individual.^29,30 We calculated the healthcare cost per case by multiplying the case-specific RIW by the cost per weighted case provided by the Canadian Institute for Health Information (CIHI).

Productivity costs were measured by the Institute for Medical Technology Assessment Productivity Cost Questionnaire (iPCQ).³¹ This validated questionnaire adopts a four-week recall period. To capture productivity loss over a year following RT injury, we administered this questionnaire at 2, 4, 6, and 12 months post-injury. Participants were also asked about their productivity 4 weeks prior to the accident using the iPCQ. The iPCQ includes three modules measuring productivity losses of paid work due to i) absenteeism and ii) presenteeism and iii) productivity losses related to unpaid work. This questionnaire also includes general questions regarding participants’ work status such as the weekly hours of paid work and the number of working days per week. The absenteeism module identifies the occurrence and length of short-term and long-term absenteeism from paid work. Absences beginning before the recall period are considered long-term absences. The presenteeism module identifies reduced productivity at work. In the unpaid work module, participants are asked if they performed less unpaid work because of physical or psychological problems. The productivity losses for the recall period were adjusted for the length of the period between two consecutive measurements using an extrapolation technique. The human capital approach (HCP) was used to calculate productivity costs for a year following RT injury.³² The costs of productivity were calculated by multiplying the missed work or lost productivity hours by the average hourly wage rate obtained from Statistics Canada’s CANSIM database for a person in the same age group, sex, and interview time in BC.³³ The costs of unpaid productivity losses were calculated by multiplying the total hours of lost productivity by the average hourly wage for household care at the interview time in BC.³⁴ The missing numbers of days lost or lost productivity were assumed to be equal to zero. If the number of paid hours per week that a participant worked was missing, the national mean based on sex was used, 38 hours per week for males and 34 hours per week for females.³⁵ Absenteeism and presenteeism losses in paid work were considered to be zero for participants who reported pre-injury absence from work, were not in the working population (aged 15–64), or did not have paid employment. All costs were adjusted for inflation to 2023 CAD dollars using the Statistics Canada Consumer Price Index.³⁶

Predictor Variables

We considered 24 potential predictors at the baseline interview for economic costs in the year following RT injury which can be grouped into five domains including sociodemographic, psychological, health, crash, and injury factors. Sociodemographic factors included age, sex, employment status, living situation, level of education, ethnicity, and years living in Canada. Psychological factors comprised pre-injury general somatic symptom severity, pain catastrophizing, psychological distress. HRQoL and the number of pre-injury comorbidities were the health factors measured at baseline. Crash and injury factors included ED visits, ED discharge disposition, crash time, road user type, Injury Severity Score (ISS),^37,38 Glasgow Coma Scale (GCS),³⁹ and injury location. Pre-injury somatic symptom severity, pain catastrophizing, and psychological distress were measured with the Patient Health Questionnaire-15 (PHQ-15),⁴⁰ the Pain Catastrophizing Scale,⁴¹ and the Patient Health Questionnaire-4 (PHQ-4),^42,43 respectively. Higher scores on these scales indicate greater symptom severity, higher levels of catastrophic thinking, and more severe depression and/or anxiety, respectively. Pre-injury HRQoL was evaluated using the Short Form 12 survey (SF-12),⁴⁴ where higher scores indicate better health. Comorbidities included eye disease, arthritis, diabetes, respiratory disease, heart disease, hypertension, stroke, epilepsy, kidney disease, psychiatric disorders, and other diseases self-reported by participants during the interviews.

