Neural bases of individual variation in decision time

Abstract

People make decisions by evaluating existing evidence against a threshold or level of confidence. Individuals vary widely in response times even when they perform a simple task in the laboratory. We examine the neural bases of this individual variation by combining computational modeling and brain imaging of 64 healthy adults performing a stop signal task. Behavioral performance was modeled by an accumulator model that describes the process of information growth to reach a threshold to respond. In this model, go trial reaction time (goRT) is jointly determined by the information growth rate, threshold, and movement time (MT). In a linear regression of activations in successful go and all stop (Go+Stop) trials against goRT across participants, the insula, supplementary motor area (SMA), pre‐SMA, thalamus including the subthalamic nucleus (STN), and caudate head respond to increasing goRT. Among these areas, the insula, SMA, and thalamus including the STN respond to a slower growth rate, the caudate head responds to an elevated threshold, and the pre‐SMA responds to a longer MT. In the regression of Go+Stop trials against the stop signal reaction time (SSRT), the pre‐SMA shows a negative correlation with SSRT. These results characterize the component processes of decision making and elucidate the neural bases of a critical aspect of inter‐subject variation in human behavior. These findings also suggest that the pre‐SMA may play a broader role in response selection and cognitive control rather than simply response inhibition in the stop signal task. Hum Brain Mapp 35:2531–2542, 2014. © 2013 Wiley Periodicals, Inc.

INTRODUCTION

Humans make decisions to survive a life‐threatening situation or simply to maneuver daily routines. Some decisions are simple to make, such as driving through a green light. On the other hand, if the traffic light turns yellow as one approaches the intersection, one will have to screech to stop or risk rushing through a red light. Without knowing when the light will turn red and weighing the possibility of getting caught against the need to arrive at a destination on time, one takes time to decide. Thus, when information or evidence is less transparent or when there is a higher bar for deciding, decision time is prolonged. Not surprisingly, individuals vary to a great extent in how they accumulate information and set the threshold. However, the neural basis of such inter‐subject variation is not clear.

Previous research has been successful in modeling reaction times in two‐choice tasks in terms of a rate of evidence accumulation—the speed at which information builds up, a threshold—the confidence level for making a decision, and a movement time (MT) [Brown and Heathcote, 2008; Leach and Carpenter, 2001; Ratcliff, 1978; Ratcliff and Van Dongen, 2011; Usher and McClelland, 2001]. In these models, evidence starts to accumulate after the stimulus is encoded; when accrued evidence reaches the threshold, a decision is made. These models were used to examine speed accuracy trade‐off during perceptual decisions [Forstmann et al., 2008; Kayser et al., 2009; Mansfield et al., 2011; Philiastides et al., 2011; Watanabe and Munoz, 2010]. For example, applying a linear ballistic accumulation model to participants' performance on motion discrimination, Forstmann et al. [2008] reported activation of the pre‐supplementary motor area (pre‐SMA) and striatum in association with a lower response threshold, especially when speed was emphasized (see also Mansfield et al., [2011]). Watanabe and Munoz applied a model of linear rise to threshold with ergodic rate to monkeys' performance on saccadic eye movements and showed that microstimulation of the caudate results in longer reaction time by reducing the rise rate during saccade initiation.

Here, we used a modified accumulator (see details in Methods section) to model the reaction times of 64 healthy adults performing a stop signal task, a paradigm widely used to study cognitive control [Logan and Cowan, 1984], during functional magnetic resonance imaging (fMRI). In this task, a go signal instructs participants to quickly respond and an occasional stop signal that follows instructs participants to withhold the response. With the need to respond to the go signal quickly and, at the same time, be ready to stop when the stop signal appears, participants have to decide how fast they should act [Li et al., 2009]. As a result, individuals vary in their go trial reaction times (goRT). This provides an ideal scenario to examine the neural processes underlying individual's variability in threshold setting and the rate of evidence accumulation. Using an accumulator model, we estimated the best parameters representing evidence accumulation and threshold setting. In a regression of successful go and all stop (Go+Stop) trials against goRT and model parameters, we hypothesized that neural activities that are correlated with goRT will be differentially involved in these two dimensions of decision making.

