An empirical study of race times in recreational endurance runners

We explored factors associated with race performance and methods for race time prediction. One approach would have been to download databases of race times from large races. However, this would have allowed us to examine only sex and age, and not other predictors such as training, body mass index or history of injury. Accordingly, we developed a questionnaire that was implemented via the Internet to collect routinely available data from a large sample of recreational runners. Since our goal was to predict race performance, we used race velocity in meters per second (m/s) as the dependent variable. Independent variables were runner characteristics (sex, age and BMI) and training characteristics (typical training mileage and the use of sprint and tempo runs) that are known to be associated with race time.

The study questionnaire included items about age, sex, height and weight, training, injury, type of runner (on a 10 point scale from 1 - “endurance runner” - to 10 - “speed demon”), and type of running shoe (normal running shoe, minimalist, or Vibrams/sandals/barefoot). Questions on training specifically concerned the typical number of miles run per week and the maximum number of miles run in a single week leading up to the longest race, and whether the runner did interval training or tempo runs most weeks during training. Interval training involves short and intense periods at maximal effort followed by equal length or longer recovery periods of less strenuous exercise. Tempo runs are done at a steady pace at or slightly above the anaerobic threshold [13]. Participants were then asked to enter data for two or three recent races: distance and time plus subjective assessments of course difficulty (wind, hills, temperature) and fitness on race day.

The questionnaire was implemented via the Web on the news website Slate.com attached to a news story about race time prediction. Readers of an article criticizing the current approach to race time prediction were encouraged to “help Slate build a better predictor” and clicked on a link to the questionnaire. The text of the article made it completely clear that the deidentified data entered by participants would be used for data analysis by the study PI (AV). The project was discussed with the chair of the Memorial Sloan Kettering Cancer Center Institutional Review Board. Given that the project involved analysis of deidentified data and there was no potential for harm, it was deemed that no oversight was required. A copy of the questionnaire is included in the Additional file 1 (see “Questionnaire Text” on page 5- 8).

Use of the Internet to distribute a survey naturally raises questions as to two key aspects of survey validity: representativeness and selection bias. As regards representativeness, we planned to compare our sample with data on sex, age and race time from 50,266 participants in the 2013 New York Marathon and also with US national figures from Running USA [14]. We do not think that there is an important risk of selection bias in this study: there is no reason to believe that, say, a runner whose marathon time was faster than average given a certain 10 km time was any more or less likely to complete the questionnaire than a runner whose velocity slowed more dramatically from 10 km to marathon distance.

Procedures for data inclusion, for instance, if a respondent included more than one race at a given distance, are given in the Additional file 1. We aimed to assess the association between race velocity (in m/s) and age, sex, BMI, and training (average mileage, intervals, tempo runs) [15, 16]. Since velocity varies by race distance, we created separate linear regression models for each variable of interest for each race distance. A change in velocity of the same magnitude has different implications based on initial velocity, so we explored modeling the logarithm of velocity, representing relative change in velocity. However, this model was found to explain less variation than using velocity untransformed. We tested for non-linearity in age, BMI and typical weekly mileage, and included cubic splines for these three covariates to account for non-linearity.

We used BMI as a correlate of body fat. Body fat likely affects race time directly – additional weight reduces run velocity over long distances – but is itself affected by training, with intense training leading to fat loss [17]. In addition, for individuals who are not overweight – about 80 % of the current sample – women tend to have lower BMI than men. Hence, we took a slightly different approach for examining the influence of BMI on race time. We modeled the association between velocity and age, velocity and BMI, and velocity and typical weekly mileage. The models were adjusted for age, typical weekly mileage, sex, intervals and tempo runs. Since BMI is highly correlated with weekly mileage and race training, the models for the association between velocity and age and between velocity and weekly mileage were calculated with and without BMI. Since the models with and without BMI produced comparable results, we included BMI in our descriptive models.