Statistical Analysis

We first explored the presence of multicollinearity among predictors using the Variance Inflation Factor (VIF) to identify highly collinear predictors.⁴⁵ At this stage, the variable of ED discharge disposition was deleted from the analysis of this study. A nonparametric machine learning technique, QRFs, was performed to identify determinants of economic costs across different quantiles of its distribution. QRFs are a generalization of random forests (RFs).⁴⁶ The RFs algorithm can be viewed as a meta-algorithm that outlines a method for constructing models using a model builder, such as decision trees.⁴⁷ The RFs with a decision tree algorithm construct multiple decision trees to assess the relationship between the conditional mean of the response and predictors through a technique known as bagging which involves randomly sampling observations into a bag as the training dataset to build each decision tree. Generally, about two-thirds of the observations are selected to be included in the bag, while the remaining, known as the out-of-bag (OOB) observations, serve as the testing sample for performance evaluation. QRFs grow an ensemble of decision trees as in the standard RFs algorithm to estimate the conditional quintiles of the response in relation to predictors. The key difference between QRFs and RFs is that for each node in each tree, QRFs keep the value of all observations that fall into the node, whereas RFs only keep the mean of those observations and neglect all other information. A backward stepwise variable selection algorithm developed by Hu et al⁴⁸ was implemented to select major determinants for the 10th, 50th, and 90th quantiles as the threshold for the low-, medium- and high-cost patients, respectively.⁴⁸ Figure 2 illustrates this iterative algorithm. At each step, the importance score for each potential predictor in the QRFs model on the OOB sample is calculated. The importance score for each predictor is measured based on the decrease in accuracy of the QRFs model when the predictor’s values are randomly permuted on the OOB sample. The predictor with the least importance score is removed and a QRFs model is rebuilt with the remaining predictors. Also, at each step, the OOB average quantile loss (AQL) is recorded until no predictor is left in the model. The AQL, as proposed by Wang et al,⁴⁹ is used as the measure of prediction accuracy of τ-th conditional quantile (eg, τ = 0.9). A set of major determinants for τ-th conditional quantile of economic costs is identified as the set of predictors in the QRFs model corresponding to the elbow point. This point achieves the optimal balance between minimizing OOB AQL and selecting a parsimonious set of predictors. Finally, a classical quantile regression was used to quantify the effects of selected major determinants for the 10th, 50th, and 90th quantiles by QRFs on the quantile of economic costs. The results of multivariable quantile regression were presented as β with 95% confidence intervals. Diagnostic criteria were evaluated to assess the robustness of the quantile regression model. Pseudo R² values across quantiles indicated moderate to strong model fit. Additionally, Rank Score Tests were statistically significant at all quantiles, suggesting that the covariates collectively explain variation in costs across the distribution and supporting the overall adequacy of the model specification.

Figure 2.

Algorithm for variable selection using quantile regression forests.

The percentage of missing data for baseline predictors was low (<2.5%) except for GCS (6.8%). The Markov Chain Monte Carlo (MCMC) method was used to impute missing values for baseline covariates. All baseline variables were considered in the imputation model. Statistical analyses were performed using R version 4.0.5 with the “quantregForest” and “quantreg” packages.

Results

Description of Patients and Economic Costs

Of the 1480 participants enrolled in this study, 1372 were included in the analysis (Figure 3). The mean age of included participants was 42.9 years (range 16 to 103 years). Out of 1372 participants, 736 (53.6%) were male, 915 (68.3%) employed, 147 (11.0%) students, and 188 (14.0%) retired. Most participants (72.2%) had minor injury severity (ISS ≤ 8). Almost half of RT survivors were drivers (45.8%), 19.3% were pedestrians, 14.8% were passengers, 12.2% were cyclists, and 7.9% were motorcyclists (Table 1). The economic costs for 1372 patients totaled $31,856,845, with a mean of $23,219. The median economic cost was $7,403, with the 10th and 90th quantiles at $1,141 and $49,537, respectively (Table 1). The description of the economic cost distribution based on baseline characteristics of patients is summarized in Table 1. The results indicate higher mean economic costs in the following groups: males ($23,517); elderly patients (age ≥ 64) ($29,833); patients living alone ($26,552); those with somatic symptoms ≥15 ($50,097), pain catastrophizing ≥ 30 ($42,102), and psychological stress ≥ 9 ($37,033); severely injured patients (ISS ≥ 16) ($81,480); and those with head ($25,829), torso ($34,367), spine/back ($26,094), and lower extremity ($27,501) injuries.

Figure 3.

Recruitment flow diagram of the current study.

Table 1.

