Motor vehicle collisions are a major source of morbidity and mortality around the world, causing about 20–50 million injuries and 1.2 million deaths every year.¹ Motor vehicle crashes are the leading cause of death among people aged 15–29 years worldwide.¹² Young novice drivers have the highest crash rate; per kilometre driven, the crash rate for 16-year-old drivers is approximately four times greater than that for drivers ages 30–59 years in the USA.³ This excess crash risk is mainly due to inexperience and risky driving behaviours.^4–6 To address this issue, some European countries including Sweden, Norway, France and Belgium have implemented an extended learner permit phase, requiring supervised driving under all conditions for adolescent drivers before they reach 18 years.^7–9 Australia, Canada, New Zealand, the USA and Israel have implemented Graduated Driver Licensing (GDL) laws in some or all states/provinces.^10–15 GDLs in Australia and Canada apply to drivers of all ages. In the USA, New Jersey’s GDL system applies to people entering the licensing process who are under 21, and some other states apply some restrictions to people 18 and older. In the USA, GDL regulates licensing and driving behaviours among adolescents younger than 18 years in three phases: the extended learner phase, requiring supervised driving under any conditions for 3–12 months; the intermediate phase, allowing unsupervised driving under low-risk conditions such as daylight or when carrying less than one young passenger; and the full licensure phase, permitting unsupervised driving all the time.

GDL policies could reduce crash rates if they reduced risky driving behaviours among those covered by GDL restrictions. Or they might reduce crash rates by reducing the amount of driving by adolescents. Either of these mechanisms would reduce crash rates for drivers age 16 and 17 years, while rates for those 18 years would be unchanged. One effect of GDL programmes is to delay licensure by increasing the minimum permit and/or intermediate license age, or by placing additional requirements on teenagers such as minimum practice driving hours. Another possibility is that GDL laws may reduce crash rates among those age 16 and 17 years, but crash rates might increase among drivers age 18 years because they drive fewer miles and therefore learn fewer driving skills while driving under GDL laws at ages 16 and 17. In states that do not apply GDL to people 18 and older, teenagers might delay licensure until 18 and thereby have higher crash rates at age 18 because they are new drivers. Another possibility is that crash rates are reduced among drivers age 16 and 17 years, and also among drivers age 18 years because teenagers licensed under GDL would have a greater amount of practice and gradual introduction to riskier driving situations and thereby lower crash rates at age 18.

GDL implementation has been reported to reduce involvement in a vehicle crash as a driver by approximately 15–40% for adolescents aged 16 years.^16–20 One reason that studies have found different effects of GDL is because some GDLs have more restrictive provisions than others. In addition, the definition used for a GDL law has varied among studies.^16171920 Few studies have examined 17-year-old and 18-year-old drivers. The potential effect of GDL on teenagers age 18 years is primarily a US issue, as other countries have older licensing ages and most apply GDL to drivers of all ages. A study of US fatal crashes reported similar crash rates among teenagers age 18 years before and after GDL implementation,¹⁷ but another study of fatal US crashes found that GDL was associated with a 10% increase in fatal crash involvements among those age 18.¹⁶ We conducted a systematic review and meta-analysis to estimate age-specific associations between GDL implementation and crash rates.

We searched MEDLINE, Transportation Research Information Service, Web of Science, Google Scholar, and internet sites maintained by the Insurance Institute for Highway Safety, the National Highway Traffic Safety Administration, the Centres for Disease Control and Prevention, and the American Automobile Association Foundation for Traffic Safety for studies of GDL polices from January 1991 through December 2011. Our search terms included: (1) GDL; (2) (graduate* or gradual* or delay* or driver or provisional) in combination with (permit* or licen* or restrict* or delay*); (3) (teen* or you* or adolescen*) in combination with (driv*). We used the ‘related citations’ feature to capture additional references for selected articles. We examined the references of articles and reviews.¹⁸²¹²²

Full-text versions of articles or reports were reviewed by the first author (MZ). Data extraction was conducted by the first author (MZ) and verified by the second author (PC). For this study we defined GDL as a new law with a learner phase of at least 3-months plus an intermediate phase that restricts driving at night and/or restricts the number of passengers allowed.²³ To be included in the meta-analysis, a study had to use counts of crashes, injuries or deaths as the outcome. It had to provide age-specific incidence rate ratio estimates with CI or information that allowed us to calculate these quantities. We included studies from individual states or provinces. We excluded national studies because within a given nation they have overlapping time periods and therefore their results are not independent. In addition, national studies overlap the time intervals of most studies from smaller jurisdictions such as states. We examined rates based on person-years, and did not consider rates based on license-years because license data for adolescents is not often or consistently reported by states/ provinces. Some jurisdictions include adolescents in the intermediate phase, while others count only those fully licensed.

