^{1}

^{*}

^{2}

^{3}

^{4}

^{3}

^{4}

^{2}

^{3}

^{5}

^{2}

^{3}

^{5}

^{1}

^{2}

^{3}

^{4}

^{5}

Edited by: Roberto Monastero, University of Palermo, Italy

Reviewed by: Alden L. Gross, Johns Hopkins University, United States; Quinn Kennedy, Naval Postgraduate School, United States

This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

Reaction time (RT) and RT variability are core components of cognitive performance that can be captured through brief and easy-to-administer tasks of simple RT and choice RT. The current study aims to describe age-related differences in cognitive performance, toward better characterizing normative performance across the lifespan. We examined mean and variability of response times on a simple RT and choice RT tasks in a large and diverse web-based sample (10,060 visitors to

Performance variability has been linked with cognitive decline (e.g.,

Variability in cognition may be a particularly important metric, in that individuals with lower baseline performance levels may be able to normally compensate across cognitive domains, whereas when their performance becomes less stable (more variable), such compensatory processes may fail (e.g.,

Lower RT variability (more consistent performance) has been associated with better cognitive control (

There have been inconsistencies in the literature, however, in terms of how RT variability differs across the lifespan. Different measures of RT variability can produce different results (

Here, we sought to compare and contrast different measures of cognitive performance and variability in a large, diverse sample to better characterize patterns of age-related change, establish updated norms, and compare across standard performance metrics in a large, well-powered sample. Specifically, we wanted to better understand the potential effect of aging on mean RT, RT variability (raw and mean-adjusted), and accuracy in measures of simple and choice RT. Our large sample size allowed us to look at differences in performance year-by-year across the lifespan, addressing sample size limitations of prior studies, and allowing us to estimate potential trajectories of cognitive change for different indices. We also compared different approaches to capturing RT variability to see whether they produced similar or discrepant results (

We hypothesized that older age would be associated with increased mean RT and RT variability, for both raw RT variability metrics and mean-adjusted RT variability metrics, consistent with the literature described above. Like

Participants were 12,327 visitors to

Participants’ ages ranged from 10 to 96 years old; the average age was 27.36 (

Participants were asked to press the space bar or touch the screen whenever a red WAIT sign changed to green GO!. Participants completed three practice trials before 30 task trials. The task takes approximately 1.5 min and estimates basic psychomotor response speed with high reliability (split-half reliability based on mean RT: 0.93). Participants had 2000 ms to respond on each trial, with a variable inter-trial interval of 700–1500 ms between trials. The task was designed to capture basic psychomotor speed. For each participant, we calculated mean RT, median RT, standard deviation RT, intraindividual coefficient of variability (ICV; standard deviation in RT/mean RT), as well as mean residualized standard deviation in RT (residualized SD RT).

For both tasks, we calculated mean RT and median RT. We also calculated standard deviation RT, coefficient of variability (ICV; standard deviation in RT/mean RT), and mean residualized SD RT (

Participants were asked to indicate the direction of an arrow that is a different color from the rest, see

For the choice RT, we also calculated accuracy (proportion correct) and the inverse efficiency score (IES). IES provides an accuracy-adjusted measure of response speed to account for speed accuracy trade-offs (mean RT/accuracy) (

Descriptive statistics and sex differences for all performance indicators.

Mean RT | 318 (64) | 0.32 [0.28, 0.36] | 919 (250) | 0.24 [0.2,0.28] |

Median RT | 301 (60) | 0.29 [0.25, 0.33] | 869 (222) | 0.24 [0.2,0.28] |

Standard deviation RT | 104 (61) | 0.1 [0.06, 0.14] | 246 (162) | 0.19 [0.15,0.23] |

Coefficient of variability (ICV) | 0.33 (0.17) | − |
0.25 (0.11) | 0.14 [0.10, 0.18] |

Residualized SD RT | 0 (57) | − |
0 (95) | |

Proportion correct | NA | NA | 0.95 (0.07) | 0.12 [0.09, 0.17] |

Inverse efficiency score (IES) | NA | NA | 979 (305) | 0.17 [0.13, 0.21] |

Residualized median RT | NA | NA | 0 (201) | 0.13 [0.91, 0.17] |

Simple reaction time performance.

Choice reaction time performance.