Distribution of Economic Costs (in Dollars) by Baseline Characteristics of Patients

	n (%)	Minimum	10th Percentile	Median	Mean	90th Percentile	Maximum
All	1372 (100)	197.8	1,141.9	7,403.1	23,219.2	49,537.5	704,160.1
Sociodemographic factors
Age
16-34	582 (42.4)	209.1	583.8	3,842.0	15,191.2	33,035.5	589,067.0
35-64	589 (42.9)	197.8	1,658.9	9,104.4	26,480.9	56,255.2	704,160.1
≥65	201 (14.7)	209.1	2,571.7	11,335.4	29,833.2	62,441.8	491,651.3
Sex
Male	736 (53.6)	197.8	899.7	6,413.0	23,517.7	53,564.8	379,284.5
Female	636 (46.4)	209.1	1,176.7	6,476.5	20,638.3	41,049.8	704,160.1
Employment status
Employed	915 (68.3)	197.8	1,156.3	6,473.4	20,976.6	43,092.3	589,067.0
School	147 (11.0)	209.1	354.3	1,864.9	10,345.9	13,631.8	363,890.0
Retired	188 (14.0)	245.4	2,398.5	11,539.0	28,065.9	60,540.2	704,160.1
Others	90 (6.7)	250.0	1,532.0	10,343.2	35,305.3	92,716.6	491,651.3
Living situation
Alone	334 (24.5)	209.1	1,344.5	9,870.3	26,552.5	63,789.5	284,218.7
With others	1031 (75.5)	197.8	910.7	5,895.7	20,764.4	43,673.5	704,160.1
Education level
Less than high school	89 (6.5)	250.0	1,938.3	15,798.8	39,715.8	116,080.3	280,742.8
High school and vocational	528 (38.6)	197.8	958.8	5,901.5	24,987.9	56,192.4	589,067.0
University	751 (54.9)	209.1	1,053.3	6,422.3	18,168.3	37,271.7	704,160.1
Ethnicity
Caucasian	706 (51.6)	197.8	1,177.6	8,575.4	27,202.5	63,732.4	704,160.1
Asian	328 (24.0)	209.1	965.5	5,060.0	13,192.9	27,786.3	363,890.0
Others	335 (24.5)	209.1	1,016.4	5,416.5	20,535.3	46,535.2	589,067.0
Years lived in Canada
>10 years	1176 (85.9)	197.8	1,177.6	7.532.5	24,061.4	49,394.2	704,160.1
≤10 years	193 (14.1)	209.1	493.9	3,297.9	10,983.9	29,538.9	168,386.5
Psychological and health factors
Somatic symptoms severity (PHQ-15)
<15	1310 (98.0)	197.8	1,017.0	6,223.8	21,558.2	45,092.7	704,160.1
≥15	27 (2.0)	1,139.4	4,376.2	22,850.4	50,097.6	133,934.4	304,059.9
Pain catastrophizing
<30	1280 (95.5)	197.8	989.8	6,361.4	20,892.8	45,035.6	704,160.1
≥30	61 (4.5)	287.4	1,389.2	10,498.3	42,102.6	116,705.2	589,067.0
Psychological distress (PHQ-4)
<9	1335 (97.4)	197.8	1,011.5	6,364.9	21,603.9	45,782.7	704,160.1
≥9	35 (2.6)	1,015.6	1,438.7	10,498.3	37,033.9	105,887.5	312,716.0
HRQoL (SF-12)
<100	272 (20.0)	209.1	1,652.1	11,816.6	32,381.5	84,197.7	589,067.0
≥100	1087 (80.0)	197.8	888.1	5,663.7	19,225.0	40,848.4	704,160.1
Pre-injury comorbidities number
0	635 (46.3)	197.8	694.7	4,091.7	15,729.5	37,427.5	363,890.0