Not all studies used the same outcomes. When extracting rate or rate ratio estimates, we first used an estimate based on the count of crashes in which a teenager was involved as a driver and at least one person was injured. Our second choice was a count of injured teenage drivers. Third was a count of crashes with a teenage driver and at least one death. Fourth was a count of fatally injured teenage drivers. We extracted the year of GDL implementation and information about each law. We identified whether the original manuscript adjusted the rate ratio for temporal trends using rate data from drivers age 21 and older, who should not be affected by GDL laws but should be affected by other factors that influence crash rates, such as changes in speed laws, seat belt use or vehicle design.

Our goal was to pool age-specific rate ratios from each study to summarise the association between GDL presence and crash rates. One study (North Carolina)¹⁵ provided an age-specific rate ratio that was not adjusted for other changes in rates over time; an adjusted estimate was not available. The needed rate ratios with CIs were not in the remaining studies. We therefore extracted from each study age-specific counts of outcome events (crashes with an injury or counts of injured drivers) and population estimates before and after GDL passage. These data came from 1–5.5 years of time before GDL passage and 1–6.5 years after passage. We used age categories of 16, 17 and 18 years as well as a category for older drivers if that information was available. We estimated adjusted rate ratios (aRR) for the association of GDL laws with crash rates using Poisson regression, with age-specific person-time as offsets. We adjusted for age group and included interaction terms between GDL presence and adolescent age groups. Except for NC, we adjusted for time as a linear term in all models to account for changes in crash rates over time due to factors other than GDL laws; but the method used depended upon the available data. For five studies (California,²⁴ Florida,²⁵ Georgia,²⁶ Nova Scotia,¹¹ New York²⁷) the adjustment for time was based on changes in the crash rates of older drivers, because we had only data from one time period before the GDL law and one after. For four studies (Maryland,²⁸ Michigan,²⁹ Pennsylvania,³⁰ Texas³¹) the adjustment for time was based on data for teenage drivers, because data was available for three or more time intervals. For two studies (Ohio,³² Wisconsin³³) the adjustment used data from both teenage and older drivers, because data was available for three or more time intervals and data was available for older drivers.

To clarify we will describe what we did for one study from Nova Scotia.¹¹ For our meta-analysis we needed to extract an adjusted rate ratio with a CI for the association between the GDL law and the crash rate. But this information was not in the paper. Table 3 of that paper provided data for counts of Nova Scotia drivers in crashes with an injury for the years 1993 and 1995 (table 1). Nova Scotia’s new GDL began in October of 1994. We created variables for age group and the interaction between age 17 and the GDL law. Then we used Poisson regression to estimate the aRR. This was adjusted for the linear trend in crash rates for persons age 25 years and older from 1993 to 1995.

We used inverse-variance methods to produce pooled estimates of the aRRs from each study using Stata V.12. Random-effect estimates were calculated using the method of DerSimonian and Laird.³⁴ Fixed effect estimates were also obtained. The Cochran Q-statistic was used to test the hypothesis that rate ratios were homogeneous across studies.³⁵ We calculated I², an estimate of the percent of total variation between studies due to heterogeneity rather than chance.³⁶ To test for publication bias, funnel plots were inspected and Egger’s test for asymmetry was used.³⁷ To identify characteristics associated with heterogeneity, we used subgroup analyses for the following variables (1) outcome type (crash with injury, crash with death), (2) entry age for learner permit phase (<16 years, 16), (3) entry age for intermediate phase (16 years, >16), (4) night-time driving restriction (starting at 21 : 00 or 22 : 00; starting at 23 : 00, midnight or 1 : 00; none or no change) and (5) number of young passengers allowed (0, 1, 2, 3 or more). Because tests of heterogeneity are statistically weak when there are few studies, we used p<0.1 as our criteria to judge that there was heterogeneity.