We first excluded participants based on data quality. Choice RT trials with very short response times (RT < 500 ms) were trimmed, based on the finding that accuracy falls to chance for trials <500 ms (see

Data were analyzed in

To evaluate cross-sectional changes in performance across the lifespan, we performed segmented regression analyses with age as our independent variable and (1) median RT, (2) SD RT, (3) ICV, and (4) residualized SD RT as dependent variables. We chose median RT to minimize the effects of outliers. Results of segmented regression are given in

Results of segmented regression.

Median RT | Two segment | 14.1 [13.4, 14.9] | NA | −5.7 [−8.5, −2.8] | 1 [0.94, 1.1] | NA |

Standard deviation RT | Two segment | 17.8 [16.7, 18.9] | NA | −4.2 [−5.5, −3] | 0.44 [0.33, 0.54] | NA |

Coefficient of variability (ICV) | Two segment | 20.1 [19.2, 21.2] | NA | −0.0091 [−0.011, −0.007] | 0.0039 [0.00005, 0.00072] | NA |

Residualized SD RT | Two segment | 20.7 [19.2, 22.3] | NA | −2.7 [−3.54, −2] | 0.11 [−0.004, 0.22] | NA |

Median RT | Two segment | 17.2 [16.6, 17.8] | NA | −21.1 [−9.7, −25.3] | 5.2 [4.8, 5.5] | |

Standard deviation RT | Three segment | 16.5 [15.9, 17.1] | 68 [66.7, 69.4] | −17.3 [−21.3, −13.2] | 1.2 [0.93, 1.5] | |

Coefficient of variability (ICV) | Three segment | 16.5 [15.3, 17.7] | 35.7 [28.3, 43.1] | −0.008 [−0.01, −0.006] | −0.0008 [−0.0014, −0.0002] | 0.0006 [0.00017, 0.0011] |

Residualized SD RT | One segment | NA | NA | −0.0008 [−0.00094, −0.00069] | NA | NA |

Proportion correct | Two segment | 26.2 [24, 28.4] | NA | 0.0026 [0.0022, 0.0031] | 0.00025 [0.00006, 0.00043] | NA |

Inverse efficiency score (IES) | Three segment | 15.8 [14.9, 16.6] | 28.8 [25.1, 32.5] | −33 [−42.2, −23.8] | 0.28 [−1.7, 2.3] | 6.2 [5.5, 7] |

Residualized median RT | Three segment | 16.2 [15.5, 17] | 36.3 [31.4, 41.2] | −20.2 [−25.2, −15.2] | 1.35 [0.4, 2.3] | 5.3 [4.4, 6.1] |

We first examined the relationship between age and median RT, with the goal of replicating previous findings of improvements in speed-related aspects of performance in adolescence, followed by declines through most or all of adulthood. Segmented regression analyses demonstrated that the relationship between age and median RT was best fit by a two segment (one breakpoint) linear function with a breakpoint at 14 years (age of best performance). Median RT decreased before 14 years and increased thereafter. These findings are consistent with the notion that processing speed declines with age (e.g.,

Age-related differences in measures of performance variability for simple RT were largely convergent. Segmented regression analyses for SD RT, ICV, and residualized SD RT all were best fit by a two segment model, with reductions in variability from ages 10 to ages 18–21 years, and performance variability increasing thereafter for the remainder of the lifespan.

To evaluate cross-sectional changes in performance across the lifespan, we again performed segmented regression analyses with age as our independent variable and (1) median RT, (2) SD RT, (3) ICV, (4) residualized SD RT, (5) proportion correct, (6) IES, and (7) residualized median RT as dependent variables. We again chose median RT to minimize the effects of outliers. Results of segmented regression for choice RT data are also given in

For median RT, segmented regression analyses again demonstrated that the relationship between age and median RT was best fit by a two segment (one breakpoint) linear function with a breakpoint at 17 years (age of best performance). Median RT decreased before 17 years and increased thereafter. Proportion correct was also best fit by a two segment (one breakpoint) model, with a breakpoint at 26 years, and improvements in performance across the entire lifespan.

The opposite and opposing relationship between response speed and accuracy indicates that speed–accuracy trade-offs play a large role in lifespan-related differences in performance and, individually, may not appropriately capture cognitive control abilities. To account for speed–accuracy tradeoffs, we looked at the relationship between age and IES (accuracy corrected RT: mean RT/proportion correct). When looking at IES, a three segment model provided best fit, with breakpoints at 16 and 29 years. IES improved from ages 10 to 16 years, increased slightly from ages 16 to 29 years, and then increased more sharply thereafter.

Median RT for choice RT controlling for median RT on simple reaction gave results that were more similar to IES than to median RT. A two breakpoint model best fit this data, with breakpoints at 16 and 36, with improvements in residualized median RT from ages 10 to 16 years, increases from 16 to 36 years, and then sharper increases from 36 to 70 years.

