INTRODUCTION

Few biomechanical measures of running have received as much interest in both the scientific and lay press as that of the runner’s step frequency (SF). This is likely due to its simplicity of measurement (i.e., steps per minute, spm, synonymous with cadence) as well as to its role as a system-level output of the individual’s mechanical patterns (2, 16, 24). Thus, SF has received attention from clinicians and coaches alike for its utility as a point of manipulation for both therapeutic and performance-related purposes (9, 35).

Elite and recreational runners alike select SFs in the range of 160–200 spm for the typical distance running velocities of 3.0–5.5 m/s (14, 38). It has been repeatedly demonstrated that runners self-select a SF close to that which is energetically optimal (6, 21, 36). Elevating the SF of a runner, however, can reduce energy absorption at the hip and knee joints (19) as well as the leg stiffness and peak vertical forces (28). Conversely, reducing SF in runners increased their vertical loading rates (20). Observations of high-performing runners (9) and anecdotally injury-free runners (13) have given rise to a pervasive notion in popular running culture that runners should select a SF of at least 180 spm. It has yet to be fully elucidated as to why or if runners “should” run at a given SF and why such large interindividual variability persists (4).

Despite the large body of work surrounding SF in running, little of the research has been field-based, with much of the work done in the laboratory setting. This could be problematic, as runners have been observed to select a higher SF on treadmills compared with overground running (34, 40). Furthermore, much of the work, whether in laboratory or in field, has assessed SF as a stationary variable for an individual at given speeds (35). Given the intraindividual variability in the biomechanics of a runner day to day (3) and within a run (12), the temporal dynamics within a running session should not be ignored. Stride times of runners (i.e., inverse of SF) have been shown to exhibit persistent long-range correlations over a time series (23), and experienced runners have demonstrated lower coefficients of variation (CVs) in their stride lengths while running (32).

Elite runners (i.e., individuals competitive at national and international levels) offer a valuable study population when exploring regulation of SF due to their high level of training, experience, and capacity, which manifest as a greater degree of “self-optimization” in mechanical patterns (3, 5, 41). Ultralong-distance races (e.g., ultramarathon) represent a unique model to study the responses of the body to prolonged loading and fatigue and the consequences of such stress on running mechanics (27, 29). Competitive racing scenarios also provide internal and external motivations for success that can yield performances not possible with isolated laboratory observations (42), revealing valuable insights of how a mechanical variable such as SF behaves under fatigue as the body approaches peak performance capacity.

Investigation of SF during elite-level competitive road races has been minimal. Buckalew et al. (1) captured kinematics of national-class female runners at four points in a marathon. They reported that SF tended to remain constant throughout the race (i.e., average of 186 spm at miles 9, 19, and 20), declining only slightly to 182 spm with a corresponding decrease in horizontal velocity (i.e., 4.4 to 3.7 m/s) in the final stage of the race at mile 24. Similarly, in an elite 5-km road race, Hanley et al. (18) reported decreases in speed and SF across three checkpoints in the race. While yielding important information, their study provided a coarse view of how highly trained elite runners modulate SF in a high-performance setting. Several studies have reported increases in SF during and after ultramarathons compared with a pre-competition state (17, 10, 26, 29, 30), but none have recorded SF continuously during the race, and all used prescribed speeds for measurement.

Continuous recording of environment-specific SF has recently been made possible with the advent of wearable technology, specifically with accessible low-profile accelerometers. Commercially available wrist-worn devices with embedded triaxial accelerometers have been shown to be valid and reliable tools for measurement of step count in running (7, 15, 33), but no studies have used this technology to investigate SF dynamics during running.