1	409 (29.8)	209.1	1,096.6	7,202.0	22,373.0	49,360.3	589,067.0
≥2	328 (23.9)	209.1	2,352.9	12,326.0	34,439.7	84,605.3	704,160.1
Crash and injury factors
ED visited
VGH (Vancouver)	1120 (81.6)	197.8	1,017.0	6,074.3	21,292.0	44,940.8	704,160.1
RCH (New Westminster)	212 (15.5)	209.1	1,177.6	11,961.5	28,295.8	64,890.0	363,890.0
KGH (Kelowna)	40 (2.9)	320.5	974.1	4,517.9	14,731.4	45,980.8	81,207.0
Crash time (Season)
Spring	241 (17.6)	209.1	1,213.3	7,723.3	23,999.6	55,718.5	704,160.1
Summer	258 (18.8)	209.1	1,046.1	9,465.0	27,340.9	77,143.6	379,284.5
Fall	444 (32.4)	209.1	997.2	6,226.0	21,955.4	45,724.2	589,067.0
Winter	429 (31.3)	197.8	861.9	5,236.4	18,295.9	38,625.9	312,716.0
Road user type
Driver	629 (45.8)	209.1	1,068.4	5,426.6	19,537.0	40,594.7	589,067.0
Passenger	203 (14.8)	245.4	821.0	5,258.8	21,266.7	44,312.7	491,651.3
Motorcyclist	108 (7.9)	197.8	1,263.8	12,675.5	28,545.0	59,696.1	331,706.7
Pedestrian	265 (19.3)	250.0	1,345.8	9,795.3	29,314.1	79,683.5	704,160.1
Cyclist	167 (12.2)	250.0	626.3	6,580.9	17,832.4	37,494.8	379,284.5
ISS
≤3	770 (56.2)	197.8	729.7	3,946.1	8,533.5	20,722.5	244,915.3
4-8	220 (16.0)	209.1	1,250.2	5,421.5	10,758.9	24,043.0	178,568.6
9-15	193 (14.1)	250.0	2,799.4	17,254.1	32,014.2	92,723.7	289,361.4
≥16	188 (13.7)	250.0	8,590.4	40,250.4	81,480.4	227,274.4	704,160.1
GCS
3-8	22 (1.7)	13,352.4	59,137.6	106,896.5	136,563.9	270,211.4	379,284.5
9-13	23 (1.8)	4,934.4	9,846.9	34,270.4	114,106.9	343,420.4	589,067.0
≥14	1233 (96.5)	209.1	1,170.8	6,375.3	19,243.9	42,026.7	704,160.1
Head injury
Yes	521 (38.0)	209.1	1,253.2	8,213.5	25,829.3	49,490.6	704,160.1
No	851 (62.0)	197.8	722.2	5,802.9	19,950.6	44,175.9	491,651.3
Neck injury
Yes	517 (37.7)	209.1	901.8	5,205.1	15,754.6	32,000.6	357,404.0
No	855 (62.3)	197.8	1,158.8	8,044.6	26,070.0	59,717.3	704,160.1
Torso injury
Yes	534 (38.9)	250.0	1,372.9	11,681.8	34,367.1	90,718.5	704,160.1
No	838 (61.1)	197.8	765.5	5,060.0	14,418.8	32,856.9	491,651.3
Spine/back injury
Yes	471 (34.3)	209.1	1,207.3	6,955.2	26,094.8	49,448.8	704,160.1
No	901 (65.7)	197.8	873.6	6,222.9	20,138.0	45,760.1	491,651.3
Upper extremity injury
Yes	699 (50.9)	209.1	1,139.4	8,213.5	23,227.8	45,137.3	704,160.1
No	673 (49.1)	197.8	994.8	5,501.0	21,097.7	48,888.1	491,651.3
Lower extremity injury
Yes	664 (48.4)	197.8	1,072.0	8,332.0	27,501.5	65,014.0	704,160.1
No	708 (51.6)	209.1	977.0	5,608.0	17,194.9	38,190.7	379,284.5