We also sought to develop a prediction model for race time. We first split the data 2:1 into a training (n = 1443) and validation set (n = 721), using a randomization algorithm that ensured a similar distribution of marathon runners and marathon times in each cohort. We then used the training cohort to explore multiple models for race time prediction using linear regression. The survey asked runners to tell us the difficulty of each race for which they reported a race time (very difficult, difficult, average, fast, or very fast). Very difficult or very fast races (5.6 % of the data) were not included in the data used to build the prediction models on the grounds that few runners reported race difficulty at the extremes. For difficult and fast races, we adjusted times to be more representative of a runner’s time for an “average” difficulty race of that distance. To do so, we created a model to predict race velocity in m/s for each race distance separately, adjusted for race difficulty (difficult, average or fast), sex, age, typical weekly mileage, intervals and tempo runs. The differences in velocity between difficult and average races, and between fast and average races, calculated in these models were used to adjust reported race times for difficult and fast races by adding the model coefficient to the runner’s true velocity, and then calculating a difficulty-adjusted time using the difficulty-adjusted velocity. The coefficients used to adjust for fast and difficult races for each race distance are reported in the Additional file 1 along with the prediction models.

A common formula for predicting race time for a longer race y based on a shorter race x was published by Riegel and is of the form: time for race y ÷ time for race x = (distance of race y ÷ distance of race x) ^k [6]. The constant k represents the amount a runner’s velocity decreases as race distance increases (a “fatigue factor”). The models we tested included k as both a predictor, using a constant of k = 1.07, [6] and as the dependent variable, that is, using variables such as mileage to predict the correct value of k. Riegel cited a k of 1.08 for elite runners and 1.05 to 1.06 for male recreational runners aged 40 to 70 [6], while the Runner’s World online calculator uses a constant of 1.06 [1]. We chose a constant of 1.07 as an average, and performed sensitivity analyses using k = 1.06 and k = 1.08.

Models were compared using two metrics: mean squared error (MSE) and penalized mean squared error. Since overestimation of a runner’s velocity is more detrimental than underestimation – a runner who starts too slow can speed up during a race whereas a runner to starts too fast will slow dramatically - the penalized mean squared error was calculated by adapting the mean squared error formula, so that an overestimate of velocity had double the weight of an underestimate. All statistical analyses were two-sided, and significance was defined as p < 0.05. All analyses were conducted using Stata 13 (Stata Corp., College Station, TX). The full data set used in the analysis is provided as Additional file 2.

The survey opened on April 24, 2014 and was closed for analysis on June 16, 2014. There were 2,497 responses. Runners were excluded if all races reported were the same distance (n = 171), if they were duplicates (n = 8), if they had only one unique race with full data (n = 7), as data from these participants cannot be used in a race time predictor. Participants were also excluded if they reported running a longer race in a shorter time than a race of a shorter distance (n = 6), or if the responses were obviously erroneous or falsified (n = 2). Erroneous data also lead to exclusion of race times from individual races reported by a participant (see Additional file 1 for details). The final cohort included 2,303 runners, with all runners having full data on at least two races of different distances. Only 3 men and 4 women reported a time that met Olympic qualifying standards for at least one race distance, hence our sample consists almost completely of recreational runners.

Characteristics of the study cohort are given in Table 1. Time and training data are given in Table 2. A detailed comparison of our cohort to US national data and the New York marathon is given in the Additional file 1. The average age and male:female ratio of our sample is similar. Although running velocities are faster on average, we did have good representation of slower runners, with over 300 runners in our cohort reporting marathon times longer than 4 h.

We first examined multivariable associations between runner characteristics and race time, both overall and separately by sex. We excluded data from the 5 and 10 mile distance due to the relatively limited number of runners at these distances. Sex, age, BMI, typical training mileage, interval training and tempo runs were all statistically significant predictors of race time at p < 0.0005. Results adjusted for BMI were very similar to those without BMI. Interaction analyses were used to determine whether the effect of these characteristics differed between males and females. Significant interactions were found between sex and age, typical mileage, BMI and interval training (p = 0.001, p = 0.004, p = 0.025 and p = 0.024, respectively), although differences were small. Men who ran intervals had a marathon time faster by 4:46 min on average, compared to 3:07 for females. For a one unit in BMI, increase in marathon time was about 40 s more for men. The difference in marathon time for training 50 vs. 30 miles per week was 25:32 for men and 31:41 for women; for age 50 vs. 30 the difference in time was 16:18 for men and 21:37 for women. If these associations are seen to be causal, then we could say that interval training has a greater effect on men, training volume has a greater effect on women, BMI has a greater effect on men and aging has a greater effect on women.