When a GDL is passed, its effects may not be immediate for several reasons. One is that some people might already have a license and may be grandfathered in under the law. For example, a 16-year-old may be allowed to continue to drive under the old law provisions after the new law is passed. For this reason, some studies use data until the time of GDL implementation. Then after the GDL goes into effect, they omit 1 year of data, using information for 16 year olds only after the law has been effect for over a year. We attempted to deal with this issue in our analyses. For example, for age 16 years there were 12 studies. Of these, six omitted the first year after the law went into effect or presented data in such a way that we could omit that first year. For a further two studies, the first year after the law went into effect could not be omitted, but there were 5 years of data available after the law went into effect, and thus 16-year-olds were fully covered by the law in four of the 5 years of post-law data. This means that any dilution of the GDL effect induced by including the first year after implementation should be relatively small in 5 years of data. But in four jurisdictions, Florida, Nova Scotia, Ohio, and Pennsylvania, we could not exclude a full year of data after the law went into effect and we did not have 5 years of data after the law went into effect. For Florida, a 6 month period after the law went into effect was excluded. For Nova Scotia, 2 months were excluded. For Ohio, 6 months were excluded. No post-law time was excluded for Pennsylvania. Thus, in theory, some dilution of the GDL effect might be found in these four studies. To find out if there was evidence for this dilution of effect, we used a test of heterogeneity to compare the pooled aRR in the eight studies where one post-law year was excluded, or 5 years of post-law data was used, with the pooled aRR in the four studies where this could not be done.

For drivers age 17 years, we followed the same procedure, except we tried to exclude 2 years of data after GDL implementation. There were eight studies for this age group. For three, we could either exclude 2 years of data after GDL passage or use 5 years of data after passage. We could not do this for the other five studies. Again, we compared the aRR estimates from the first three studies with the other five.

For drivers age 18 years there were only four studies. California excluded the third year and used the fourth year to the eighth year after GDL implementation. Florida excluded only 6 months. Georgia excluded no time, but used 5.5 years of data after implementation. Wisconsin excluded 2.25 years of post-law data. We did not perform the test of heterogeneity in this small group of four studies.

The literature search identified 1397 papers, but 1253 were excluded as not relevant based upon their titles (figure 1). Another 97 papers were removed after reading the abstract and the remaining 47 papers were read. We identified 12 studies that could provide crash counts and population estimates for age 16,^111524–33 eight for age 17,^{1124–27303233} and four for age 18.^24–2633 We conducted new analyses to estimate adjusted rate ratios with CIs for GDL presence for 11 of the 12 studies for adolescents age 16 years. For North Carolina¹⁵ the data were not available for a new analysis. We conducted new analyses for all eight studies for age 17, and all four studies for age 18.

Our meta-analysis estimated that adolescents aged 16 years experienced a 22% (95% CI 16% to 28%) reduction in crash rates, while among 17-year-olds the rate reduction was smaller, 6% ( 95% CI 4% to 7%), and among teenagers aged 18 years the rate changed little after GDL implementation. The apparent difference in GDL effects among those age 16 and those age 17 years is probably not due to the different study jurisdictions, as the pooled random-effects aRR for 16-year-olds was 0.78 using the eight studies with results for both ages 16 and 17, the same as the pooled aRR of 0.78 using all 12 studies.