In addition to identifying regional activations associated with the decision process, an auxiliary aim of this study is to further investigate the functions of pre‐SMA in cognitive control. Our previous studies implicated the pre‐SMA in response inhibition in the stop signal task [Chao et al., 2009; Duann et al., 2009; Li et al., 2006; Zhang and Li, 2012b]. However, other studies suggest that the pre‐SMA is involved in a broader role of action selection [Rushworth et al., 2002, 2004]. To distinguish these hypotheses, we estimated the stop signal reaction time (SSRT) using the race model [Logan and Cowan, 1984]—as an index of individual's capacity of response inhibition—for a regression against the Go+Stop trials. A response of pre‐SMA to inter‐subject variation in the SSRT would support a role of this medial prefrontal structure in response inhibition. In contrast, a response to inter‐subject variation in both goRT and SSRT would support a broader role of pre‐SMA in action selection.

METHODS

Participants and Behavioral Task

Sixty‐four healthy adults (34 men; 27.6 ± 5.5 years) participated in the study. Participants met the following criteria for recruitment: age 18 and older; right‐handed and able to read and write English; no current or history of diagnosis of any Axis I psychiatric or substance (except nicotine) use disorders [First et al., 1995]; no current use of psychotropic medications; no significant current medical including neurological, cardiovascular, endocrine, renal, hepatic, or thyroid disorders. Women who were pregnant or breast feeding were not recruited. Participants tested negative for cocaine, amphetamines, opioids, and benzodiazepines prior to fMRI. All participants signed a written consent after they were given a detailed explanation of the study in accordance with a protocol approved by the Yale Human Investigation Committee.

All participants performed a stop signal task (Fig. 1 a, Chao et al., 2009; Li et al., 2006; Logan and Cowan, 1984]. In this task, two types of trials, “GO” and “STOP” were randomly presented with an inter‐trial‐interval (ITI) of 2 s (the time between the end of the previous trial and the start of the current trial). A fixation dot appeared on the screen to signal the beginning of each trial. The dot was replaced by a circle—the go signal—after a time interval, the “fore‐period,” which varied from 1 to 5 s (with uniform distribution, Li et al., 2005]. The randomized fore‐period minimized anticipation and allowed jittering of events of interest. Participants were instructed to press a button on a hand‐held button box using the right index finger when they saw the circle. The circle disappeared at button press or after 1 s if the participant failed to respond.

Figure 1.

The stop signal paradigm and accumulator model. (a) Stop signal paradigm. In GO trials (approximately 75%, Trial N), participants responded to the go signal (a circle) and had to withhold the response when they saw the stop signal (an X) in STOP trials (approximately 25%, Trial N+1). In both trials, the go signal appeared after a randomized time interval between 1 and 5 s (the fore‐period or FP, uniform distribution) following the appearance of the fixation point. The stop signal followed the go signal by a time delay—the SSD. The SSD was updated according to a staircase procedure, increasing or decreasing by 67 ms following a stop success or stop error trial, respectively. The ITI was 2 s. (b) Accumulator model for the stop signal task. GS: go success trial, in green; SS: stop success trial, in cyan; SE: stop error trial, in red; θ: threshold; Enc.Time1: encoding time after go signal onset; μ _G: growth rate after go signal encoding; Enc.Time2: encoding time after stop signal onset; μ _R: reduction rate after stop signal encoding; k: leakage; MT: movement time. (c) Simulation results for one participant, whose model fit reached a chi‐square of 0.1072. Green: GS trials; Red: SE trials; Cyan: average of SS trials; Blue line: threshold θ. GS and SE trials were binned at 50 ms. Dotted lines are the average of the GS or SE trials for each bin; solid lines are the average of all GS or SE trials. (d) Histogram of empirical (top panels) and simulated (bottom panels) reaction time distributions of GS and SE trials of the same participant. The median of empirical and simulated values of goRT are 566 ms and 542 ms; stop error reaction time (SERT): 496 ms and 486 ms; go response rate GS%: 98% and 98%; stop success rate SS%: 50% and 50%, respectively. y‐axis is the proportion of trials. [Color figure can be viewed in the online issue, which is available at http://wileyonlinelibrary.com.]

