We pre-registered our meta-analysis at PROSPERO (ID number: CRD42021236384) and followed the PRISMA guidelines (Page et al., 2021). In the analysis, separate effects were estimated for HIIT and for conventional training as control, then combined to estimate the net effect of HIIT; we, therefore, also included studies of HIIT without control groups because the controlled trials effectively provided controls for these studies (Weston et al., 2014). Studies were eligible for inclusion if the outcome was some measure of endurance or sprint performance in highly trained endurance athletes [runners (Smith et al., 2003; Stöggl and Sperlich, 2014; Menz et al., 2015; Salazar-Martinez et al., 2018), cyclists (Stöggl and Sperlich, 2014; Menz et al., 2015; Salazar-Martinez et al., 2018; Clark et al., 2014; Hanstock et al., 2020; Laursen et al., 2002; Skovereng et al., 2018; Stenqvist et al., 2020; Sylta et al., 2016; Rønnestad and Vikmoen, 2019; Stepto et al., 1999), duathletes or triathletes (Smith et al., 2003; Stöggl and Sperlich, 2014; Menz et al., 2015; Salazar-Martinez et al., 2018; Hanstock et al., 2020; Laursen and Jenkins, 2002), cross-country skiers (Sandbakk et al., 2013; Stöggl and Sperlich, 2014; Johansen et al., 2021; Sandbakk et al., 2011), and rowers (Stevens et al., 2015)], with mean baseline V ˙ O_2max values ≥60 mL kg^–1·min^–1 for men (Sandbakk et al., 2013; Smith et al., 2003; Stöggl and Sperlich, 2014; Menz et al., 2015; Salazar-Martinez et al., 2018; Clark et al., 2014; Hanstock et al., 2020; Laursen et al., 2002; Skovereng et al., 2018; Stenqvist et al., 2020; Sylta et al., 2016; Rønnestad and Vikmoen, 2019; Stepto et al., 1999; Johansen et al., 2021; Sandbakk et al., 2011; Stevens et al., 2015) and ≥55 mL kg^–1·min^–1 for women (Sandbakk et al., 2013; Stöggl and Sperlich, 2014; Menz et al., 2015; Sandbakk et al., 2011), or in healthy other elite athletes [first league, national team, or international level in ball (Wells et al., 2014; Helgerud et al., 2001; Akdoğan et al., 2021; Iaia et al., 2015; Purkhus et al., 2016; Soares-Caldeira et al., 2014; Thomassen et al., 2010; Venturelli et al., 2008; Chtara et al., 2017; Hermassi et al., 2018; Selmi et al., 2018; Delextrat et al., 2018; Dupont et al., 2004), racket (Liu et al., 2021), canoe sports (Sheykhlouvand et al., 2018a; Sheykhlouvand et al., 2018b; Sheykhlouvand et al., 2016; Yang et al., 2017), and alpine skiers (Breil et al., 2010)]. We included one study on the upper-body exercise of male endurance athletes with V ˙ O_2max values of ∼55 mL kg^–1·min^–1 as their running V ˙ O_2max values met our inclusion criteria (Johansen et al., 2021). We accepted endurance athletes from different disciplines but did not include a mixture of endurance and other types of athletes within a single study. Studies in which training and testing were unrelated to the specific sport under investigation were excluded (Seo et al., 2019; Ojeda-Aravena et al., 2021; Herrera-Valenzuela et al., 2021; Thom et al., 2020; Sarkar et al., 2021). There was no eligibility for age, sports, training periodization, sample size, or international ranking of a national sports team. For two research groups that each presented the same outcomes for the same training and athletes in separate studies, the more recent study was excluded (Stöggl and Bjorklund, 2017; Laursen et al., 2005). Two different training studies of partially the same athletes were included as separate studies (Menz et al., 2015; Salazar-Martinez et al., 2018). The inclusion criteria for the HIIT interventions were at least two training sessions per week for at least a week and at least five sessions. Measures of V ˙ O_2max estimated from time trials were excluded. HIIT studies that included other types of interventions (e.g., ingestion of stimulants or other supplements, change in the environmental condition, or the use of modalities like compression garments) were excluded. HIIT exercise intensities had to be ≥90% of the maximal heart rate (Sandbakk et al., 2013; Breil et al., 2010; Stöggl and Sperlich, 2014; Menz et al., 2015; Johansen et al., 2021; Sandbakk et al., 2011; Helgerud et al., 2001; Seo et al., 2019), ≥90% of V ˙ O_2max (Smith et al., 2003; Salazar-Martinez et al., 2018; Sheykhlouvand et al., 2018b; Sheykhlouvand et al., 2016; Yang et al., 2017), or classified as “all-out” efforts for speed (Clark et al., 2014; Hanstock et al., 2020; Wells et al., 2014; Akdoğan et al., 2021; Iaia et al., 2015; Purkhus et al., 2016; Soares-Caldeira et al., 2014; Thomassen et al., 2010; Venturelli et al., 2008; Chtara et al., 2017; Hermassi et al., 2018; Selmi et al., 2018; Delextrat et al., 2018; Dupont et al., 2004) or power (Laursen et al., 2002; Skovereng et al., 2018; Stenqvist et al., 2020; Sylta et al., 2016; Rønnestad and Vikmoen, 2019; Stepto et al., 1999; Stevens et al., 2015; Sheykhlouvand et al., 2018a). The outcomes for several groups in multi-arm studies were excluded due to low exercise intensity (Supplementary Tables S3–S9). We accepted peer-reviewed articles published in English or German.

