Driver see, driver crash: Associations between televised stock car races’ audience size and the incidence of speed-related motor vehicle collisions in the United States

Abstract

Several studies have examined road rage, but few studies have examined other psychosocial factors that may contribute to the number of motor vehicle collisions (MVCs). One study found increases in MVCs in West Virginia following televised NASCAR races but did not account for audience size. This study examined associations between NASCAR’s television viewership ratings and the incidence of speed-related MVCs in the USA using generalized estimating equations that controlled for seasonal effects, intoxication, road surface conditions, and lighting conditions. A 1% increase in the number of US households watching NASCAR races per month was associated with a 6.3% (95% confidence interval [CI], 3.0% to 9.7%; P < 0.001) increase in the incidence of speed-related MVCs—approximately 4911 (95% CI, 2353 to 7470) speed-related MVCs per month or one speed-related MVC per 595 (95% CI, 382 to 1354) viewers. As expected, similar results were not found for the total number of MVCs. These data suggest that televised NASCAR races may be associated with substantial increases in the incidence of speed-related MVCs. Making drivers aware of psychological factors that may increase risky driving behaviors could prove beneficial because self-monitoring can result in safer driving.

There have been multiple initiatives to prevent motor vehicle collisions (MVCs) by raising awareness about the dangers of various behaviors, such as intoxicated or distracted driving and road rage. One area with less empirical attention, however, is the potential effects of mass media on drivers’ aggression and risk-taking. One study found increases in aggressive driving MVCs in West Virginia following televised National Association for Stock Car Auto Racing (NASCAR) events.¹ The findings were attributed to priming and social contagion effects, where seeing behaviors rewarded increases the likelihood that observers will imitate that behavior—particularly when the actor is relatable to the viewer.^2–5 However, the West Virginia study faced several limitations, namely, its failure to account for television viewership ratings, which limited the ability to attribute the results to media exposure.¹ This study attempted to extend the West Virginia study and address many of its limitations. Specifically, we examined the association between NASCAR television audience size and the incidence of MVCs in the USA. Because speed-related MVCs constitute only a small portion of total MVCs, we predicted a positive association between NASCAR viewership and speed-related MVCs but not total MVCs.

Methods

Viewership ratings represent the percentage of US households with a television that watched the races as measured by the Nielsen company and have been used in similar studies of local events.⁶ Race dates and viewership ratings of NASCAR’s most popular racing division for 2004 through 2013 were obtained from an Entertainment and Sports Programming Network (ESPN) subsidiary devoted to covering NASCAR.⁷ After extraction, a 50% random sample was validated against a professional trade journal with less readily available ratings.⁸ There were no discrepancies between sources when comparing the ratings.

The estimated number of viewers was calculated from the ratings as Number of viewers = HH_TV × PPH × R%, where HH_TV is Nielsen’s annual approximated number of households with televisions^9–11; PPH is the annual average number of people per household based on Census data¹²; and R% is the rating in percent form. The ratings and estimated number of viewers were nearly perfectly correlated (r = 0.997, P < 0.001).

The MVC data were obtained from the US Department of Transportation, National Highway Traffic Safety Administration’s General Estimates System (GES).¹³ The GES uses police crash reports from a sample of police jurisdictions to create a sample-weighted database representative of the nation’s traffic ways. Though the GES data set is limited to crashes resulting in police reports, approximately half of all MVCs in the USA, it is unlikely that societally or health care–relevant crashes would not result in police reports.

To assess the potential association between NASCAR viewership and the incidence of MVCs, the GES data set was aggregated into monthly frequencies and NASCAR ratings were summarized into monthly means. Although this does result in some loss of granularity, the GES data do not provide the date of each crash, making it impossible to obtain more granular data that are still representative of US roadways. Additionally, because NASCAR races occurred nearly every week in the study period, exposure to the races was largely continuous. Thus, the effect of this loss of granularity on the soundness of the data is believed to be minimal. There were no live broadcasts of NASCAR races in the months of January or December in this 10-year period. As such, these months were excluded.

