REF-1607 · Halperin et al., 2022 · Sports Medicine · Paper · PubMed

Pooled over 12 studies (414 healthy participants, 262 effect sizes), people who guessed how many reps they had left before failure were off by about one rep on average, guessing too few (0.95 reps, 95% CI 0.17-1.73), with very large heterogeneity (I² 97.9%); the guesses were slightly better on sets of 12 reps or fewer and in later sets, the gain nearer failure was small and its interval crossed zero, and training status and upper- vs lower-body exercise showed no clear effect. Exploratory meta-analysis, from the abstract

Main caveat: Abstract only; no risk-of-bias or certainty assessment is reported in it.
moderate evidenceMeasured on non-athletes to highly trained athletesRead from the abstract only

Halperin and colleagues collected studies in which healthy people said, before or during a set, how many more reps they could do, and then carried the set on to task failure so the guess could be checked. Of 16 eligible publications, 13 (12 studies, 414 participants) entered a multilevel meta-analysis of 262 effect sizes in 12 clusters.

On average the guesses fell short of what people then managed, by 0.95 reps, and the studies disagreed a great deal.

Meta-regressions linked slightly better guesses to sets that allowed fewer reps (heavier loads), to later sets and, with an interval that crossed zero, to guesses made nearer failure; training status and upper- versus lower-body exercise showed no clear effect, and people differed little from one another.

The authors conclude that participants were imperfect at predicting proximity to failure whatever their training background, that it is not yet known whether this much error is acceptable, and that predictions improve near failure, with heavier loads and in later sets. They ask future studies to report how failure was explained to participants and how it was confirmed.

Participants

414 healthy participants across 12 studies. Age, sex and training levels are not given in the abstract.

Protocol

Inclusion: healthy participants predicting the reps they could complete to task failure in a resistance exercise, before or during a set that was then performed to failure.
Analysis: a multilevel model over all 262 effect sizes (12 clusters); meta-regressions on proximity to failure, reps performed to failure (≤12 vs >12), set number, training status and upper vs lower body; between-participant standard deviation of accuracy.

Key Results

OutcomeEstimate (95% CI)
Mean error (underprediction of reps to failure)0.95 reps (0.17 to 1.73)
HeterogeneityQ(261) = 3060, p < 0.0001, I² = 97.9%
Prediction made closer to failureβ = −0.025 (−0.05 to 0.0014)
Sets of ≤12 reps to failureβ = 0.06 (0.04 to 0.09)
Sets of >12 reps to failureβ = 0.47 (0.44 to 0.49)
Later set numberβ = −0.07 reps (−0.14 to −0.005)
Training statusβ = −0.006 reps (−0.02 to 0.007)
Upper minus lower body−0.58 reps (−2.32 to 1.16)
Between-participant SD of accuracy1.45 reps (0.99 to 2.12)
Direction: people guessed fewer reps than they then completed, so on average the guess was conservative.
Near failure: the authors describe a slight improvement; its interval runs from just below to just above zero (our reading), so this effect is small and uncertain.
Heavier loads: the abstract does not say what the two rep-range coefficients are measured per; the authors read them as better accuracy when fewer reps were possible, i.e. with heavier loads.
Training status and body region: both intervals include zero. That is not evidence that experience makes no difference, only that none was detected.
Between people: the authors call the variation between participants minimal (SD 1.45 reps).

Practical Application

Across 12 studies of healthy participants (ages not given in the abstract), guesses of reps left before failure were about one rep off on average (people thought fewer were left than there were), and training experience showed no clear effect. The calls were slightly better on heavier sets of 12 reps or fewer and in later sets, so those are the sets where a rating deserves most trust; training background alone did not make the calls accurate.
Inference, not tested here: with an average error toward saying too few, a lifter who stops at "2 left" has probably stopped with a little more in hand, but the spread between studies is so large that any one person can be off in either direction.
Whether practice makes the calls better is not tested in this review: in adults new to the method, six weekly sessions with every set taken to failure cut the error (REF-1609); in nine trained men, six weeks did not change it significantly (REF-1610).
The base's own practice, not this paper's finding: SYNTH-101 writes the next load from the last done sets and their RPE; this review is a reason to take that RPE from a heavy set and after the first set, rather than from a long light set.
It says nothing about rating a ride or a run; for endurance ratings see REF-1611 and REF-1612.

Limitations

Abstract only; no risk-of-bias or certainty assessment is reported in it.
A scoping review with a meta-analysis the authors label exploratory; 12 studies with very large heterogeneity.
How failure was explained and confirmed varied between studies, which the authors flag as a cause of the spread.
The search ended in January 2021, so later studies such as REF-1485, REF-1609 and REF-1610 are not in it (our reading of the dates).
Participant ages, sexes and training levels, and which exercises were used, are not in the abstract.
How this entry was checked and changed
Why this level

capped by rule 1: risk of bias not read (abstract only), and the abstract reports no risk-of-bias or certainty assessment; the authors themselves call the meta-analysis exploratory, and the spread between studies is very large (I² 97.9%)