Fewer than 2 % of our cohort reported wearing Vibrams, sandals, or running barefoot. We excluded those runners from analyses of footwear. After adjustment for sex, age, BMI and training, runners wearing minimalist shoes had significantly faster race velocity than those reporting conventional footwear (p < 0.0001). Difference in race time was close to 0.5, 1.5, 2 and 3 min for 5 km, 10 km, half-marathon and full marathon respectively. There was no evidence that the association between footwear and velocity differed by race distance (p = 0.4 for interaction term).

An additional question is whether the associations between race velocity and runner characteristics or training depend on race distance. For instance, we wanted to know whether the difference between male and female runners is greater for shorter compared to longer races. The results of these analyses are shown in Table 3 and Fig. 1a, b and c. Females do relatively better at longer distances, whereas tempo runs are associated with faster times more strongly for shorter distances. The association between interval training and race velocity was not found to differ by race length. Adjusting for BMI did not importantly influence these interactions. Differences between men and women were larger with adjustment for BMI, on the grounds that men are generally heavier; differences by intervals and tempo runs are smaller, suggesting that these types of training lower BMI. There was no significant interaction between race distance and age (p = 0.13) or between race distance and BMI (p = 0.4). Although there was a significant interaction between race distance and typical weekly mileage (p = 0.025), the magnitude was small, and the effect of typical weekly mileage on velocity did not differ importantly between race distances. These effects can be seen in Fig. 1a, b and c, where the curves for each race distance are approximately parallel.

We found evidence that the Riegel formula predicted race time reasonably well for distances up to a half marathon (see Additional file 1: Figure S1 and S2), but was poor for marathon prediction. Using a linear regression model for each race distance with the intercept set to 0, we tested whether the coefficient for the Riegel race time predictions was equal to 1, which would indicate good calibration. The coefficient for Riegel marathon predictions was significantly different than 1 (p < 0.0001) and therefore poorly calibrated, while there was no evidence of a difference for half marathon and 10 k times, implying good calibration (p = 0.3 and p = 0.9, respectively).

We explored various models for marathon time prediction using the training set. We settled on two approaches. Model 1 was used for runners who provided data from only one prior non-marathon race. Model 1 predicted marathon velocity in m/s using typical weekly mileage and the predicted marathon velocity calculated using the Riegel formula where k = 1.07 and that runner’s longest non-marathon race. For runners who provided data from two prior non-marathon races, we calculated k between the two races provided. Model 2 used k between the two shorter races and typical weekly mileage to predict k between the longer reported race distance and a marathon. Variables strongly associated with velocity on univariate analysis were not found to improve prediction after controlling for training mileage and prior races times. Hence our models included only prior race time and typical weekly mileage.

Both velocity and k were then converted to time in minutes so that mean squared error and penalized mean squared error could be calculated. We also compared the error in these models to the predicted times using the Riegel model where k = 1.07 and the predictor time is from the runner’s longest reported non-marathon race. The formula for both models is given in the Additional file 1 (see “Formulae for Prediction Models” on pages 3 – 4).

All mean squared errors were calculated on the same sample of runners in the validation data (N = 156) who reported running a marathon and two other races of differing non-marathon distances. Mean squared errors for model 1, 2 and Riegel were 227.6, 208.3 and 380.7, respectively. MSEs penalizing underestimation of race time were 646.1, 525.0 and 1429.8. Figure 2a, b and c show calibration plots. Calibration of the novel models (Fig. 2a and b) is excellent; in contrast, the Riegel formula (Fig. 2c) shows clear miscalibration with actual running times considerably slower than those predicted. Table 4 shows the distribution of differences between predicted and observed times. Predicted marathon times from Riegel are 10 min or more too fast for about half of all runners; for the new models, this drops to about 25 % of runners. Sensitivity analyses using different coefficients did not change our conclusion. Using the Riegel formula with k = 1.08 did improve estimates slightly, but MSEs were much higher than those from the novel models with marathon velocity overestimated for about 75 % of runners.