In approximately one quarter of the trials, the circle was followed by a “cross”—the stop signal. Participants were instructed to withhold button press when they saw the stop signal. The trial terminated at button press or after 1 s if the participant successfully inhibited the response. The time between the go and stop signals, the stop signal delay (SSD), started at 200 ms and varied from one stop trial to the next according to a staircase procedure, increasing and decreasing by 67 ms each after a successful and failed stop trial. With the staircase procedure, we anticipated that participants would succeed in withholding the response half of the time. We computed a critical SSD that each participant would require to succeed in half of the stop trials [Li et al., 2008a]. The SSRT was computed by subtracting the critical SSD from the median goRT for each participant [Logan and Cowan, 1984].

Participants were trained briefly on the task before imaging to ensure that they understood the task. They were instructed to quickly press the button when they saw the go signal while keeping in mind that a stop signal might come up in some trials. In the scanner, they completed four sessions of the task, each lasting 10 minutes, with approximately 100 trials in each session.

Accumulator Model

We used a modified accumulator model [Boucher et al., 2007; Usher and McClelland, 2001] to estimate information accumulation or growth rate, threshold, and MT for each individual participant. The growth rate, with a noise embedded in the accumulating process, represents the speed at which information is gathered; as time goes by without seeing a stop signal, participants are more inclined to make a response. The threshold represents the criterion level of information, on the basis of which people are confident in making a decision. Because the go and stop processes interact prior to the go signal onset [Hu and Li, 2012 ], the threshold can be interpreted as a “go” or “quit‐stopping” threshold, as in an earlier study of visual search [Wolfe and Van Wert, 2010 ]. That is, participants accumulate evidence to ascertain the absence of the stop signal, a process similar to visual search where participants search to confirm the absence of a target. However, while the threshold in Wolf and Van Wert's model was of timing for no response and existed only for target absent trials, our threshold was shared by all trials.

The evidence accumulates after go signal encoding (Enc. Time 1) with a growth rate μ _G [± standard deviation (SD) σ _G]. Once the threshold (θ) is reached, a go response is determined. While θ was fixed for the convenience of simulation in some studies [e.g., Boucher et al., 2007 ], it was estimated in our model. In stop trials, the evidence accumulates as it does in go trials before the stop signal appears. After the stop signal is encoded with an Enc. Time 2, evidence accumulation starts to reduce at a rate μ _R (± σ _R). Thus, μ _R represents the growth rate of noisy evidence against a decision to go. All parameters serve to constrain the model and determine whether the stop signal would prevent the go response: for example, if μ _G is large and Enc. Time 2 is small, the go response would be difficult to stop; moreover, if σ _R is large (i.e., high uncertainty), evidence for a go response may still reach the threshold even after the stop signal has been encoded. The goRT was computed as the sum of the time for the accumulator to reach threshold and the non‐decision time including the Enc. Time 1 and MT. MT is a variable derived from a uniform distribution, as suggested by Ratcliff [ 1978 ], and conventionally interpreted as a non‐decision time (as with the encoding times) and thus not part of the canonical process of decision making. Given that reaction time (RT) = decision time + non‐decision time [Ratcliff, 1978 ], MT is integral to modeling RT. A leakage term (k) was incorporated in the accumulation process to account for information decay as suggested by previous research [Usher and McClelland, 2001]. Therefore, the model contained nine parameters: μ _G, σ _G, μ _R, σ _R, θ, MT, Enc. Time 1, Enc. Time 2, and k. Figure 1 b shows a schematic of the model.