The present study recorded the SF patterns of highly trained ultramarathon runners in an elite-level competition and explored the effects of fatigue (i.e., distance into the race), pace (i.e., running speed), and intrinsic characteristics (i.e., anthropometry, age, and experience) on SF. The null hypotheses tested were: H1) SF remains constant throughout the race, H2) SF is not influenced by running speed, H3) SF is not influenced by an individual runner’s intrinsic characteristics [i.e., height (a), weight (b), age (c), and years of experience (d)]. To explore the patterns of SF variability in the runners, the same hypotheses were tested on the SF CV throughout the race, denoted as H4, H5, and H6, respectively.

METHODS

Race characteristics.

The subjects were competitors in the 2016 International Association of Ultrarunning (IAU) 100-km World Championship. Competitors were selected for participation under qualification guidelines set by their respective national federations and in accordance with the IAU. The competition took place in Los Alcazares, Spain, on November 27, 2016. The race course was a flat 10-km loop at sea level (i.e., altitude range was 1–6 m) that was repeated 10 times. The 10-km loop (i.e., map with kilometer marks; Fig. 1) consisted of a surface with 7.5 km of paved asphalt and 2.5 km of paved floor tiles. Temperatures ranged between 10°C at the start of the race (0600 GMT) to 17°C when all participants had finished (1500 GMT). Runners (n = 137) from 33 nations finished the race. In addition to individual competition, the cumulative time from each nation’s top three performers was used for the team competition among nations, incentivizing race time in addition to placing.

Download figureDownload PowerPoint

Subjects.

The top 25 male and female finishers in the race were recruited for participation in the study. Individuals were accepted for inclusion if he/she used a continuously recording wrist-worn device with an embedded accelerometer that reported step count (e.g., Garmin Forerunner 235; Garmin, Olathe, KS) during the entire competition. Subjects were recruited and consented to participation in accordance with the ethical guidelines approved by the University of Michigan’s Institutional Review Board.

Data collection.

Following verification of device qualification, subjects provided the investigators with the watch’s record from the event, where the time, SF, and sampling rate during the session were extracted for analysis. Device manufacturers included Garmin (n = 12; Garmin Olathe, KS), Suunto (n = 7, Suunto Oy, Vantaa, Finland), and Polar (n = 1, Polar Electro, Bethpage, NY). Mean sampling periods varied from 1 to 6 s across devices (i.e., per individual device settings, where the sampling period is the elapsed time for each point of data capture). Finishing time, place, and individual 10-km lap times for each subject were extracted from the official time-keeping records provided by the race.

Data processing and statistical analysis.

Each subject’s record of SF and time from the wrist-worn device were matched to its corresponding lap in the race by using the race’s official recorded lap times. The mean SF for each lap was calculated along with its standard deviation over each lap. A linear mixed-effects model was used to study the individual SF patterns across the race. With SF as the dependent variable, the fixed effects were distance into the race, running speed, sex, height, and years of running experience. Individual subjects were modeled as random effects, and R² values for the mixed-effects model were estimated (31). The model was structured as follows:

$$\begin{array}{l}SF={\beta }_0+{\beta }_{\text{distance}}\times {\beta }_{\text{speed}}+{\beta }_{\text{sex}}\times {\beta }_{\text{height}}+{\beta }_{\text{age}}+{\beta }_{\text{weight}}+\\ {\mathrm{\hspace{1em}\hspace{0.17em}\hspace{1em}\beta }}_{\text{experience}}+Random\left(Subject\right).\end{array}$$

A second model was used with the same fixed and random effect structure to study the variability in SF throughout the race. The CV of the lap SF was used as the dependent variable (i.e., 10-km lap’s standard deviation/mean). For all models, the linearity assumption was verified by randomly distributed residual plots, and the normality of their distributions were verified with q-q plots. Coefficients of the model were tested with a Wald test against the null hypothesis that β = 0, and the effect was deemed significant if α < 0.05. Additionally, an independent-samples t-test was used to characterize general sex differences. All data processing and statistical analyses were performed with R Statistical Programming (R Foundation for Statistical Computing, Vienna, Austria).