Outcomes were the total number of MVCs per month and the number of speed-related MVCs per month. The total number of MVCs per month was calculated by summing the sampling weights for each month in the GES data set. Speed-related MVCs were defined as MVCs where law enforcement indicated that speed was a contributing factor in the collision. This includes racing, exceeding the speed limit, driving too fast for the conditions, and other or unspecified scenarios where vehicle speed was at least partially responsible for the collision occurring. The number of speed-related MVCs per month was calculated by summing the sampling weights of cases meeting the above definition.

To control for clustering effects caused by the longitudinal nature of the data and to obtain population averaged inference, generalized estimating equations (GEEs)¹⁴ were used to model associations with the monthly incidence of MVCs. Based on inspection of quantile–quantile plots, the GEEs used Poisson distributions. Identity links were used to determine associations with the raw counts of outcome variables and log links were used to evaluate associations with incidence rates.

To account for seasonal effects and nonlinear effects of month (e.g., increases in MVCs resulting from holidays), months were used as the clustering variable. Thus, the GEEs treated calendar month as the subject variable and year as a repeated measurement within each month. Given these repeated measurements, the GEEs were specified with autoregressive working correlation structures, which controlled for within-month serial correlation. Robust standard errors were used to account for possible misspecification of the working correlation structure and link functions.

Covariates, which were mean-centered and standardized to make interpretation easier given the vastly different scales, included the number of MVCs with police-reported alcohol or drug involvement, the number of MVCs occurring on each day of the week with Fridays as the reference, the number of MVCs occurring on dry surfaces, and the number of MVCs occurring in daylight.

To determine the impact of including the ratings in the models, the GEEs were fit with and without NASCAR ratings in the equation, and both the change in the marginal R² analogue¹⁵ and Cohen’s f ^2,16 were computed. Cohen¹⁶ recommended cutoffs of ≥0.02, ≥0.15, and ≥0.35 to indicate small, medium, and large effect sizes, respectively. The marginal R² for the GEE including rating is denoted as $R_{\text{with}}^2$, the GEE without rating as $R_{\text{without}}^2$, and the difference as $\mathrm{\Delta }{R}^2$. All marginal R² values are from models using log links and follow the same interpretation as the traditional R².

As a sensitivity analysis, the above analyses were repeated with the months of January and December included. For these months, ratings were assumed to equal 1.00, which is less than half of the minimum rating in the actual data (2.38). Though arbitrary, this was deemed a reasonably low value that would not overly penalize any potential association of ratings.

Both data collection and analysis were performed in R (version 3.3.3, R Core Team, Vienna, Austria). Significance was assessed at P < 0.005, which has been recommended as the new cutoff for statistical significance.¹⁷

Results

In the 100-month data period, there were 390 live NASCAR races. These races had a mean ± standard deviation monthly rating of 4.1 ± 0.97 with an estimated 11,831,162 ± 2,668,967 viewers per race per month.

There were 470,594 ± 37,344 total MVCs per month. As expected, the incident rate ratio (IRR) of Table 1 shows no significant association between mean monthly NASCAR ratings and the rate of MVCs (IRR = 0.996; 95% confidence interval [CI], 0.988 to 1.003; P = 0.234) or the number of MVCs (−1455; 95% CI, −4816 to 1907; P = 0.396). However, this could be a result of the model without ratings almost perfectly predicting the number of MVCs and leaving very little variance unexplained ($R_{\text{with}}^2$ = 0.929, $R_{\text{without}}^2$ = 0.929, $\mathrm{\Delta }{R}^2$ < 0.001, f² = 0.008). As such, an unplanned GEE was performed with ratings as the only predictor; this model also found no significant association between ratings and total MVCs (IRR = 1.006; 95% CI, 0.999–1.022; P = 0.452; $R_{\text{with}}^2$ = 0.012; f ² = 0.012).