Abbreviations: PHQ-15, Patient Health Questionnaire-15; PHQ-4, Patient Health Questionnaire-4; HRQoL, Health-related quality of life; SF-12, Short Form 12 survey; ED, emergency department; VGH, Vancouver General Hospital; RCH, Royal Columbian Hospital; KGH, Kelowna General Hospital; ISS, injury severity score; GCS, Glasgow Coma Scale.

QRFs Model

The estimated OOB AQL from QRFs model at each iteration in the backward stepwise algorithm at the 10th, 50th, and 90th quantiles of economic costs distribution were shown in Figure 4 (panels A, C, and E, respectively). The optimal QRFs model at the 10th, 50th, and 90th quantiles suggests 6, 7, and 17 major predictors of costs, respectively. ISS, GCS, and age were three most influential variables among low-cost, medium-cost, and high-cost patients. ISS, GCS, age, sex, employment status, and living situation were common major determinants at all quantiles. Ethnicity was selected as an important determinant at the 50th and 90th quantiles. Education level, years lived in Canada, somatic symptoms severity, psychological distress, HRQoL, road user type, and head, torso, spine/back, and lower extremity injuries were selected only for high-cost patients (90th quantile). Figure 4, panels B, D, and F show the importance scores for those selected determinants in optimal QRFs models at the 10th, 50th, and 90th quantiles, respectively. Importance scores quantify each predictor’s contribution to cost prediction. ISS was the most important determinant of cost across all quantiles and GCS and age were the next most influential factors at all quantiles. Among high-cost patients, the fourth and fifth most important factors were employment status and sex, followed by road user type, ethnicity, head injury, torso injury, years lived in Canada, lower extremity injury, living situation, education level, HRQoL, spine/back injury, somatic symptoms, and psychological distress.

Figure 4.

The variable selection and rank of the selected major determinants for the 10th, 50th, and 90th quantiles of economic costs. Panels (A) and (B) show, respectively, the estimated out-of-bag average quantile loss for each QRFs model at each iteration and the importance scores of the selected major determinants for the 10th quantile of costs. At each iteration, the QRFs model includes the remaining k variables (numbered 1, 2, …, k) after sequentially removing variables (numbered k+1, …, 23). Importance scores indicate each predictor’s contribution to cost prediction. Panels (C) and (D) present the corresponding results for the 50th quantile of costs, while panels (E) and (F) display the results for the 90th quantile.

Abbreviations: HRQoL, Health-related quality of life; ED, emergency department; ISS, injury severity score; GCS, Glasgow Coma Scale.

Quantile Regression Model

Further analysis using classical quantile regression to assess the effect of selected determinants on costs across the 10th, 50th, and 90th quantiles of cost distribution is provided in Table 2. According to the findings, being female, age ≥ 35, ISS ≥ 9, and GCS ≤ 8 were significantly associated with higher economic costs among low-cost patients. Being female, age ≥ 65, being retired, Caucasian ethnicity, ISS ≥ 4, and GCS ≤ 8 were significantly associated with higher economic costs among medium-cost patients. Age ≥ 65, being retired, having less than high school education, higher somatic symptoms (PHQ-15 ≥ 15), lower HRQoL (SF-12 < 100), ISS ≥ 9 and GCS ≤ 13 were associated with higher costs among high-cost patients.

Table 2.

Results of Multivariable Quantile Regression on Economic Costs Following Road Trauma