RESULTS

Twenty subjects participated in the study (12 male, 8 female; Table 1). Finishing times for the 100-km race ranged from 6:37:23 to 6:56:56 (males) and 7:34:25 to 8:23:39 (females).

The average SF of all participants during the race was 182.0 ± 12.1 spm (mean ± SD). Males ran with an average of 177.6 ± 12.0 spm, while females ran with an average of 188.5 ± 9.0 spm (P < 0.001). After control for all other factors, SFs were not significantly different between the sexes (P = 0.47), and SF progression for each runner over the course of the race is illustrated in Fig. 2.

In the modeling all factors together, the full mixed-effects model had a conditional R² = 0.99 (i.e, fixed effects and random effects of subjects) and a marginal R² of 0.54 (i.e., only fixed effects considered). The SFs of the individual runners were significantly affected by their running speed (β_speed = 5.6 spm/m·s⁻¹, P = 0.01), with faster speeds eliciting higher SFs (Fig. 3). A runner’s stature also had a significant effect on SF (β_height = −123.3 spm/m, P = 0.01), with taller runners having lower SFs. SFs of these runners did not significantly change across 100 km of running (P = 0.25). Neither age, weight, nor years of running experience exerted a significant effect on the SFs of these runners. Significant interactions between running speed and distance or between stature and sex were not observed (Table 2).

The average SF CV throughout the race was 4.2 ± 3.6%. Similar to SF results, SF variability was significantly affected by running speed (β_speed = −3.9% spm·km⁻¹·min⁻¹, P = 0.008), with faster speeds eliciting lower variation in SF. Distance had a significant effect on the SF variability, as within-lap variability decreased throughout the race (β_speed =−0.4% spm/10-km lap, P = 0.003). Sex, stature, age, weight, and years of running experience did not have significant effects on SF variability (Table 3).

DISCUSSION

This study analyzed continuous recordings of SF taken by elite-level runners during an ultramarathon road race in order to evaluate the extent to which distance, speed, and runner characteristics affected SF profiles and variability during competition. These elite runners revealed that SF increased when running speed increased but was not affected by distance into the race, supporting null hypothesis H1 but rejecting the null H2. Individual SF was linked to body stature, but not to age, weight, or experience, rejecting H3a and supporting H3, b–d. Variability in SF, measured by CVs, was influenced by both speed and distance, with SF variability decreasing throughout the race and faster running speeds being associated with lower variations in SF, rejecting null hypotheses H4 and H5. The characteristics of the runner had no influence on the SF variation, supporting H6.

These findings align with observations of recreational and trained runners in laboratory settings with regard to the influences of speed and body characteristics. Cavanagh and Kram (4) found that a 1 m/s increase in speed on a treadmill corresponded to a 7-spm increase in SF in recreational runners, similar to the observation in the current study that a 1 m/s increase was associated with a 5.6-spm increase in SF. However, Cavanagh and Kram found that the CV of the SF remained approximately the same, whereas in the present study it decreased with faster speeds. The SF at each of the five speeds in the Cavanagh and Kram study (i.e., 3.15–4.12 m/s) was calculated from the average stride length (SL) over 2 min at each speed, so the duration of capture and/or the environment may explain the discrepancy in their results. Using speeds similar to those in the current study (i.e., 3.3–4.4 m/s), Tokmakidis et al. (38) found that elite runners on a treadmill demonstrated an average rise in SF of 6 spm, similar to our current findings. Related to anthropometric variables, Tokmakidis and colleagues reported weak correlations of 0.26–0.20 between SL (and correspondingly SF) and stature and body mass, respectively, at 3.8 m/s. However, Cavanagh et al. (5) found a moderate correlation (r = 0.67) between SL and leg length (i.e., closely related to height) in elite runners but not in subelite runners ( = −0.10). The significant relation we observed in these elite ultramarathoners corroborates their conjecture that elite runners have “self-optimized” their stride characteristics.

