We used the DCE to understand Americans’ preferences for hypothetical teen driving policies. DCEs describe policy alternatives according to their key characteristics, or attributes (e.g., testing requirement for driver’s license), and levels of these attributes (e.g., no test, easy on-road test, or hard on-road test). Different levels of each attribute are combined to create choice alternatives. These alternatives are then combined to construct choice profiles (e.g., with two or more alternatives per profile) that are presented sequentially to respondents as choice scenarios. In each choice scenario, respondents are asked to state which of the presented alternatives they prefer. This gives an indication for participants’ underlying preferences and enables analysts to estimate their relative preferences for and tradeoffs they are willing to make between attribute levels.

DCE methodology is rooted in the economic theory of utility maximization; it belongs to a family of preference-based survey methods rooted in multi-attribute valuation, specifically, choice modeling. [7, 8] DCEs are increasingly used in health policy applications to elicit preferences for and tradeoffs between attribute levels of various healthcare resources. [9]

Our DCE was part of a national road safety survey conducted by the Johns Hopkins Center for Injury Research and Policy and fielded by the University of Chicago’s National Opinion Research Center (NORC). We developed a dual-mode DCE survey (web- and phone-based, described below), programmed and fielded by NORC from May 4 to June 10, 2022, in both the English and Spanish languages, in 808 participants ≥16 years. This sample size was deemed sufficiently large to detect a change of 0.01 in preference weights at the 95% significance level, with an 80% statistical power. Participants were recruited from NORC’s AmeriSpeak Panel, a probability-based panel of randomly sampled US households who participate in surveys, online via the AmeriSpeak App, or by phone. *10+ NORC’s sampling frame, representative of over 99% of US households, has been used in well-known studies, like the National Longitudinal Surveys of Youth. [10] The DCE was conducted as a special module of the larger road safety national survey. Participants in the DCE module were recruited from NORC’s AmeriSpeak Panel, to be nationally representative of the U.S. ≥16-year population, after applying survey weights. Participants comprised adults (≥18), selected using 48 demographic strata for representativeness, and teens (16–17) randomly chosen from each household without overlapping with adults. Survey weights were developed for this sample, adjusting for selection probabilities, nonresponse, and population benchmarks, to produce nationally representative estimates from the DCE sample.

We used multinomial logit (MNL) and mixed logit (MIXL) models to estimate preferences for different levels of attributes of teen driver licensure policy. (1) PrYn,t=i∣Xn,t,Zn,j,t=expUYn,i,t∑j=1JexpUYn,j,t where: (2) UYn,i,t=βn,Opt-in+βn,HardTestHardTestn,i,t+βn,EasyTestEasyTestn,i,t+βn,LicensurePeriodLicensurePeriodn,i,t+βn,MonitoringPeriodMonitoringPeriodn,i,t+γnzn,i,t+ϵn,i,t, In equation 2, ∈_n,i,t is assumed to be independently and identically distributed (iid) extreme value type I for the MNL model and can have any distribution for the MIXL model. Here, we assumed ∈_n,i,t to be multivariate normal. The coefficients estimate the effects of policy attributes and individual characteristics on the choice probability. The MIXL model also accounts for repeated choices by the same respondent, allowing estimation of respondent-level random coefficients to capture preference heterogeneity. [19]

Marginal rates of substitution and relative importance of attributes We calculated the marginal rates of substitution (MRS) between attribute levels, indicating how much of one attribute respondents are willing to trade for another. It is the ratio of the mean preference weights of the compared attribute levels: (3) EWTPkr=−Eβk/Eβr where β_r denotes the estimated mean preference weight for the reference attribute level.

Confidence intervals (95%) were estimated using the method of Krinsky and Robb (1986, 1990).[20–23]

We also calculated the average relative importance of attributes given by respondents as the ratio of each attribute’s part-worth utility range to the sum of all ranges, multiplied by 100: (4) ARIi=100×Ui¯−Ui¯∑j=1JUj¯−Uj¯ where Ui¯=maxβi,kZi,k,∀k∈Ki and Ui¯=minβi,kZi,k,∀k∈Ki denote, respectively, the largest and smallest utility level of attribute i where Z_i,k is the k^th attribute leveI of the i-th attribute (k ∈ K_i, with K_i denoting the number of attribute levels of attribute i), β_i,k is the coefficient of attribute level Z_i,k, and minβi,kZi,k,∀k∈Ki=0 when an attribute is continuous rather than discrete, and β_i,k = 1 when Z_i,k is the reference level for the attribute. [15, 24–26] To characterize uncertainty in the attribute importance metric, we conducted bootstrap sampling with N=50 sampled values and R=500 replications. Values for the coefficients were drawn from a multivariate normal distribution, using the mean and variance-covariance matrices from the MNL and MIXL models.

Unless specified otherwise, all estimates are based on a survey-weighted sample. All statistical analyses were conducted in the R programing language. [27]

This study did not involve patients. As noted above, the survey instrument was pilot tested with 22 respondents to ensure its relevance and clarity.