	10th Quantile	50th Quantile	90th Quantile
	β (95% CI)	β (95% CI)	β (95% CI)
Sex (Ref: Male)
Female	410.05 (71.11, 748.97)*	1138.34 (148.23, 2128.45)*	3785.53 (−557.14, 8128.21)
Age (Ref: 16–34)
35-64	707.48 (329.07, 1085.89)***	2359.59 (1255.91, 3463.26)***	7844.87 (3077.69, 12,612.04)**
≥65	2045.53 (1293.11, 2797.94)***	603.53 (−1591.35, 2798.43)	6141.01 (−3268.52, 15,550.55)
Employment status (Ref: Employed)
School	−458.59 (−1028.35, 111.15)	−1503.11 (−3172.73, 166.51)	−4001.05 (−11,188.31, 3186.20)
Retired	307.53 (−408.25, 1023.32)	2957.79 (870.57, 5045.01)**	9554.51 (717.02, 18,392.01)*
Others	86.33 (−581.44, 754.12)	480.10 (−1467.20, 2427.40)	8102.09 (−343.59, 16,547.78)
Living situation (Ref: With others)
Alone	207.73 (−182.73, 598.19)	545.69 (−602.22, 1693.61)	−141.26 (−5050.99, 4768.45)
Education level (Ref: University)
Less than high school	†	†	9523.94 (745.17, 18,302.72)*
High school and vocational	†	†	−324.18 (−4753.89, 4105.55)
Ethnicity (Ref: Caucasian)
Asian	†	−1683.28 (−2895.07, −471.48)**	−4393.92 (−9709.70, 921.85)
Others	†	−1345.40 (−2545.89, −144.91)*	−684.71 (−5865.89, 4496.43)
Years lived in Canada (Ref: >10 years)
≤10 years)	†	†	4382.63 (−2051.75, 10,817.02)
Somatic symptoms severity (Ref: <15)
≥15	†	†	32,036.70 (16,678, 47,395.21)***
Psychological distress (Ref: <9)
≥9	†	†	3295.38 (−10,482, 17,072.91)
HRQoL (Ref: ≥100)
<100	†	†	18,472.94 (13,035.26, 23,910.63)***
Road user type (Ref: Driver)
Passenger	†	†	1364.38 (−4755.75, 7484,51)
Motorcyclist	†	†	−534.11 (−8965.74, 7897.52)
Pedestrian	†	†	−38.12 (−6127.74, 6051.52)
Cyclist	†	†	−1179.17 (−8182.32, 5823.98)
ISS (Ref: ≤3)
4-8	224.05 (−242.81, 690.91)	1559.22 (198.45, 2919.89)*	2298.22 (−3531.01, 8128.17)
9-15	3019.88 (2516.22, 3523.53)***	14,249.46 (12,777.88, 15,721.03)***	54,425.46 (48,008.29, 60,842.55)***
≥16	8127.41 (7595.37, 8659.45)***	31,586.11 (30,033.04, 33,139.18)***	169,313.29 (162,172.58, 176,454.01)***
GCS (Ref: ≥14)
3-8	58,347.60 (56,985.15, 59,709.04)***	74,709.98 (70,800.42, 78,619.54)***	62,649.63 (45,846.30, 79,453.32)***
9-13	−273.77 (−1416.74, 868.79)	902.17 (−2379.09, 4184.25)	49,357.64 (35,245.25, 63,470.02)***
Head injury (Ref: No)
Yes	†	†	505.44 (−3780.66, 4791.55)
Torso injury (Ref: No)
Yes	†	†	60.71 (−4396.88, 4518.31)
Spine/back injury (Ref: No)
Yes	†	†	2393.88 (−1979.88, 6767.65)
Lower extremity injury (Ref: No)
Yes	†	†	1874.68 (−2505.16, 6254.52)
Pseudo R-squared	0.15	0.26	0.47
Rank Score Test	<0.001	<0.001	<0.001

Notes: † Not selected. *P < 0.05, **P < 0.01, ***P < 0.001.

Abbreviations: Ref, reference; HRQoL, Health-related quality of life; ISS, injury severity score; GCS, Glasgow Coma Scale.

Additionally, our results show that the selected major predictors affect low-cost, middle-cost and high-cost participants differently. For example, the estimated effect for age 35–64 on the distribution of costs was more pronounced among high-cost patients compared with low-cost patients (β = $707.48; 95% CI = $329.06-$1,085.89 for the 10th quantile; β = $7,844.87; 95% CI = $3077.69-$12,612.04 for the 90th quantile). Patients with severe injury (ISS ≥ 16) incur $169,313.29 (95% CI = $162,172.58-$176,454.01) more than those with minor injury (ISS ≤ 3) at 90th quantile and $8,127 (95% CI = $7,595.37-$8,659.45) more at the 10th quantile. Patients with GCS ≤ 8 incur $58,347 more in costs compared to those with GCS ≥ 14 among high-cost patients and $62,649 more among low-cost patients.