Age-related differences in measures of performance variability for choice RT did not converge across measures. Both SD RT and ICV across the lifespan were best fit by a three segment function, with a first breakpoint at age 16 years. For SD RT, performance variability decreased from 10 to 16, and then increased thereafter. Segmented regression identified a second breakpoint for SD RT at age 68 years, where variability increased very steeply after age 68 years. ICV, on the other hand, decreased from 10 to 16, and then continued to decrease until a second breakpoint at age 36 years, before increasing over the remainder of the lifespan. Results for residualized SD RT were markedly different, however, with a

Men and women differed significantly across most measures of performance and performance variability, although with small effect sizes (

This is the largest study to compare and contrast different indices of performance and performance variability in simple RT and choice RT tasks. Different measures of performance capture different characteristics of human behavior, and here we found that different measures, including measures that putatively measure the same constructs, exhibited different patterns of age-related performance. For example, we found that the peak age for cognitive performance based on median RT was 17, whereas based on coefficient of variability scores for choice RT, we found highest performance at age 36 years. Interestingly, different measures of RT variability also showed distinct patterns across age. While measures of raw variability (SD RT) and coefficient of variability (SD RT/mean RT) showed similar differences with age, for choice RT mean residualized variability (SD RT controlling for mean RT) showed a distinct relationship with age – with apparent decreases in variability across the lifespan, contrary to the literature. This contrasted with simple RT where mean-residualized variability produced similar (but attenuated) age effects when compared with other variability measures.

Our large sample size allowed us to finely characterize variations in RT and RT variability across the lifespan. Such results can provide normative models for cognitive performance on such tasks across the lifespan, which could provide a basis for revealing abnormal trajectories (

Our findings of different ages of peak performance based on mean RT and RT variability suggest that different measures of variability might yield different information about lifespan-related processes. What is most striking about our data here was how remarkably consistent the linear changes in residualized variability were across the lifespan for choice RT, with an almost entirely linear decrease across age that could not be explained by differences in accuracy. We are not certain why the unusual residualized SD RT result appears only in the choice RT task but not simple RT; however, we suspect that it is due to the fact that responses in the choice RT reflect a more complex cognitive process than simple RT that contributes to increases in the mean with age but not SD. Although mean residualized standard deviations have become the primary method of quantifying RT variability, it may be that adjusting for the mean can sometimes obscure true differences in variability. If mean RTs increase with older age due to multiple additive processes, not all of which are associated with changes in variability of RTs, then removing variance associated with the mean could result in an overcorrection that might explain the results observed here. For example, a tendency to respond less impulsively with age would tend to offset greater variability in RTs in a way that is due to shifts in strategy that increase response time without concomitant increases in RT variability. The fact that our finding contrasts sharply with the literature may be due to task differences, file drawer effects (due to violation of

We found a small effect for differences in variability (based on raw variability and coefficient of variability) between men and women, with men demonstrating slightly more consistent performance. We cannot assume that the differences are due to sex differences in processing speed alone [men are typically faster than women, see

There are several limitations of the current study. First, our cross-sectional design unfortunately prohibits us from making any strong conclusions related to individual lifespan trajectories as findings could be due to cohort effects or ascertainment biases that vary by age. Interpretation of correlations (slopes in segmented regression) is problematic with such unequal age bins, and thus, we emphasize patterns of change in terms of changes in slope (breakpoints) rather than correlations as point estimates. Second, given that our sample was self-selected, it is possible that our results are biased toward higher functioning older adults with more expertise using computers. The expectation is that this would cause us to

Despite these limitations, our study provides a potential foundation for future research on lifespan performance and performance variability, how best to conceptualize variability, as well as a richer characterization of how performance metrics differ even in relatively simple task designs. Findings from web-based samples such as this one have been shown to match traditional findings from the literature, and are being used more often to recruit larger, more diverse samples (i.e.,

The primary results reported in this manuscript have been presented at local and national conferences between 2018 and 2019. The data are deposited to the Open Science Framework, which can be accessed with the following link:

All datasets generated for this study are included in the article/

The studies involving human participants were reviewed and approved by Harvard University.

LR and LG contributed to the conception and design of the study, organized the database, and performed the statistical analysis. LR wrote the first draft of the manuscript. All authors contributed to the manuscript revision, and read and approved the submitted version of the manuscript.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

The Supplementary Material for this article can be found online at: