Notes and comment
How to measure post-error slowing: A confound and a simple solution

https://doi.org/10.1016/j.jmp.2012.04.001Get rights and content

Abstract

In many response time tasks, people slow down after they make an error. This phenomenon of post-error slowing (PES) is thought to reflect an increase in response caution, that is, a heightening of response thresholds in order to increase the probability of a correct response at the expense of response speed. In many empirical studies, PES is quantified as the difference in response time (RT) between post-error trials and post-correct trials. Here we demonstrate that this standard measurement method is prone to contamination by global fluctuations in performance over the course of an experiment. Diffusion model simulations show how global fluctuations in performance can cause either spurious detection of PES or masking of PES. Both confounds are highly undesirable and can be eliminated by a simple solution: quantify PES as the difference in RT between post-error trials and the associated pre-error trials. Experimental data are used as an empirical illustration.

Highlights

► We show that the traditional measure of post-error slowing can be confounded. ► We illustrate this confound with both simulations and real data. ► We offer a solution that is both simple and adequate.

Introduction

People tend to slow down after they commit an error, a phenomenon known as post-error slowing (PES). Ever since the classic article “What does a man do after he makes an error?” (Rabbitt, 1966), the PES phenomenon has received considerable attention in the response time (RT) literature and several explanations have been proposed to explain its existence (e.g, Laming, 1968, Laming, 1979, Notebaert et al., 2009, Rabbitt & Rodgers, 1977, see Dutilh, Vandekerckhove, Forstmann, & Wagenmakers, 2012), for an empirical comparison). The most popular account of PES states that it reflects an error-induced increase in response caution that allows a participant to maintain a relatively constant level of accuracy (e.g., Botvinick, Braver, Barch, Carter, & Cohen, 2001, Smith & Brewer, 1995).

Specifically, this account holds that participants continually monitor their performance and interpret errors as a sign that the chosen response threshold was too liberal. Consequently, participants heighten their threshold following an error in order to increase the probability of a correct response on the next trial. The heightened threshold leads to fewer errors but also causes slower responding (i.e., the PES phenomenon).

At the same time, participants interpret correct responses as a sign that the chosen response threshold was too conservative, and therefore they are assumed to lower their threshold following each correct response. Thus, participants become more cautious after an error and slightly more daring after a correct response; in this way the system self-regulates to a state of homeostasis characterized by fast responses and few errors. Fig. 1, based on fictitious but representative data, illustrates the typical pattern of modest post-correct speed-up and pronounced post-error slowing (e.g., Brewer and Smith, 1989, Smith and Brewer, 1995).

This response-monitoring interpretation of PES suggests that the amplitude of PES can be used as a direct measure of cognitive control.1 Although the response monitoring/cognitive control interpretation might not be appropriate in all cases (e.g., Dutilh, Forstmann, Vandekerckhove, & Wagenmakers, submitted for publication, Notebaert et al., 2009), in many studies it is assumed to be correct from the outset. Consequently, the magnitude of PES is often treated as an important dependent variable that is correlated with neurophysiological variables such as anterior cingulate activity (Danielmeier et al., 2011, Li et al., 2006), error-related negativity (ERN) and positivity (Pe; Hajcak, McDonald, & Simons, 2003b), and cortisol levels (e.g., Tops & Boksem, 2010).

In this article we discuss how PES can best be measured. First we explain how, although straightforward and intuitive, the traditional method to quantify PES can create spurious PES or mask real PES as a result of global changes in performance. We illustrate this confound with two simulation studies and then show how the confound can be eliminated. The final section illustrates both spurious and masked PES in a real data set.