Discussion

The aim of the current study was to employ a robust and reproducible machine learning approach to identify major determinants of economic cost among low-cost, medium-cost, and high-cost RT survivors, rank their importance, and then quantify their effects on cost distribution using quantile regression. We first estimated economic costs including healthcare and productivity costs over 12 months following minor to severe RT injury in a multicenter, multilingual, prospective cohort study of 1372 RT survivors in BC, Canada. In our cohort, the mean costs over a year following RT were estimated at CAD $23,219.20. The high costs observed in our study align with findings from existing literature in other countries. For instance, a previous report demonstrated that hospital costs for RT patients were approximately US $20,800 (CAD $26,832) on average in Saudi Arabia in 2017.¹⁰ An Australian study indicated that RT patients generated mean healthcare costs of AUD $10,153 (CAD $10,369) over a 12-month period in 2011.¹⁵ Another study conducted in the Netherlands found that the mean costs including healthcare and productivity costs of RT injuries amounted to €11,400 (CAD $16,644) over 24 months after injury in 2017.¹³

We collected a range of sociodemographic, psychological, health, crash, and injury factors of RT patients as potential determinants of costs. Our results illustrated that the effects of these factors across population are not uniform and quantile regression provided a comprehensive picture about effects of factors on costs in whole population. Identifying major factors of costs among different subpopulations at different levels of costs and ranking their relative importance using a rigorous machine learning-based approach provides information that policymakers can use to develop tailored strategies to address the needs of high-cost patients and save more potential costs. We identified 6, 7, and 17 significant cost-drivers of RT injuries among low-cost, medium-cost, and high-cost RT survivors and estimated the costs they incurred. Therefore, our findings can provide higher-resolution insight into the upper limit of costs that can be saved by investing in intervention programs among RT survivors. Since most of these costs were paid by public health insurance in BC, improving traffic safety is an opportunity to reduce a significant portion of government expenses. Among variables, ISS, GCS, and age were three predominantly influential variables among low-cost, medium-cost, and high-cost patients. The impact of ISS and GCS is explained by the fact that patients with more severe injuries (higher ISS or lower GCS) typically require extended stays in hospitals that offer elevated levels of care, which is in line with previous research.^{10,11,13,14,16,50} In addition, age-related factors could exacerbate RT injury and lead to prolonged recovery times, extended hospital stays, and consequently, increased costs.⁵¹ ISS, GCS, age, sex, employment status, and living situation were common major determinants at all quantiles. The results of previous studies from the Netherlands from 2017 showed higher healthcare costs for females across the entire injury population.^13,52 Ethnicity was picked only at the 50th and 90th quantiles. Education level, years lived in Canada, road user type, and head, torso, spine/back, and lower extremity injuries were selected only for high-cost patients. Notably, somatic symptom severity, psychological distress, and lower HRQoL emerged as key factors within this group. This finding aligns with the adverse effects of poor pre-injury health on the recovery process and return-to-work post-injury.^53,54

Major determinants selected at different quantiles had a bigger impact on the 90th quantile (upper tail) than on the 10th quantile (lower tail). For example, increased ISS may impose exorbitant costs among high-cost subpopulations, while it is associated with negligible costs among low-cost subpopulation. Among our patients with ISS 9–15, costs increased by $3019.88, $14,249.46, and $54425.46 at the 10the, 50th, and 90th quantiles compared to patients with ISS ≤ 3, respectively (Table 2). Alghnam et al also found that the effect of older age on the healthcare costs distribution was greater at the upper quantiles.¹⁰ This finding highlights the heterogeneity of cost within the population and the fact that the effect of different predictors on cost varies for high versus low cost cases. This population heterogeneity is ignored in mean-based models, potentially leading to biased estimates.

Identifying key determinants of higher costs among the high-cost subpopulation enables policymakers to implement targeted cost-containment strategies and optimize resource allocation. For example, older age, being retired, having less than a high school education, higher somatic symptoms, lower HRQoL, higher ISS, and lower GCS scores were all associated with higher healthcare costs in this subgroup. While some factors, such as age and injury severity, are non-modifiable, others—particularly HRQoL, somatic symptoms, and psychological distress—can be addressed through early interventions. Integrating comprehensive rehabilitation services, mental health support, and patient education initiatives into post-injury care could help mitigate long-term healthcare costs. Furthermore, this evidence reinforces the impact of pre-injury health status on post-injury recovery and its financial implications. By implementing preventive healthcare strategies—such as promoting workplace wellness programs, chronic disease management, and community-based health screenings—policymakers can enhance HRQoL before an injury occurs, potentially reducing economic costs. Overall, a better understanding of the modifiable characteristics of high-cost patients provides an opportunity to shift healthcare policies from a reactive cost-management approach to a more sustainable, preventive model, ultimately leading to improved patient outcomes and reduced economic burden on the healthcare system.

