Limits...
Validation of a Bayesian Adaptive Estimation Technique in the Stop-Signal Task

View Article: PubMed Central - PubMed

ABSTRACT

The Stop Signal Task (SST), a commonly used measure of response inhibition, uses standard psychophysical methods to gain an estimate of the time needed to withhold a prepotent response. Under some circumstances, conventional forms of the SST are impractical to use because of the large number of trials necessary to gain a reliable estimate of the speed of inhibition. Here we applied to the SST an adaptive method for estimating psychometric parameters that can find reliable threshold estimates over a relatively small number of trials. The Ψ adaptive staircase, which uses a Bayesian algorithm to find the most likely parameters of a psychophysical function, was used to estimate the critical stop signal delay at which the probability of successful response inhibition equals 0.5. Using computational modeling and adult participants, estimates of stop signal reaction time (SSRT) based on the Ψ staircase were compared to estimates using the method of constant stimuli and a standard staircase method of adjustment. Results demonstrate that a reliable estimate of SSRT can be gained very quickly (20–30 stop trials), making the method very useful for testing populations that cannot maintain concentration for long periods or for rapidly obtaining multiple SSRT estimates from healthy adult participants.

No MeSH data available.


Results from horse-race model simulations with fixed Go RT distribution and varied stop RT.Stop RT was systematically varied around the starting point of the stepwise-adjusted staircase and across the parameter space considered by the Bayesian Ψ staircase. (A) Mean absolute deviation from the control SSRT as a function of staircase stop trial. (B) Mean correlation between SSRT estimates as a function of staircase stop trials. (C) Slope of the best fitting affine function describing the relationship between actual SSRT and estimated SSRT (a slope of less than 1 suggests overestimation of fast SSRT and underestimation of slow SSRT). (D) Mean SSRT estimate after 30 stop trials plotted against control SSRT.
© Copyright Policy
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC5120782&req=5

pone.0165525.g001: Results from horse-race model simulations with fixed Go RT distribution and varied stop RT.Stop RT was systematically varied around the starting point of the stepwise-adjusted staircase and across the parameter space considered by the Bayesian Ψ staircase. (A) Mean absolute deviation from the control SSRT as a function of staircase stop trial. (B) Mean correlation between SSRT estimates as a function of staircase stop trials. (C) Slope of the best fitting affine function describing the relationship between actual SSRT and estimated SSRT (a slope of less than 1 suggests overestimation of fast SSRT and underestimation of slow SSRT). (D) Mean SSRT estimate after 30 stop trials plotted against control SSRT.

Mentions: We were interested in two aspects of the results from these simulations. First, how quickly and reliably the adaptive staircase method settled on an accurate estimate of SSRT relative to a typical stepwise-adjusted staircase, second whether the estimates were accurate or showed evidence of systematic bias across a range of SSRTs. To examine how quickly the staircases could find an accurate SSRT estimate, we calculated a control estimate of SSRT using the whole population of Go RTs in the distribution, then calculated the mean absolute difference between this control SSRT and the estimate after each trial of the staircase. For this first set of simulations, the stepwise-adjusted staircase always began with SSD of 300 ms, while the Ψ staircase considered a range of values for α that corresponded to an SSD between 50 and 550 ms. Fig 1A shows this mean absolute deviation from the control SSRT as a function of staircase stop trial, averaged across the simulations and across the different stop RTs. As is evident, both the method of stepwise adjustment and the Ψ method converge quickly on an accurate SSRT estimate. Both methods reached a mean deviation of less than 20 ms in less than 20 trials, and within 30 trials the mean deviation was less than 15 ms for both methods. Between 30 and 100 trials, the mean deviation improved marginally from 12 ms to 10 ms. For each simulated experiment, we calculated the correlation between the SSRT estimate from each staircase with the control SSRT (i.e. the correlation between measures amongst 50 simulated participants with systematically varied stop RT). These correlations were highly consistent across simulations; the means across the simulations are shown in Fig 1B. This shows that after 20 trials the correlation was approximately 0.92 and by 30 trials this had risen to 0.95, with marginal increase beyond 30 trials.