Of the 808 respondents who participated in the survey, 730 participants completed all DCE choice tasks (figure 1 and table 2). Unweighted sample mean age was 31.2 years (standard deviation [SD], 15.0 years), including 30.7% teenagers (16–17 years), and consisted predominantly of females (62.4%), non-Hispanic Whites (56.8%) with less than a high school degree (50.0%). Unsurprisingly, the majority of participants were not married (67.7%) and were not in the labor force (45.2%). The majority of respondents (51.9%) resided in a household earning less than $60,000. The median household size was 5 persons (interquartile range [IQR], 3.00,6.00). The median number of household members younger than 18 years was 2 (IQR, 1.00,3.00). Respondents resided predominantly in metropolitan areas (81.9%) and in the South (35.4%), Midwest (26.5%) and West (21.5%). Survey-weighted estimates are also provided in table 2 and are largely qualitatively consistent with the unweighted estimates.

Survey-weighted average preference weights (unexponentiated) for the attribute levels are reported in table 3 and depicted in supplemental figure 1.

They indicate correct parameter signs in both MNL and MIXL models. Key results are presented and interpreted for the MIXL models (Model 3 and Model 4), based on goodness of fit statistics (McFadden pseudo-R-squared, Akaike and Bayesian information criteria). [14, 28] They indicate a significantly high baseline support for changes to teen driving licensure policies (Model 3: 1.98; 95% CI, 1.64–2.33; Model 4: 5.61, 95% CI, 3.27–7.95). Baseline support for the policy change varied significantly by certain socioeconomic and sociodemographic characteristics. For example, likelihood of support was negatively associated with being women, identifying as other or multiple race and ethnicity, having completed a bachelor’s or equivalent degree, living in a household with income $60,000 or more, residing in the Midwest or South, or in a metropolitan area. In contrast, identifying as Hispanic or non-Hispanic Asian, or not being in the labor force were associated with greater likelihoods of support for a policy change.

Preferences were stronger for testing over no testing, (easy test: 0.88; 95% CI 0.68,1.07; hard test: 0.31; 95% CI 0.09,0.52), suggesting that the public perceived testing as a “good”. There was distaste for an extended driver monitoring period (−0.26; 95% CI: −0.47,−0.04 for each additional year) or longer intermediate licensure periods (−0.89; 95% CI: −1.15,−0.63 for each year), suggesting that extended monitoring and longer intermediate licensure periods are perceived by the public as “bads”.

Baseline preference weights for policy change were substantially heterogeneous based on the magnitude and statistical insignificance of the standard deviations and varied by certain individual characteristics like sex, race/ethnicity, education, marital status, income, and geography, but not age, noticeably.

Table 4 shows the marginal rates of substitution between attribute levels.

These are also illustrated in online supplemental figure 2 as a heat map, with darker colors indicating larger magnitudes: navy for substitutions between two “goods” or two “bads” and red for substitutions between a “good” and a “bad”. They consistently suggest that the substation rates are larger between 1) the length of the extended driver monitoring and intermediate licensure periods, 2) the length of the monitoring period and easy tests, 3) hard test and length of the intermediate licensure period, and 4) hard test and easy test. Based on Model 4, the public was willing to give up at least 3.46 years (95% CI: −5.81,−1.11) in intermediate licensure period for a year reduction in monitoring period. They were willing to give up support for an easy test requirement if this reduced the duration of the monitoring period by at least 3.41 years (95% CI: 0.25,6.57). Similarly, the public was willing to give up support for a hard test requirement for a reduction of at least 2.90 years (0.28,5.51) in licensure period; support for an easy test was 2.85 times (95% CI: −4.57,−1.13) that for a hard test.

The length of the intermediate licensure period, followed by testing requirements, were the most important attributes in Americans’ decisions about a teen driving licensure systems, among the set of attributes included in this study (table 5 and online supplemental figure 3).

The length of the intermediate licensure period was nearly twice as important as testing requirement (60.5% vs 30.8%), which itself was over three-folds as important as the length of the extended driver monitoring period (30.8% vs 8.6%).

These results highlight public preferences for teen driving licensure reforms. Respondents favored easy tests and shorter monitoring and intermediate licensure periods, emphasizing the public’s commitment to safe driving. Research indicates that teen crashes stem from inexperience and risk-taking rather than ignorance of road rules. [40] An effective driving test aligning with scientific evidence could enhance teen driver safety. Aversion to extended licensing and monitoring periods may reflect concerns about limiting teen mobility or privacy. While GDL provides safety benefits, extended monitoring and intermediate driving licensure periods may burden license-seekers. Striking a balance between public preferences and broader policy goals is crucial. Policymakers should consider these preferences, to optimize mobility and maximize safety. Our findings can therefore guide effective communication, resource allocation, and continuous policy adjustments based on evolving preferences and circumstances.