This study has notable strengths, including its prospective design spanning 12 months with five rounds of data collection on lost productivity. We used validated scales to assess HRQoL, somatic symptom severity, psychological distress, and pain catastrophizing. The use of administrative data to estimate healthcare costs is another strength. Moreover, to improve generalizability, no limitations were placed on road user type, injury severity level, or language. It is important to recognize some limitations as well. First, self-reported pre-injury HRQoL, somatic symptom severity, and psychological distress may be subject to recall bias, although most baseline interviews were conducted within seven days after injury to minimize this bias. Second, the current estimate of costs likely underestimates the actual economic burden of RT injuries, as private-pay healthcare costs, legal costs, property damage costs, life years lost costs, and compensation costs relating to pain, stress, and suffering have not been included. In addition, our study did not assess the long-term healthcare costs beyond a year following RT injury; injured persons often have increased usage of long-term healthcare services in comparison with the general population.⁵⁵ Older individuals and those experiencing significant pain, who likely utilize more healthcare resources, may have been less inclined to participate in the study. Lastly, our estimates exclude costs from events involving lone cyclist, e-bikes, and pedestrian-only injuries which is a limitation given the rising incidence of serious injuries in these groups due to the increasing use of active transportation in urban environments. Third, our inclusion criterion requiring presentation within 24 hours of the collision may introduce selection bias by excluding individuals who delayed seeking care. This may disproportionately affect patients from rural or underserved areas, or those with delayed symptom onset, and may limit the generalizability of our findings to the broader population of RT survivors. Moreover, we may have missed survivors with minor injuries who never sought hospital treatment or who were discharged rapidly from the ED. Fourth, the absence of important socioeconomic variables such as household income, urban versus rural residence, and health insurance coverage limits our ability to fully capture cost drivers. Future studies should incorporate these variables to provide a more comprehensive understanding of healthcare cost determinants in this population. Finally, we estimated lost productivity costs using average national hourly wages rather than participants’ actual wages, which may introduce some measurement error.

Conclusions

A data-driven machine learning approach provided valuable insights into the key cost drivers of RT injuries across different quantiles of the cost distribution. Our findings indicate that high-cost patients were more likely to be older, retired, have lower education levels, higher somatic symptom severity, lower HRQoL, higher ISS, and lower GCS. These insights may help policymakers prioritize resource allocation and implement targeted interventions for high-risk groups—particularly older adults and those with more severe physical and psychological conditions—to reduce healthcare costs.

Funding Statement

This study was funded by a research grant from the Canadian Institutes of Health Research (grant # 461108).

Data Sharing Statement

Access to data provided by the Data Stewards is subject to approval but can be requested for research projects through the British Columbia Ministry of Health (the designated Data Steward) or their authorized service providers (Population Data BC). The following data sets were used in this study: Fee for Service (MSP); Discharge Abstract Database (DAD); National Ambulatory Care Reporting System (NACRS); Pharmanet. You can find further information regarding these data sets by visiting the PopData project webpage at: https://my.popdata.bc.ca/project_listings/21-100/collection_approval_dates. All inferences, opinions, and conclusions drawn in this publication are those of the author(s), and do not reflect the opinions or policies of the Data Steward(s).

Ethics Approval and Consent to Participate

This study was approved by the research ethics board of the University of British Columbia (H18-00284). Participants provided informed written or verbal consent. For minors (16–18 years old), parental/guardian permission was obtained in addition to participant assent. For participants unable to provide consent (eg, comatose), proxy consent was obtained from a designated caregiver. All methods were carried out in accordance with relevant guidelines and regulations.

Disclosure

The authors report no conflicts of interest for this work.