Table 1.

Associations with total number of MVCs per month

	IRR (95% CI)	B (95% CI)	Wald’s χ2 (P value)
NASCAR rating	0.996 (0.988 to 1.003)	−1455 (−4816 to 1907)	1.42 (0.234)
MVCs with alcohol involvement	1.012 (1.003 to 1.021)	5518 (1270 to 9765)	6.87 (0.009)
MVCs with drug involvement	0.994 (0.985 to 1.004)	−3183 (−8067 to 1701)	1.47 (0.225)
MVCs on dry surfaces	0.998 (0.983 to 1.013)	−1335 (−6907 to 6904)	0.07 (0.795)
MVCs during daylight	0.999 (0.992 to 1.006)	−1121 (−4398 to 2156)	0.04 (0.850)
MVCs on Saturday	1.034 (1.028 to 1.040)	15,479 (12,737 to 18,221)	114.68 (<0.001)
MVCs on Sunday	1.012 (1.006 to 1.019)	6122 (3036 to 9207)	14.07 (<0.001)
MVCs on Monday	1.027 (1.023 to 1.032)	12,480 (10,679 to 14,281)	152.17 (<0.001)
MVCs on Tuesday	1.023 (1.018 to 1.027)	10,504 (8252 to 12,757)	92.51 (<0.001)
MVCs on Wednesday	1.019 (1.011 to 1.026)	9175 (5820 to 12,531)	23.20 (<0.001)
MVCs on Thursday	1.035 (1.031 to 1.038)	16,166 (14,638 to 17,695)	381.51 (<0.001)

B indicates regression coefficient from generalized estimating equation using identity link; CI, confidence interval; IRR, incident rate ratio from generalized estimating equation using log link; MVC, motor vehicle crash. Wald’s χ² statistics and P values are from generalized estimating equation using log link.

There were 78,381 ± 16,001 speed-related MVCs per month. As hypothesized, there was a significant positive association with ratings and the incidence rate of speed-related MVCs. The IRR of Table 2 indicates that each unit increase in monthly ratings was associated with a 6.3% (95% CI, 3.0% to 9.7%; P < 0.001) increase in the incidence of speed-related MVCs—the equivalent of 4911 (95% CI, 2353 to 7470) speed-related MVCs per month. This equates to 1680 (95% CI, 739 to 2620) speed-related MVCs per million viewers or one additional speed-related MVC per 595 (95% CI, 382 to 1354) viewers. The inclusion of ratings resulted in a 3.5% increase in the proportion of variance explained by the model, with Cohen’s f² indicating a small to moderate effect size ($R_{\text{with}}^2$ = 0.556, $R_{\text{without}}^2$ = 0.521, $\mathrm{\Delta }{R}^2$ = 0.035, f² = 0.079). Unlike with total MVCs, NASCAR ratings had a significant positive association with the number of speed-related MVCs that explained a large proportion of variance in the absence of covariates (i.e., unadjusted; IRR = 1.095; 95% CI, 1.068 to 1.123; P < 0.001; $R_{\text{with}}^2$ = 0.310; f² = 0.449).

Table 2.