Section snippets

The measurement of post-error slowing

There are several methods to quantify PES. The most insightful method plots the fluctuations in mean RT surrounding an error (e.g., Brewer and Smith, 1989, Smith and Brewer, 1995; see Fig. 1 for an example). The resulting graph shows mean RT for error trial E, mean RT for subsequent trials E+1,E+2, etc., and mean RT for preceding trials E1,E2, etc. The form of the graph depends slightly on what trials are included in the calculations. For example, one may choose to include pre-error trials

A confound

The traditional method of quantifying PES, PEŜtraditional=MRTpost-errorMRTpost-correct, has strong face validity. However, the method is vulnerable to a confound that was already hinted at by Laming (1979, p. 205) when he suggested …

…the possibility that errors and the increased RT on trials which follow them are jointly due to a local deterioration in performance. Suppose, for example, that the subject suffers short periods of relative inattention to the CR [choice response] task …During

Simulation studies

In two simulation studies we used the diffusion model (Ratcliff, 1978, Ratcliff and McKoon, 2008, Wagenmakers, 2009) to make the scenarios described above more concrete. The diffusion model produces both RTs and percentage correct. Most importantly for this study, the model can describe the specific influences of motivation and response caution on response time data.

A simple solution

The above confounds arise because post-correct and post-error trials (i.e., the trials used to calculate PEŜtraditional) are not evenly distributed across the time series. The confounds can be eliminated when we compare post-error trials to post-correct trials that originate from the same locations in the time series. One natural option is to use post-correct trials that are pre-error trials at the same time. So, instead of comparing the mean RTs of all post-error trials to those of all

Towards a principled solution

The measure PEŜrobust is designed to be immune against global performance fluctuations that may adversely affect the widely used measure PEŜtraditional. However, two important problems remain. First, PES measures capture error-induced changes in a specific variable, namely mean RT; the measures ignore changes in accuracy and changes in RT distributions. Such changes can be captured by the diffusion model described above. The diffusion model, however, currently does not have a mechanism for

Empirical Illustration: the confound is real

The simulations above showed that PEŜtraditional can detect spurious PES and mask real PES. We now provide an empirical illustration of these two situations. For this illustration, we selected data from two individual participants from a larger study that will be published elsewhere.

Concluding comments

Over the last two decades, cognitive control has become a major research topic in experimental psychology. PES is assumed to be an indicator of cognitive control and as such it is often used as an important behavioral variable. Recently, however, some studies have questioned whether PES really reflects cognitive control (e.g., Dutilh et al., 2012, Notebaert et al., 2009). In this study, however, we focused on a more elementary issue regarding the application of PES as a dependent variable: the

References (47)

  • C.N. White et al.

    Using diffusion models to understand clinical disorders

    Journal of Mathematical Psychology

    (2010)
  • C. Alain et al.

    Neurophysiological evidence of error-monitoring deficits in patients with schizophrenia

    Cerebral Cortex

    (2002)
  • M.M. Botvinick et al.

    Conflict monitoring and cognitive control

    Psychological Review

    (2001)
  • N. Brewer et al.

    Developmental changes in processing speed: influence of speed-accuracy regulation

    Journal of Experimental Psychology

    (1989)
  • K.H. Britten et al.

    The analysis of visual motion: a comparison of neuronal and psychophysical performance

    Journal of Neuroscience

    (1992)
  • C. Danielmeier et al.

    Posterior medial frontal cortex activity predicts post-error adaptations in task-related visual and motor areas

    The Journal of Neuroscience

    (2011)
  • E.R.A. De Bruijn et al.

    Drug-induced stimulation and suppression of action monitoring in healthy volunteers

    Psychopharmacology

    (2004)
  • Dutilh, G., Forstmann, B.U., Vandekerckhove, J., & Wagenmakers, E.J. (2012). A diffusion model account of age...
  • G. Dutilh et al.

    Task-related vs. stimulus-specific practice: a diffusion model account

    Journal of Experimental Psychology

    (2011)
  • G. Dutilh et al.

    Testing theories of post-error slowing

    Attention, Perception, & Psychophysics

    (2012)
  • G. Dutilh et al.

    A diffusion model decomposition of the practice effect

    Psychonomic Bulletin & Review

    (2009)
  • L.K. Fellows et al.

    Is anterior cingulate cortex necessary for cognitive control?

    Brain

    (2005)
  • W.J. Gehring et al.

    Prefrontal-cingulate interactions in action monitoring

    Nature Neuroscience

    (2000)
  • Cited by (179)

    View all citing articles on Scopus
    View full text