We did not find a relation between SF and distance covered throughout the race. Given the nature of the competition, we considered distance covered to be a reasonable surrogate for fatigue accumulation (37). The lack of change in SF in response to fatigue is contrary to findings in both laboratory and trail ultramarathon studies exploring the relation between SF and fatigue. Hunter and Smith (22) observed a small but significant decrease of 2 spm in SF at a given speed during a 1-h fatiguing treadmill run, whereas Morin et al. (29) observed an increase in SF during a 24-h treadmill run. Similarly, mixed results have been reported in SF before and after mountainous ultramarathons. Increases in SF measured following competition (10, 17, 26) have been interpreted as the runners adopting a “safer” and “smoother” running style to reduce impact loads in response to the prolonged neural fatigue and accumulated muscle damage of the race. For example, Giandolini et al. (17) attributed an observed decrease in SF after a trail ultramarathon as being due to peripheral neuromuscular fatigue of the plantar flexors. Other investigations, however, revealed no change in SF following the completion of a trail ultramarathon (11, 39), but none of those investigations recorded the SF within the competition or run but rather used standardized speeds with measurements taken either away from the site or on a runway set up at the site; thus, it is unclear how the current within-race findings compare with those earlier reports. Furthermore, there are unique physical challenges of a mountainous ultramarathon that may influence SF, including greater elevation changes (e.g., greater eccentric muscle loading), prolonged bouts of hiking, slower average running speeds, and exposure to altitude and varied environmental conditions. We posit that SF was unrelated to distance covered in the current study because of the unique population of runners studied. Given their high-performance capacity in the ultramarathon, their stride characteristics were ostensibly more fatigue resistant than recreational or even highly trained but nonelite runners. Costill et al. (8), for example, found that elite distance runners had a greater proportion of fatigue-resistant muscle fibers than other highly-trained but not elite distance runners. Potentially, skeletal muscle composition of the individuals in the current elite population allowed them to maintain SF patterns throughout the ultramarathon. Alternatively, perhaps this population has an increased resistance to the central or peripheral neural fatigue that otherwise caused a decrease in SF in nonelite ultramarathon runners (17).

Interestingly, in the present study, SF variability was significantly related to fatigue, as the CV decreased over the race duration. Nakayama et al. (32) found that trained runners had lower CVs in their SLs across speeds compared with novices. Meardon et al. (25) found no change in the CV of SL when runners ran to fatigue on an indoor track, but the subjects in that study ran an average of 5.7 km at an average speed of 3.5 m/s, substantially slower than the subjects in this study. The decrease in SF CV with increasing running speeds observed in the current ultramarathon parallels that which was observed in SLs by Jordan et al. (23), where the CVs similarly decreased as subjects ran faster. Nevertheless, the interactive effects of fatigue, experience, and injury on movement variability in running are still being elucidated. Perhaps the decrease in SF variability throughout the ultramarathon is a manifestation of the “safer” or “smoother” running style proposed by Morin et al. (30) and unique to the fatigue in road running or in high-performance elite runners’ patterns.

In interpreting the results of the current study, the method of SF measurement must be considered. The data were collected from a variety of commercially available, accelerometer-embedded GPS sport watches. Several studies have explored the accuracy of the step counts of Fitbit and Jawbone devices with similar hardware (7, 33). The step counts of the wrist-based accelerometers were found to be highly error prone in free-living and walking; but at faster speeds associated with running, wrist-mounted consumer-based devices had errors of less than 1.5% compared with optical measurement of step counts. The better accuracy of wrist-based accelerometers for step counts in running was suggested to be due to the greater peak angular momentum of the arm swing in running, providing devices with distinct peak acceleration signals for each stride cycle (7). Although no validations have been published on the devices included in this study, their analogous structure and function to validated wrist-based consumer-grade devices provide evidence for the quality of their measurements. With respect to the users of these devices, it should also be noted that the data used in the study came from a population of runners who were voluntarily using them, which may invite a hypothetical sampling bias; perhaps the runners supplying the data are more aware of SF than their counterparts who did not use a device that recorded this metric. Another consideration with respect to the devices was the varied sampling rate. Several of the devices sampled every second, but most had a variable rate (i.e., range of mean sampling periods among the 20 devices was 1–6 s). Although this may have been insignificant in the aggregation of SF measurements over each lap used in the current study (i.e., 36–53 min/lap), the nonuniform sampling prevented us from exploring any structure to the SF variability beyond the trends in the CV [e.g., detrended fluctuation analysis (23) to examine the long-range correlations of the SF patterns]. Future field studies using wrist-based accelerometers for SF measurement should ensure that sampling periods are uniform and minimal (e.g., 1-s sampling).