Associations with number of speed-related MVCs per month

	IRR (95% CI)	B (95% CI)	Wald’s χ2 (P value)
NASCAR rating	1.063 (1.030 to 1.097)	4911 (2353 to 7470)	14.59 (<0.001)
MVCs with alcohol involvement	1.125 (1.041 to 1.216)	10,752 (4967 to 16,535)	8.80 (0.003)
MVCs with drug involvement	0.913 (0.856 to 0.973)	−8584 (−13,517 to −3650)	7.75 (0.005)
MVCs on dry surfaces	0.934 (0.898 to 0.971)	−5666 (−9535 to −1797)	11.83 (<0.001)
MVCs during daylight	0.996 (0.979 to 1.014)	−547 (−1,951 to 858)	0.19 (0.663)
MVCs on Saturday	1.022 (0.975 to 1.071)	1755 (−1725 to 5,236)	0.81 (0.369)
MVCs on Sunday	1.012 (0.987 to 1.038)	830 (−1303 to 2963)	0.85 (0.356)
MVCs on Monday	1.038 (1.018 to 1.059)	3144 (1533 to 4755)	13.40 (<0.001)
MVCs on Tuesday	1.047 (1.018 to 1.076)	3629 (1210 to 6049)	10.25 (0.001)
MVCs on Wednesday	1.016 (0.987 to 1.047)	1287 (−1051 to 3626)	1.17 (0.279)
MVCs on Thursday	1.052 (1.005 to 1.101)	4217 (879 to 7555)	4.77 (0.029)

B indicates regression coefficient from generalized estimating equation using identity link; CI, confidence interval; IRR, incident rate ratio from generalized estimating equation using log link; MVC, motor vehicle crash. Wald’s χ² statistics and P values are from generalized estimating equation using log link.

Sensitivity analyses of the above results revealed similar results. There was no significant association between ratings and total MVCs (IRR = 0.999; 95% CI, 0.996 to 1.002; P = 0.566; $R_{\text{with}}^2$ = 0.950, $R_{\text{without}}^2$ = 0.949, $\mathrm{\Delta }{R}^2$ < 0.001, f² = 0.010). However, there was still a positive association between ratings and speed-related MVCs, although not a significant one (IRR = 1.013; 95% CI, 1.001 to 1.025; P = 0.038; $R_{\text{with}}^2$ = 0.711, $R_{\text{without}}^2$ = 0.709, $\mathrm{\Delta }{R}^2$ = 0.002, f² = 0.007).

Discussion

Consistent with predictions and extant literature, this study found significant positive associations between NASCAR viewership ratings and the national incidence of speed-related MVCs over the course of 10 years. While controlling for the effects of drug and alcohol intoxication, road conditions, lighting conditions, and season, a one percentage point increase in the average number of US households watching NASCAR’s most popular division per month was associated with a national increase of 4911 speed-related MVCs per month, approximately one additional speed-related MVC per 595 viewers. For comparison, the number needed to treat to prevent one death from cardiovascular disease in patients with a history of heart attack or stroke using daily, low-dose aspirin is 333, and the number needed to treat to prevent a first heart attack or stroke is 1667.^18,19 Also as expected, NASCAR ratings were not significantly associated with the number of total MVCs per month.

The link between mass media and changes in the frequency behaviors that are measurable on an epidemiologic level is not new. Multiple studies since the late 1970s through the late 1990s have found increases in fatal single-passenger–single-vehicle MVCs, which are suggestive of suicide, on the day of and in the days following highly publicized suicides.^20–24 The same is true for presumed murder-suicide by plane crash after highly publicized murder-suicides.²⁵ In addition to the temporal relationships, these behavioral imitation effects are often highly specific. For example, two studies^25,26 have shown that homicide rates increased following heavyweight boxing matches, and the increases were primarily in cases where the victim’s race corresponded to that of the losing fighter (e.g., when a Black boxer lost, there were more Black victims). Additionally, the effects can be quite large, as in one frequently cited natural experiment showing that from 1995 to 1998 eating disorders rose by 200% in Fiji due to the introduction of television and subsequent exposure to the thin ideal.²⁷

In addition to observational studies, experimental and quasi-experimental research has shown various forms of media can have substantial impacts on behavior. Most notable is the effect of violent media, with several studies finding increases in aggression after consumption of violent media. Importantly, the mean effect sizes of these studies are greater than the effect sizes of condom use in decreasing sexually transmitted HIV and of lead on children’s intellectual development.²⁸ Thus, these effects are clearly not trivial. However, due to the wide-reaching and recurrent nature of media exposures, even small effect sizes can influence society in meaningful ways, as seen in this study.