The workflow is presented in Figure 2. On 18 February 2021, the electronic bibliographic databases PubMed, Scopus, SPORTDiscus, and Web of Science were scanned for relevant studies. The search string, conducted in the title, abstract, and keywords, is presented in Supplementary Appendix. The search results were imported into EndNote 20.1 (Clarivate Analytics, Philadelphia, PA, United States) for deduplication before being exported to Rayyan QCRI software (https://rayyan.qcri.org). After another deduplication, all authors independently screened the articles based on the title and abstract, and we identified 547 papers that met the inclusion criteria. Seven additional studies identified through references cited in eligible articles were checked for inclusion. Subsequently, these full-text articles were screened, and discrepancies in 28 studies were resolved by two authors (HPW and JB).

Two authors (HPW and JB) independently rated the included studies on a quality scale proposed by Galna et al. (2009) (Supplementary Tables S1, S2). Interrater discrepancies were resolved by consensus, and no studies were excluded due to low scores.

Three authors (JB, TS, and HPW) collected subject characteristics and quantitative data from pre- and post-tests. Two authors (WGH and HPW) searched for inferential statistics to obtain standard errors of each study estimate, decided on potential moderators, and conducted all statistical analyses and data presentations using Statistical Analysis System (SAS OnDemand for Academics, version 9.4, SAS Institute, Cary, NC). Raw data, when available (Stöggl and Sperlich, 2014; Clark et al., 2014) or digitizable (Breil et al., 2010; Skovereng et al., 2018; Rønnestad and Vikmoen, 2019; Sandbakk et al., 2011; Stevens et al., 2015; Purkhus et al., 2016; Thomassen et al., 2010) (DigitizeIt, 38108 Braunschweig, Germany, or WebPlotDigitizer Pacifica, CA, United States), were re-analyzed using the statistical model in the study using SPSS Statistics V.27.0 (IBM Corporation, Chicago, Illinois, United States). Some study estimates and uncertainties were digitized from figures (Sandbakk et al., 2013; Smith et al., 2003; Hanstock et al., 2020; Laursen et al., 2002; Stepto et al., 1999; Sandbakk et al., 2011; Iaia et al., 2015; Liu et al., 2021; Sheykhlouvand et al., 2016; Yang et al., 2017); the accuracy of data extraction was verified by re-digitization (HPW) and conspicuousness checks (WGH and HPW). Otherwise, HPW asked corresponding and/or co-authors for the required values (Smith et al., 2003; Menz et al., 2015; Salazar-Martinez et al., 2018; Hanstock et al., 2020; Stenqvist et al., 2020; Rønnestad and Vikmoen, 2019; Sandbakk et al., 2011; Stevens et al., 2015; Helgerud et al., 2001; Akdoğan et al., 2021; Chtara et al., 2017; Hermassi et al., 2018; Delextrat et al., 2018; Liu et al., 2021; Seo et al., 2019). Sufficient data for some variables were obtained (Menz et al., 2015; Salazar-Martinez et al., 2018; Hanstock et al., 2020; Skovereng et al., 2018; Stenqvist et al., 2020; Rønnestad and Vikmoen, 2019; Akdoğan et al., 2021; Liu et al., 2021), but responses were not received from some authors (Smith et al., 2003; Stevens et al., 2015; Helgerud et al., 2001; Chtara et al., 2017; Hermassi et al., 2018; Delextrat et al., 2018), while other authors had either deleted their data due to national regulations (Sandbakk et al., 2013; Sandbakk et al., 2011) or could not send the required information for all requests (Akdoğan et al., 2021; Liu et al., 2021).