Validation of a Bayesian Adaptive Estimation Technique in the Stop-Signal Task
Results from horse-race model simulations with fixed Go RT distribution and varied stop RT.Stop RT was systematically varied around the starting point of the stepwise-adjusted staircase and across the parameter space considered by the Bayesian Ψ staircase. (A) Mean absolute deviation from the control SSRT as a function of staircase stop trial. (B) Mean correlation between SSRT estimates as a function of staircase stop trials. (C) Slope of the best fitting affine function describing the relationship between actual SSRT and estimated SSRT (a slope of less than 1 suggests overestimation of fast SSRT and underestimation of slow SSRT). (D) Mean SSRT estimate after 30 stop trials plotted against control SSRT.
© Copyright Policy
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC5120782&req=5

pone.0165525.g001: Results from horse-race model simulations with fixed Go RT distribution and varied stop RT.Stop RT was systematically varied around the starting point of the stepwise-adjusted staircase and across the parameter space considered by the Bayesian Ψ staircase. (A) Mean absolute deviation from the control SSRT as a function of staircase stop trial. (B) Mean correlation between SSRT estimates as a function of staircase stop trials. (C) Slope of the best fitting affine function describing the relationship between actual SSRT and estimated SSRT (a slope of less than 1 suggests overestimation of fast SSRT and underestimation of slow SSRT). (D) Mean SSRT estimate after 30 stop trials plotted against control SSRT.
Mentions: We were interested in two aspects of the results from these simulations. First, how quickly and reliably the adaptive staircase method settled on an accurate estimate of SSRT relative to a typical stepwise-adjusted staircase, second whether the estimates were accurate or showed evidence of systematic bias across a range of SSRTs. To examine how quickly the staircases could find an accurate SSRT estimate, we calculated a control estimate of SSRT using the whole population of Go RTs in the distribution, then calculated the mean absolute difference between this control SSRT and the estimate after each trial of the staircase. For this first set of simulations, the stepwise-adjusted staircase always began with SSD of 300 ms, while the Ψ staircase considered a range of values for α that corresponded to an SSD between 50 and 550 ms. Fig 1A shows this mean absolute deviation from the control SSRT as a function of staircase stop trial, averaged across the simulations and across the different stop RTs. As is evident, both the method of stepwise adjustment and the Ψ method converge quickly on an accurate SSRT estimate. Both methods reached a mean deviation of less than 20 ms in less than 20 trials, and within 30 trials the mean deviation was less than 15 ms for both methods. Between 30 and 100 trials, the mean deviation improved marginally from 12 ms to 10 ms. For each simulated experiment, we calculated the correlation between the SSRT estimate from each staircase with the control SSRT (i.e. the correlation between measures amongst 50 simulated participants with systematically varied stop RT). These correlations were highly consistent across simulations; the means across the simulations are shown in Fig 1B. This shows that after 20 trials the correlation was approximately 0.92 and by 30 trials this had risen to 0.95, with marginal increase beyond 30 trials.

View Article: PubMed Central - PubMed

ABSTRACT

The Stop Signal Task (SST), a commonly used measure of response inhibition, uses standard psychophysical methods to gain an estimate of the time needed to withhold a prepotent response. Under some circumstances, conventional forms of the SST are impractical to use because of the large number of trials necessary to gain a reliable estimate of the speed of inhibition. Here we applied to the SST an adaptive method for estimating psychometric parameters that can find reliable threshold estimates over a relatively small number of trials. The Ψ adaptive staircase, which uses a Bayesian algorithm to find the most likely parameters of a psychophysical function, was used to estimate the critical stop signal delay at which the probability of successful response inhibition equals 0.5. Using computational modeling and adult participants, estimates of stop signal reaction time (SSRT) based on the Ψ staircase were compared to estimates using the method of constant stimuli and a standard staircase method of adjustment. Results demonstrate that a reliable estimate of SSRT can be gained very quickly (20–30 stop trials), making the method very useful for testing populations that cannot maintain concentration for long periods or for rapidly obtaining multiple SSRT estimates from healthy adult participants.

No MeSH data available.