As noted in the introduction, this study was based on the West Virginia study;¹ in addition to not accounting for audience size, that study faced three major limitations. First, significant increases in aggressive driving MVCs only occurred on the fifth day after a NASCAR race; because most of NASCAR’s races occur on Sundays, this primarily indicates an increase in aggressive driving on Fridays. Second, West Virginia was chosen for its unique position as simultaneously having no NASCAR track in the state while also having the highest number of NASCAR fans per capita, which limits the generalizability of the findings to areas where NASCAR may be less popular. Lastly, the study examined collisions that were “a product of aggressive driving” (p. 493),¹ but not necessarily involving speeding or racing. By utilizing a nationally representative data set of police-reported MVCs, incorporating estimates of audience size, and showing a significant positive association with MVCs resulting from speeding, this study has avoided many of the weaknesses of the West Virginia study. Additionally, by incorporating estimates of television audience size, this study appears to be the first to examine whether there is a dose-response for this form of behavioral imitation, rather than simply establishing whether there is any increase.

Though no causal inference can be drawn from the results of this study, there is a plausible theoretical framework to explain the observed associations if one assumes that at least some small portion of the drivers involved in the speed-related MVCs watched NASCAR races in the month of the crash. Under social cognitive theory, it is possible that when viewers see NASCAR racers’ driving behaviors rewarded, this results in symbolic observational learning and creates an implicit (i.e., unconscious) association between fast driving and reward. This link could then be reactivated via priming effects when viewers get behind the wheel. This explanation satisfies the four mechanisms that govern observational learning.²⁹ First, attentional processes are fulfilled by the fact that viewers are watching the races, which are salient and—ostensibly—enjoyable. Second, the encoded association is retained because races are held weekly during the season and driving between the races can reactivate the association via priming. Third, driving—safely or not—provides an adequate behavioral production process that nearly perfectly matches the observed activity. Lastly, the prizes awarded to NASCAR drivers (e.g., acclaim, monetary reward) provide the motivational process, although likely unconscious. In addition to these required elements, the typically live broadcasts are frequently described as physiologically arousing to viewers,³⁰ which has been shown to increase the likelihood of learned imitative behaviors.³¹ Additionally, recent research suggests that repeated viewing of an action is associated with an often-untenable increase in self-perceived ability to perform the same action well.³²

Despite the above, this study faces several limitations. Foremost among these is the retrospective nature of the data and the related inability to directly link viewers to drivers. Further, there were other racing competitions televised during the study period, and the potential effects of these races were not accounted for in this study. Furthermore, sensitivity analyses suggest that these results may not generalize to January or December, months when there were no NASCAR races. Lastly, because Nielsen’s ratings represent the percentage of television-owning households in the USA that watched a show on television, (a) the ratings do not capture viewership via other media (e.g., races watched online or in person) or at other locations where large numbers of people may watch the races (e.g., restaurants); (b) within households that watched the races, it is unlikely that all household members watched the races; and (c) they do not differentiate between large groups of people gathering at a single household to watch the race versus one person watching independently. Though these last three points are important limitations, points (a) and (b) and the two scenarios of (c) are likely to counteract each other at least partially, thereby potentially reducing their effects.

Limitations notwithstanding, this study contributes to a large and growing corpus of literature showing a strong link between media exposure and subsequent, if not consequent, behavioral imitation. As such, though neither intended nor likely to make NASCAR viewers or fans of other racing competitions stop watching races, viewers should be aware of the potential effects that their viewership may have on both their safety and that of the public. Simply being aware of these possible effects might even be sufficient to reduce their risk because self-monitoring may offset the potential for harmful imitation behavior.³³ An intervention as simple as adding “Warning: professional drivers on closed course. Do not attempt” to NASCAR’s broadcasts may be sufficient and is likely warranted.