Endurance and sprint performance measures were categorized into sprint speed/power, repeated-sprint ability, time–trial speed/power, peak speed/power, aerobic/anaerobic threshold, V ˙ O_2max, and exercise economy (Figure 1). Due to a scarcity of intervention studies on different types of athletes (classical endurance or other elite athletes and male or female participants), meta-analyses for some outcome measures were restricted to specific subgroups. For sprint speed/power, the meta-analysis was limited to non-endurance athletes and based on the mean speed of linear running between 10 m and 40 m. The longest sprint duration was 7.69 s for a 40-m sprint (Selmi et al., 2018). For the repeated-sprint-ability tests, the distances were between 20 m and 40 m, sometimes including a 180° turn halfway. The number of sprints was between 2 and 15, with a mean time of a single sprint of 6.0 ± 1.4 s and a total sprint time of 44 ± 17 s. The mean or total time was used for the percent effects of the interventions. Time-trial times were all for endurance athletes and lasted between 0.5 min and ∼1 h. The trials consisted of either fixed-distance (0.4 km–40 km) or fixed-time (40 min) running or cycling. Wingate tests on a cycle or rowing ergometer of 30 s or 60 s were also categorized as time trials.

Performance effects of the time trial were obtained via changes in the mean speed or power. For this, we converted the percent effects on cycling and rowing time or speed to percent effects on power by multiplying by 2.2 for cycling and 3.0 for rowing (Hopkins et al., 2001). For peak speed/power, we used the highest speed obtained during running incremental treadmill ramp tests or YoYo Intermittent Recovery running tests at Level 1 or 2 (YoYo IR1 and IR2) or the highest power during an incremental ramp ergometer cycle test. The aerobic and anaerobic threshold data were presented by the authors of the included studies as speed or power. When fractional utilization at the threshold was presented (Sandbakk et al., 2013; Smith et al., 2003; Sandbakk et al., 2011; Sheykhlouvand et al., 2018a), the values were converted to V ˙ O₂ by multiplying by V ˙ O_2max in the same test before the calculation of the percent change, its standard error, and confidence interval. For V ˙ O_2max, absolute (L·min⁻¹) or relative changes normalized to body mass (mL·kg⁻¹·min⁻¹) were accepted; if both were available, we used the relative changes. Exercise economy was estimated from V ˙ O₂ during running, rowing, or cycling at a submaximal velocity between 8 and 13.5 km h⁻¹ and 75 and 220 W. Oxygen cost was converted to exercise economy.

The mean effects of HIIT and usual control training were calculated as percent changes. The standard error of the changes was calculated for each subgroup using p-values, t-values, or confidence intervals. Where these statistics lacked, the standard error was imputed via the typical error of measurement as follows (Carr et al., 2011; Vandenbogaerde and Hopkins, 2011): available standard errors were converted to typical errors and averaged separately (via variances weighted by degrees of freedom) for experimental and control groups for the given outcome variable; these typical errors were assumed to apply to experimental and control groups in those studies lacking standard errors; all typical errors were then converted to standard errors by multiplying by √2 and dividing by √(group sample size) (HIIT spreadsheet). Confidence limits for each study estimate shown in forest plots (Figures 3–9) and Supplementary Tables S3–S9 were calculated by log-transforming the estimates and their standard errors and then back-transforming to percent values.