Perhaps the most important limitation of this meta-analysis was the small number of studies available with age-specific data for ages 17 and 18 years; there were reports available from 12 jurisdictions for age 16 years, but only from eight jurisdictions for age 17 and four for age 18. Because we had few studies to summarise, all results are exploratory and those for adolescents age 18 years are particularly unreliable due to the small number of studies. Second, it would be ideal to study the same outcomes for every jurisdiction as it is possible that associations with GDLs might vary with the type of crash outcome. Among the 12 studies for age 16 years, the rate ratio for GDL laws varied somewhat according to whether the outcome involved injury (aRR 0.79) or death (aRR 0.72), but the difference in these rate ratios was modest and not statistically different (p=0.44, table 4). Another limitation is that we could not exclude 1 to 3 years of data after GDL implementation for all studies, allowing us to examine whether GDL effects might be stronger when all drivers of a given age are covered by the laws. To address this problem, we compared aRRs based on excluding 1 or 2 years of data after GDL passage or based on at least 5 years of data after GDL passage, with estimates from jurisdictions where the years immediately after GDL passage had to be included. These estimates were similar. A national evaluation of 1996–2007 US fatal crashes compared the immediate effect of GDL with 1-year delay for age 17, and 2-year delay for age 18, and reported similar estimates.¹⁷ Another limitation is that our estimates may be subject to residual confounding. GDL effectiveness should ideally be estimated with many repeated measures of crash rates before and after GDL implementation. We conducted new analyses for 11 of the 12 studies for age 16 by using methods to control for temporal changes in crash rates unrelated to GDL policies. However, some traffic safety factors may affect adolescents and adults differently, and our ability to control for temporal trends was limited by the available data years in the original research. Nevertheless, our estimates were comparable to nationwide US evaluations using more than 10 years of crash rates.^16171920 Another limitation is that none of the studies selected for meta-analysis contained any information about the amount of driving done by adolescents.

Our estimate of a 22% (95% CI 16% to 28%) reduction of traffic crash rates among 16 year olds is consistent with previous studies for this age group for the entire USA. A review of collision claims for vehicles less than 4-years-old during 1996 through 2006 reported a 22% reduction (95% CI 18% to 27%) for states with good GDL ratings from the Insurance Institute for Highway Safety and a 17% reduction (95% CI 12% to 31%) for states with fair GDL ratings.²⁰ In an analysis of crashes with a death during 1994–2004, GDL polices that included five or more of seven GDL components were associated with an 18–21% reduction in crash rates.¹⁹ Another study of all fatal crashes during 1996–2007 reported a 41% reduction in crash rates for states with a good GDL rating and an 18% reduction for states with a fair GDL rating.¹⁷ Another analysis of 1986–2007 fatal crashes in the USA reported that among those age 16 years, GDL was associated with a 16% reduction (95% CI 6% to 25%) in crash rates for jurisdictions with either a night-time driving restriction starting before 1 : 00 or a passenger restriction allowing no more than one young passenger, and a 26% reduction (95% CI 16% to 35%) for jurisdictions with both restrictions.¹⁶ In a Cochrane review, five GDL programmes were summarised with a median decrease of 15.5% (range 8% to 27%) for the adjusted rate of all crashes during the first year and 21% (range 2% to 46%) for the adjusted rate of crashes with an injury among 16 year olds.¹⁸

Our estimate of association (aRR 0.94, 95% CI 0.93 to 0.96) between GDL and traffic crash rates among 17 year olds is consistent with previous studies for this age group for the entire USA. A study of all fatal crashes during 1996–2007 reported a rate ratio of 0.81 for states with a good GDL rating and 0.97 for states with a fair GDL rating.¹⁷ Another study of fatal crashes during 1986–2007 reported an aRR of 0.98 (95% CI 0.92 to 1.04) for GDL with night-time driving or passenger restriction, and an aRR of 0.91 (95% CI 0.83 to 1.01) for GDL with both restrictions.¹⁶

Our study did not find that GDL implementation was related to traffic crash rates among 18 year olds (aRR 1.00, 95% CI 0.95 to 1.04). A study of all fatal crashes for 18-year-olds in the USA during 1996–2007 reported a rate ratio of 0.96 for states with a good GDL rating and 1.03 for states with a fair GDL rating.¹⁷ However, another study of all fatal crashes in the US during 1986–2007 reported that GDL with either a night-time driving restriction or passenger restriction was associated with a 10% increase (aRR 1.10, 95% CI 1.03 to 1.18) in fatal crash involvements among those age 18.¹⁶ Further research is needed to evaluate whether GDL negatively affects 18 year olds.

GDL implementation was associated with a 22% reduction in traffic crash rates among 16 year olds, but only a 6% reduction in rates among 17 year olds. GDL implementation was unrelated to crash rates among 18 year olds, but this exploratory finding was based upon a sample of only four jurisdictions and should be treated with caution.