The accumulating process is described for the go trials in Eq. (1) (1) and stop trials in both Eqs. (1) and (2). These equations specify the change in the accumulation with a time step dt and dt/τ set to 0.1 [Boucher et al., 2007]. ξ is a Gaussian noise term with a mean of zero and a variance of σ _G ² and σ _R ², respectively. The accumulating process automatically aborts when it reaches the threshold, resulting in a successful go or SE trial, or at 1,000 ms (the time window for a response), resulting in a non‐responded go or successful stop trial. Equations (1) and (2) differ with respect to the accumulation rate. While both μ_G and μ_R represent how fast evidence changes, μ_G represents the accumulation of internal information and μ_R is influenced additionally by the stop signal. That is, μ_G is required for all trials and μ_R is only present for the stop trials; at the trigger of the stop signal, μ_R drives the accumulator away from the threshold to prevent a go response.

$$\mathrm{d}A\mathrm{(t)=}\frac{\mathrm{d}t}{\tau }\cdot [{\mu }_{\mathrm{G}}-k\cdot (A\left)\right]+\sqrt{\frac{\mathrm{d}t}{\tau }}\cdot {\xi }_G\{{}_{t\hspace{0.17em}<\hspace{0.17em}\text{SSD+Enc.Time 2 in stop trials}}^{t\hspace{0.17em}\le \hspace{0.17em}1,000\hspace{0.17em}\text{ms in go trials}}$$ (1)

$$\mathrm{d}A\mathrm{(t)=}\frac{\mathrm{d}t}{\tau }\cdot [{\mu }_{\mathrm{R}}-k\cdot (A\left)\right]+\sqrt{\frac{\mathrm{d}t}{\tau }}\cdot {\xi }_{R}t\ge \mathrm{SSD+Enc}\mathrm{.Time}\mathrm{2}$$ (2)

We fit the model to the behavioral data by searching for parameters to minimize a Pearson chi‐square statistic [Ratcliff and Tuerlinckx, 2002]. The observed reaction times of go trials and SE trials were separately grouped into six bins (at 0.1, 0.3, 0.5, 0.7, and 0.9 quantiles). The range and step of the parameters for the simulated data were as follows: μ _G from 0.01 to 2.00 in 0.01 steps; μ _R from −2.00 to −0.01 in 0.01 steps; σ _G and σ _R both from 0.01 to 3.00 in 0.01 steps; θ from 1 to 30 in 1 steps; k from 0.01 to 0.30 in 0.01 steps; MT from 100 to 450 ms in 50 ms steps; Enc. Time 1 at 30 ms [Boucher et al., 2007]; Enc. Time 2 from 30 to 100 ms in 10 ms steps. Model parameters were estimated by minimizing chi‐square using the Nelder‐Mead simplex method [Nelder and Mead, 1965]. The simulations were run in MATLAB 7.9 on a near‐supercomputer cluster supported by the Information Technology Service of the Yale University. The chi‐square statistic was obtained as the sum of (O–E)²/E across the six bins for each reaction time distribution where O and E was the frequency of observed and simulated data, respectively. We performed the chi‐square goodness of fit each for goRT and SERT distribution. For each participant, the summed chi‐square value from the two distributions served as an index of overall model fit, with a smaller value indicating a better model fit.

There are several reasons why we used a modified accumulator rather than one of the published models. First, the leaky competing accumulator [LCA, Usher and McClelland, 2001], linear ballistic accumulator [LBA, Brown and Heathcote, 2008], and drift diffusion [Ratcliff, 1978] models were built originally for decision tasks with multiple choices and corresponding boundaries while our model was built with one boundary involving a single measurable, overt response. Second, previous models assumed a competition in the decision process, whether explicit (LCA and LBA) or implicit (drift diffusion), with different levels of inhibition between the competitors. For example, the LCA, as well as the interactive race model [Boucher et al., 2007], has a mutual inhibition parameter controlling the level of inhibition. The drift diffusion model implements the competition by having an upper and lower boundary with the two processes being reciprocal; that is, drifting toward one threshold is tantamount to drifting away from the other threshold. In our model of the SST, the GO and STOP processes are not two independent processes competing to reach the threshold. As discussed earlier, the GO and STOP processes are interactive [Boucher et al., 2007], and such interaction is in place even before the go signal appears [Hu and Li, 2012; Lo and Wang, 2010]. Third, the data structure in our SST is different from those in choice decision tasks. Previous models required distinct RT distributions for both correct and incorrect responses. In our SST, there is a RT distribution from correct go and incorrect stop trials, but not correct stop trials.