Tables 1–3 provide an overview of the sample size of the included studies and the moderators included in each meta-analysis. A few missing values of moderators were imputed, as shown in Supplementary Tables S1, S2. Nominal moderators were type of athlete (endurance or other), type of test (incremental or YoYo for peak speed/power), whether HIIT was performed as extra training or replacement, and phase of training, with values of 0, 0.5, or 1 for on-season, a mixture of on and off-season, and off-season. The type of HIIT was included as a numeric linear moderator [coded 1–6, thus from aerobic HIIT using classical intervals to SIT, as shown in Figure 1 and discussed in detail by Stöggl et al. (2024)]. Some studies performed a progressive HIIT training phase, for example, changing from HIIT type 1 of a long aerobic interval to HIIT type 2 of a short interval duration but higher intensity during the intervention period. For such studies, we included an average of two adjacent integers (e.g., 1.5). To account for the amount of training in each study, we used the weeks of training, which applied equally to the HIIT and control groups.

Table 3 provides an overview of the pre-test moderators. Distinct physiological processes might determine sprint and time-trial performance of different distances and durations; therefore, we included pre-test time as a numeric linear moderator. For peak speed/power, adjusted test duration was included to account for the design of the incremental test, in which longer stages and smaller increments result in lower values (Morton, 1994; Luttikholt et al., 2006). To account for that, we adjusted the duration to that of a time trial in which the same amount of work was done using the formula: adjusted duration = (pre-test duration) × (pre-test mean speed/power)/(pre-test peak speed/power). For repeated-sprint ability, the reported percent decrement representing fatigue arising from the number and duration of sprints was included. For the aerobic/anaerobic threshold, the possible modifying effect of different threshold intensities was taken into account by including the pre-test threshold as a percent of V ˙ O_2max. For V ˙ O_2max, the inclusion of the pre-test mean accounted for possible ceiling effects on improvements. For exercise economy, the intensity at which economy was measured was V ˙ O₂ expressed as a percent of V ˙ O_2max; for some study estimates lacking V ˙ O₂ (Skovereng et al., 2018; Sylta et al., 2016; Rønnestad and Vikmoen, 2019), it was estimated from the gross efficiency and workload by assuming the V ˙ O₂ corresponding to a respiratory exchange ratio of 0.85 (Skovereng et al., 2018; Rønnestad and Vikmoen, 2019).

Apart from extra training and the type of HIIT, all moderators interacted with treatment. After initial explorations, a moderator that distinguishes between running and cycling performance was not included. The following interactions were investigated but were unclear and not included in the final models: type of HIIT with the duration of time trials on time-trial performance; type of HIIT with the duration of sprints on sprint performance; and type of HIIT with itself (a quadratic effect) on sprint performance. We investigated the number of HIIT sessions as an additional moderator, but it did not contribute usefully to the analysis owing to its strong correlations with weeks of training. There were insufficient study estimates to include sex as a moderator for sprint speed/power, time-trial speed/power, and peak speed/power. There were also insufficient study estimates to include arm tests for sprint speed/power or any moderators for exercise economy.

Random-effects meta-analysis models were realized using the mixed-model procedure (Proc Mixed). In each meta-analysis, the dependent variable was the log-transformed-extracted percentage change (100 × natural log[1 + percent change/100]). Sample estimates were weighted by the inverse square of their log-transformed standard errors, with the residual variance set to unity in the mixed model to perform the weighting (Rønnestad and Vikmoen, 2019; Yang, 2003). The fixed effects were treatment (HIIT or control) and the moderators described above. The random effects were study identity and sample-estimate identity within studies (with separate variances estimated for HIIT and control groups). The square root of the sum of study and sample-estimate variances provided an estimate of heterogeneity [the tau statistic (Higgins et al., 2024)] for HIIT and control treatments, representing differences between settings in predicted mean effects not due to sampling variation. The random effects were estimated, allowing only for positive variance, but standard errors of the variances were used to calculate appropriate lower and upper confidence limits, assuming a normal sampling distribution for the variances.

We examined scatter plots representing random-effect solution values versus standard errors of the study estimates to identify potential outliers and publication bias (Neville et al., 2022). One extreme outlier (Hermassi et al., 2018) was deleted in the peak speed/power analysis (shown in Supplementary Table S6). Potential publication bias was evident for peak speed/power (Figure 10A). However, running the analysis with the seven data points potentially contributing to publication bias (Figure 10B) did not result in substantial changes in the predicted effects. We retained these points to avoid loss of precision. To further address publication bias, we conducted simulations in SAS and RStudio to estimate any potential bias arising from the publication of only statistically significant effects on athlete endurance performance. Our analysis indicated that for the magnitudes of the effects observed, publication bias is expected to be negligible (Wiesinger et al., 2024a). Scatter plots of random-effect solution values versus linear predictors were examined for evidence that the use of simple linear moderators was adequate; the choice of log-transformation for some moderators (pre-test V ˙ O_2max for V ˙ O_2max and pre-test time for sprint speed/power) and the use of a quadratic term (test time for sprint speed/power) was motivated partly by these plots.

