Go/No-Go Test

What the three tasks ask

Each task sets up a response you are about to make and then asks you to hold it back. How often does the press slip out, and how quickly can you stop it?

• Go/No-Go. Four consonants come one at a time, each equally often. Press for three of them and do nothing for the fourth; the rule stays under the letter. Each of the 4 blocks has 50 trials and four new letters. A letter stays for 1,250 ms and 750 ms pass before the next, as in the lab study below. That study ran 600 trials; 200 are this page's choice, so a run takes about 7 minutes.
• SART (Sustained Attention to Response Task). 225 digits, 25 of each from 1 to 9, in random order and never the same digit twice in a row. Each shows for 250 ms, then a ring with a cross masks it for 900 ms, so a new digit comes every 1,150 ms: 4.3 minutes in all. Press for every digit except 3, with speed and accuracy equally important. Five digit sizes are used, in the proportions of the original task (12 to 29 mm there), scaled to your display.
• Stop-signal. A cross for 250 ms, then an arrow: answer its direction as fast as you can, within 1,250 ms. On 25% of trials the arrow turns into XX after a delay, the stop-signal delay (SSD), and you try not to press. The next trial starts 2 s after your answer. The 4 blocks of 52 trials (13 stops each, 52 in all) take about 10 minutes.

The first run of each task starts with practice, which is not counted: 12 letters, 18 digits with 2 targets, or 16 arrows followed by 16 trials with stops.

Things to try

• After a SART run, look at the chart: do the marked errors come where your times fall?
• Run the stop-signal task once as instructed and once waiting for the XX. What happens to go RT and to the checks?
• Do one Go/No-Go run for speed and one for accuracy, then compare commission errors with go RT.
• Watch the delay in the stop-signal chart: it climbs while you stop and falls when you fail.
• Compare keyboard and touch: a longer device delay adds the same time to go RT and to SSRT. Do both rise by about as much?

How the results are calculated

Go/No-Go and SART. A commission error is a press on a no-go trial, an omission error is no press on a go trial. The interval of the commission rate is a 95% Wilson interval, and the one of the median go reaction time (RT) is a percentile bootstrap: 2000 new sets drawn from your trials, keeping the middle 95%. Presses under 100 ms count as presses, but their times are left out of go RT. Go/No-Go also gives d′, the distance between your go and no-go responses in standard units, with the log-linear correction so that 0% or 100% stays finite.

Speed before an error (SART). For each digit 3, the page averages your last four go times. It then compares targets you pressed on with targets you held back. Robertson and colleagues found that errors were predicted by a shortening of RT over the preceding responses.

SSRT. The stopping itself cannot be observed, because a successful stop leaves no time to measure. The race model treats go and stop as two runners: the response slips out when go finishes first. The tracking rule moves the SSD up 50 ms after a successful stop and down 50 ms after a failed one, also after a press before the signal or in the wrong direction, so you fail on about half of the stops. The integration method then reads the stopping time from your go times:

Here RT_n is the nth fastest of all go trials, N_go is their number, and p(respond|signal) is the share of stop trials with a press. With 200 go trials and p = 0.45, it is the 90th fastest. As the consensus paper asks, the go times include wrong-direction and premature presses, and each missed go trial counts as the longest possible time (1,250 ms here). n is rounded to the nearest whole number. For the interval the page repeats the procedure 2000 times: new go times drawn from yours, the same tracking rule and your SSRT as the stopping time, with a variation of 30 ms that the page assumes. It reads the 95% interval from how far those estimates fall from your SSRT.

When no SSRT is shown

The page withholds the number and says why in three cases. The consensus paper sets the first two and leaves the limit of the third to each study.

• Failed stops as slow as go trials. Mean RT on failed stops is at or above mean go RT. The race model predicts the opposite, so its answer would mean little.
• p(respond|signal) below 0.25 or above 0.75. The estimate is most reliable near 0.50.
• More than 20% of go trials missed. The limit is this page's choice: the simulations behind the paper went up to 20%.

Compared with a laboratory study

Hedge, Powell and Sumner (2018) tested students twice, 3 weeks apart, about 100 in total in two studies, with 600 trials per task (75% go). Each value is a group mean with the standard deviation between people in brackets (Supplementary Table B1).

Their stop signal was a tone and their task told a square from a circle, so a run here with arrows and a visual XX is not the same measurement. For the SART the page gives no group value: the only figure it quotes is how well the score repeats.

Why there is no percentile

In the same study people kept their order only moderately from one session to the next. The intraclass correlation (ICC; 1 means the same order and values both times, 0 means no relation) was 0.76 for Go/No-Go commission errors in both studies, but only 0.36 and 0.49 for SSRT by the integration method. In the original SART study the error score repeated at r = 0.76 (25 people, one week apart). Even the better figures come from small groups on other procedures, so a rank among other people would claim more than a run can show. History compares you with your own runs.

Limits

• The SART also measures your chosen speed. Press faster and errors rise, so a low error count can mean caution rather than attention.
• SSRT is an indirect estimate. It rests on the race model and moves a lot from one session to the next. 52 stop trials meet the consensus minimum of 50, which was set for group comparisons; the interval covers only chance among the trials of this run, under this page's model. It does not cover changes between sessions: the standard error of measurement of SSRT was 39 and 35 ms in the lab study.
• A visual stop signal. The page uses XX instead of a tone, because the delay of a sound on a web page is not known. Times with a visual signal may differ from those with a tone.
• Your device. The screen, keyboard or touch screen add a delay. A constant delay raises go RT and SSRT by the same amount: it is added to every recorded time, while the SSD follows only whether you stopped. So compare runs only from the same device. Frame timing limits the SSD to whole frames: a 50 ms step is 3 frames at 60 Hz and 2 frames (61 ms) at 32.6 Hz.
• Short runs and tiredness. Few no-go trials give a wide interval, and the SART has only 25 targets. Errors can also rise over a run as you tire.
• Not a medical test. This is not a test for ADHD, impulsivity or any other condition.

Controls

• Keyboard: Space for Go/No-Go and SART; F or ← for a left arrow, J or → for a right arrow. These keys work on any layout. Enter or Space on the focused button goes on; Esc stops a run.
• Touch and mouse: the buttons under the display; in Go/No-Go and SART also the display itself.
• Pauses: for 0.7 s after a result, the end of a block or Stop the buttons are dimmed and ignore keys still pressed from the task. A second click within 0.7 s of a start does not stop the run, so a double click is safe. Trials themselves have no such pause.
• A hidden tab: the trial in progress is discarded and repeated later in the block, and the run waits for Resume.
• Language: the link at the end of History switches the app to Slovak.

Sources

• Verbruggen, F., et al. (2019). A consensus guide to capturing the ability to inhibit actions and impulsive behaviors in the stop-signal task. eLife, 8, e46323. Full text (PMC6533084)
• Robertson, I. H., Manly, T., Andrade, J., Baddeley, B. T., & Yiend, J. (1997). 'Oops!': performance correlates of everyday attentional failures in traumatic brain injured and normal subjects. Neuropsychologia, 35(6), 747–758. doi:10.1016/S0028-3932(97)00015-8
• Hedge, C., Powell, G., & Sumner, P. (2018). The reliability paradox: why robust cognitive tasks do not produce reliable individual differences. Behavior Research Methods, 50, 1166–1186. Full text and supplementary tables (PMC5990556)