The current results bring several novel insights about our understanding of SF in running. It was the first field study to continuously record SF in elite runners during a road race. The subject population and environment both presented unique value, where the runners studied represented individuals that have “self-optimized,” and thus likely ran with SFs that were most appropriate for high levels of both training and performance consistent with their individual characteristics. The environment studied (i.e., a repeated 10-km road loop) is in itself an interesting study model, as it created a repeated-measures design to examine SF and its variation under relatively consistent external conditions. The repetition isolated the effect of fatigue (i.e., distance into the race) and allowed us to interpret the results of the SF progression throughout the race without addressing confounding effects of route changes. Furthermore, the road environment was a setting in which millions of recreational and competitive runners carry out their training and race running activity, giving the observations ecological and external validity. As we found, using a consumer-grade wrist-based accelerometer device is a highly accessible approach that can be replicated in diverse populations and settings to expand our understanding of SF and validate or challenge findings from laboratory studies. The continuous recording of SF with these devices allowed us to explore variability patterns in SF. Studies on running biomechanics tend to use limited measurement windows, as time, resources, data processing, and subject availability tend to limit long-duration recordings. The running activities that millions of people undertake for health or training reasons, however, are sessions that can last anywhere from 20 to 30 minutes to hours. To approach a richer understanding of a runner as a system, it is essential to capture the temporal dynamics of their biomechanical parameters throughout a run.

As a final observation, an intriguing finding for these elite runners was the amount of variability in their SF patterns explained by the measured factors. The marginal R² of the SF model was 0.53, indicating that ~53% of the variability in SF of these runners was explained by the included fixed effects: speed, distance, sex, height, age, weight, and years of training. The full model, with the random effect of the individual subjects, had a conditional R² of 0.99. This could be an indication that approximately one-half the variability in an individual’s SF could be predicted by these measureable characteristics, whereas the other half of the variation was due to unmeasured factors unique to each individual. That interpretation parallels the observation of Cavanagh and Kram (4) that “extrinsic variables” such as speed and anthropometry may partially explain an individual’s SF but that many difficult-to-identify “internal variables” (e.g., muscle fiber type, physiology, and training characteristics) likely play a significant role. Although speed and stature had predictable effects on SF in the present study, significant variations still existed between individuals that were not fully captured by “extrinsic variables” alone. Thus, runners, therapists, and coaches seeking to manipulate SF should understand how certain factors do or do not affect SF, such as speed and fatigue, and they should also seek to understand and appreciate the patterns unique to the individual and avoid overarching dogmatic or rigid prescriptions of SF.

DISCLOSURES

No conflicts of interest, financial or otherwise, are declared by the authors.

AUTHOR CONTRIBUTIONS

G.T.B. conceived and designed research; G.T.B. performed experiments; G.T.B. analyzed data; G.T.B., J.M.Z., and R.F.Z. interpreted results of experiments; G.T.B. prepared figures; G.T.B. drafted manuscript; G.T.B., J.M.Z., and R.F.Z. edited and revised manuscript; G.T.B., J.M.Z., and R.F.Z. approved final version of manuscript.