The effects of HIIT and control training were predicted for representative values of moderators for each meta-analysis, assuming that HIIT was performed during the competition phase and as replacements for some conventional training (see Tables 4–10 for details). The effects of the nominal moderators (e.g., male vs. female athletes, endurance vs. other athletes, and incremental vs. YoYo) are reported as differences between levels. The numeric moderators are evaluated as the effect of approximately two standard deviations (the mean of the pre-test study SDs of subject characteristics, such as V ˙ O_2max, or the SD of study characteristics, such as study duration).

The smallest important and other magnitude thresholds are derived from the variability in the performance of top athletes from competition to competition, which differs between sports (Malcata and Hopkins, 2014). We chose the smallest important value of ±1% for mean power in a competition performance, which is appropriate for cyclists, rowers, and kayakers and which is conservative for runners, whose smallest important difference is ±0.3%–±0.5% (Malcata and Hopkins, 2014). For the sprint performance of the other athletes, and presumably also for repeated sprints, the smallest important enhancement is ∼0.8% (Paton et al., 2001); we therefore applied the slightly more conservative threshold of 1%. The thresholds for small, moderate, large, very large, or extremely large increases are 0.3, 0.9, 1.6, 2.5, and 4.0 times the within-athlete variability in competition performance, respectively (Hopkins et al., 2009); applying these factors to log-transformed performance and back-transforming, the resulting thresholds for increases are 1.0%, 3.0%, 5.5%, 8.6%, and 14%, respectively, while those for decreases are −1.0%, −2.9%, −5.2%, −8.0%, and −12%, respectively.

The changes in laboratory and field tests that would result in the smallest and other enhancements in competition performance remain to be established. However, we assumed that percent changes in performance in these tests would directly transfer into percent changes in performance in a competition when the changes in performance in these tests are expressed as percent changes in power or its equivalent (Malcata and Hopkins, 2014). The transfer is obvious in sprints and time trials, provided that these are performed at intensities similar to competition intensities. Changes in peak power should transfer into changes in time trials performed at intensities similar to the peak power. The well-known physiological relationship that predicts endurance performance as the product of V ˙ O_2max, exercise intensity as a percent of V ˙ O_2max, and exercise economy (Joyner, 1991) implies that percent changes in each of these predictors translate directly into percent changes in time-trial speed/power provided that the other predictors do not change. Accordingly, changes in V ˙ O_2max should transfer into changes in endurance performance executed at intensities similar to V ˙ O_2max, assuming no change in exercise economy; this transfer should also apply to longer-endurance performances executed at sub- V ˙ O_2max intensities. Changes in aerobic/anaerobic threshold performance and exercise economy should also transfer into endurance performance executed at intensities similar to those of the tests. All these assumptions apply to the performance of endurance athletes, but for some of the measures ( V ˙ O_2max, peak power, and threshold), there were sufficient data to estimate effects on other athletes (ball or racket sports and alpine skiing); for these effects and athletes, we assumed the same magnitude thresholds as for endurance athletes, but it remains to be determined how percent changes in the tests transfer to changes in competition performance.

Sampling uncertainty of the estimates from each meta-analysis is presented as ± 90% confidence limits and as quantitative chances of substantial and trivial magnitudes based on a Bayesian analysis with a minimally informative prior (Hopkins, 2022). A probability (chances/100) and its complement (1 minus the probability) are p-values for tests of the hypotheses that the effect has the given magnitude and does not have the given magnitude, respectively (Hopkins, 2022). An effect on the predicted mean performance was deemed to have adequate precision if it was potentially implementable (benefit/harm odds ratio >66) or was unlikely to be beneficial (chance of benefit <25%). Mean effects with conservative, adequate precision had either a probability of harm <0.1% or a probability of benefit <5% and are highlighted in bold in tables; up to five of the former independent effects could be implemented while limiting the overall risk of harm to the most unlikely (<0.5%), while up to five of the latter could not be implemented while limiting the overall probability of benefit to the unlikely (<25%). Moderator effects and heterogeneity SDs were deemed to have adequate precision if at least one substantial hypothesis was rejected at the 5% level (probability of a substantial magnitude <5% or very unlikely). Adequate precision was conservative if the hypotheses were rejected at the 0.5% level (probability of a substantial magnitude <0.5% or most unlikely) and are highlighted in bold in tables; the overall error rate for rejecting the substantial hypotheses with up to 10 such independent effects is controlled to <5%. For effects with adequate precision, probabilities of substantial and/or trivial magnitudes >25% were interpreted qualitatively using the following scale: 25%–75% indicates possible, some, or modest evidence; 75%–95% indicates likely or good evidence; 95%–99.5% indicates very likely, very good evidence, and; >99.5% indicates most likely, strong evidence (Hopkins, 2022).

Figure 11 shows a concise summary of the predicted mean effects of HIIT and control training across all performance measures in settings defined by some practically important values of moderators. All measures, except exercise economy, showed clear substantial improvements in at least one setting, considering the uncertainties in the mean effects represented by thick error bars. For male endurance athletes, these clear effects were ∼5%. Effects for female athletes, where they could be estimated (aerobic/anaerobic threshold and V ˙ O_2max), tended to be greater than those for male athletes (∼10%). The enhancements in the team-sport-relevant performance measures (sprint and repeated-sprint ability) were similar to the enhancements in the other performance measures among male athletes. However, other athletes tended to experience greater improvements than endurance athletes in V ˙ O_2max and less improvement in the aerobic/anaerobic threshold and peak speed/power. Mean changes in performance with control training were generally less than with HIIT, and given their uncertainties, all true effects could be trivial.

Tables 4–10 complement the predicted HIIT effects summarized in Figure 11, adding detailed statistics for all subgroup comparisons, estimable moderator effects, and heterogeneity (random effect) SDs in the meta-analyses of each performance measure. Inferential comparisons of HIIT with control training confirm the effects of HIIT mentioned above.

For male endurance athletes, there was strong evidence for an increase in V ˙ O_2max (7.7%) and good to very good evidence for substantial improvements in time-trial speed/power, peak speed/power, and aerobic/anaerobic threshold (3.1%–5.3%). For female endurance athletes, the aerobic/anaerobic threshold and V ˙ O_2max showed good and strong evidence for increases (12.6% and 7.4%, respectively).

For male athletes, the effect of HIIT on sprint speed/power in a 5-s lasting sprint test was unclear when it replaced some conventional training, but the evidence for enhancements with 5 and 10-s sprint tests was otherwise good or very good (3.9%–6.7%). For these athletes, strong evidence and good evidence for enhancements were found for changes in V ˙ O_2max and repeated-sprint ability (8.6% and 3.9%, respectively), but the changes in the peak power and aerobic/anaerobic threshold were unclear. There was strong evidence of enhancement in V ˙ O_2max for female athletes (11.5%), while effects were unclear for other outcomes for these athletes (repeated-sprint ability and aerobic/anaerobic threshold).

Standard deviations representing heterogeneity in the effects of HIIT and control training between studies had adequate precision only for sprint and peak speed/power in the HIIT group, both of small magnitude with good evidence of substantial. Observed magnitudes otherwise ranged from moderate (one for HIIT and one for control) to trivial (one for control), with a suggestion of greater heterogeneity in the effects of HIIT.

Heterogeneity and its uncertainty add to the mean effects and their uncertainty to yield the prediction intervals for effects in specific new settings, as shown in Figure 11. The prediction interval is only slightly wider than the confidence interval of the mean effect for V ˙ O_2max and time-trial speed/power after 9 weeks of HIIT, but for most other measures, there is considerable widening due to heterogeneity, often a doubling of width.

These are summarized in Supplementary Table S10 as observed SDs in HIIT and control groups averaged across all studies in which they could be estimated. The means range from 1.4% to 3.5% in HIIT groups and 1.3%–4.4% in controls. The means for the HIIT groups are either similar to or greater than those of controls. Standard deviations representing individual responses to HIIT beyond those occurring with control training were derived from the errors of measurement and are also shown in Supplementary Table. SDs range from ∼1% to ∼7%.

The magnitude of the effects varies across the predictors and measures of performance among elite endurance and other athletes. The biggest effects of HIIT on V ˙ O_2max appeared to be greater than those on time-trial performances (Tables 6, 9). Formal inferential comparison of the effects on the various measures of performance is challenging because of differing moderator values and dependent-variable interdependence. Nevertheless, inspection of their confidence intervals indicates that the greater effect on V ˙ O_2max is not simply due to sampling uncertainty. It follows from the physiology of endurance performance (Joyner, 1991) that these greater improvements in V ˙ O_2max were accompanied by a reduction in either or both of the fractional utilization of V ˙ O_2max (and probably, therefore, aerobic/anaerobic threshold) and exercise economy. However, we found similar effects of HIIT on the aerobic/anaerobic threshold as time-trial speed/power and a trivial (but unclear) effect of HIIT on exercise economy. This apparent discrepancy may arise from including different study estimates and moderator selections for each outcome measure but underscores the need for mediation analysis, which is presented via relationships between the changes in these different predictors of endurance performance by Wiesinger et al. (2024b).

The moderator effects impact the implementation of HIIT in real-world sports, exploring the following key considerations: differences between the types of athlete and test; the type of HIIT; the optimal duration of HIIT intervention; the phase of implementing HIIT within the annual training cycle; and whether HIIT should augment or replace some conventional training.

Real differences between the effects of HIIT, characterized as SDs presenting between-study heterogeneity, were almost all substantial but estimated with inadequate precision. A comparison of heterogeneity for HIIT vs. control was therefore not attempted, but the observed tendency toward greater heterogeneity for HIIT is the expected result since there is likely to be more variation in the HIIT programs and, consequently, in their effects than in controls, even after adjustment for the known subject and study moderators.

Figure 11 shows that heterogeneity and its uncertainty have added substantially to the uncertainty in a mean effect for most measures and thereby weaken the strength of evidence for the effect of HIIT in an individual study setting. For some measures and mean settings (e.g., V ˙ O_2max for male and female endurance and other athletes), the 90% prediction intervals fall entirely in beneficial values, so HIIT is effective in at least 95% of individual settings. However, for some measures, HIIT could be ineffective or even harmful in a substantial proportion of individual settings, especially when the prediction interval in control training is taken into account. The most important measure for endurance athletes is time-trial speed power, and for other athletes, the most important measures are sprint speed/power and repeated-sprint ability. For these measures of performance, Figure 11 shows that there are individual settings where HIIT does better than control training, but control training could be equally effective, if not better, in other individual settings. We, therefore, cannot recommend HIIT unprovisionally for implementation in individual settings unless and until heterogeneity and its uncertainty are reduced sufficiently in an updated meta-analysis with more studies.

The mean errors in the control groups are similar to, or a little larger than, those in short-term reliability studies for trained individuals (Hopkins et al., 2001), implying relatively little contribution of individual responses to error of measurement with control training over the duration of the studies. Larger errors of measurement in the HIIT groups (for sprint speed/power, repeated-sprint ability, time-trial speed/power, peak speed/power, and exercise economy) reflect substantial contributions of individual responses to HIIT (SDs of up to 6.8%), but there is bound to be large uncertainty for these estimates arising from sampling variation. Sampling variation is also likely responsible for the lower errors of measurement with HIIT than with control training for aerobic/anaerobic threshold and V ˙ O_2max (resulting in negative individual-response SDs) because we would not expect HIIT to have fewer individual responses than control training with these two measures.

Although we found little evidence of publication bias across all the measures, our method for detecting and eliminating such bias may not be trustworthy if statistical significance was required for the publication of most of the study estimates. We, therefore, performed worst-case scenario simulations with the mean sample size (Seiler et al., 2013) and error of measurement (3.1%) for the HIIT group in the V ˙ O_2max studies to determine the published mean effect for different true mean effects when only significant effects are published. A true effect of 10% would result in negligible upward publication bias, and substantial bias becomes evident only for true effects of 5% (published mean effect 6%) [HPW and WGH (Wiesinger et al., 2024a)]. We can, therefore, be very confident that the meta-analyzed mean effects of ∼10% for V ˙ O_2max do not suffer from publication bias, but the meta-analyzed mean effects for the other measures could be overestimated by ∼1–2% in the worst-case scenario.