Imaging Protocol and Spatial Preprocessing of Brain Images

Conventional T ₁‐weighted spin‐echo sagittal anatomical images were acquired for slice localization using a 3Tesla scanner (Siemens Trio). Anatomical images of the functional slice locations were next obtained with spin‐echo imaging in the axial plan parallel to the Anterior Commissure‐Posterior Commissure (AC‐PC) line with repetition time (TR) = 300 ms, Echo time (TE) = 2.5 ms, bandwidth = 300 Hz/pixel, flip angle = 60°, field of view = 220 × 220 mm, matrix = 256 × 256, 32 slices with slice thickness = 4 mm and no gap. A single high‐resolution T ₁‐weighted gradient‐echo scan was applied on each participant. One hundred and seventy‐six slices parallel to the AC‐PC line covering the whole brain were acquired with TR = 2,530 ms, TE = 3.66 ms, bandwidth = 181 Hz/pixel, flip angle = 7°, field of view = 256 × 256 mm, matrix = 256 × 256, 1 mm³ isotropic voxels. Functional blood oxygenation level dependent signals were then acquired with a single‐shot gradient‐echo echo‐planar imaging (EPI) sequence. Thirty‐two axial slices parallel to the AC‐PC line covering the whole brain were acquired with TR = 2,000 ms, TE = 25 ms, bandwidth = 2,004 Hz/pixel, flip angle = 85°, field of view = 220 × 220 mm, matrix = 64 × 64, 32 slices with slice thickness = 4 mm and no gap. Three hundred images were acquired in each run for a total of four runs.

Data were analyzed with Statistical Parametric Mapping (SPM8, Wellcome Department of Imaging Neuroscience, University College London, UK). Images from the first five TRs at the beginning of each run were discarded to enable the signal to achieve steady‐state equilibrium between radio frequency pulsing and relaxation. Images of each participant were realigned (motion‐corrected) and corrected for slice timing. A mean functional image volume was constructed for each participant for each run from the realigned image volumes. These mean images were co‐registered with the high resolution structural image and the structural image was then segmented for normalization to an MNI (Montreal Neurological Institute) EPI template with affine registration followed by nonlinear transformation [Ashburner and Friston,1999; Friston et al., 1995a]. Finally, images were smoothed with a Gaussian kernel of 8 mm at full width at half maximum.

General Linear Model

There were four possible trial outcomes: go success (GS), go error, stop success (SS), and SE. A statistical analytical design was constructed for each participant using the general linear model (GLM) with the onsets of the go signal from every trial convolved with a canonical hemodynamic response function (HRF) and the temporal derivative of the canonical HRF [Friston et al., 1995b]. Realignment parameters in all six dimensions were also entered in the model. Serial autocorrelation of the time series was corrected by a first degree autoregressive or AR(1) model [Della‐Maggiore et al., 2002; Friston et al., 2000]. The data were high‐pass filtered (1/128 Hz cutoff) to remove low‐frequency signal drifts. The GLM estimated the component of variance that could be explained by each of the regressors. For individual participants, we derived go and stop “activations” (.con images) using a contrast of “1” each on GS and stop trials.

Because the goRT, SSRT, and model parameters are all determined by a decision process shared by GS and stop (both SS and SE) trials, we performed regressions of GS and Stop trials combined (Go+Stop) each on goRT, SSRT, and model parameters θ, μ _G, and MT in group level analyses. In an exploratory analysis, we separated the regressions for Go and Stop trials; that is, we regressed GS trials against goRT and Stop trials against SSRT. These analyses were performed on the basis that goRT and SSRT were each specific to go and stop trials. Moreover, although the stop signal task does not come with an overt measurement that could be uniquely explained by μ _R, one might speculate that the neural correlates of μ _R and μ _G are similar because both of them reflect information accumulation. We thus performed a regression of Stop trials against μ _R to explore its neural correlate.

RESULTS

Behavioral Performance and Model Fit

The mean and median goRT were 570 ± 92 (mean ± SD) and 563 ± 98 ms, respectively, consistent with the right‐skewed distribution of RT in an RT task. Participants scored 97 ± 3% of go trials and 50 ± 2% of stop trials. The mean SSRT was 195 ± 33 ms. SSRT and goRT were not correlated across participants (P = 0.2171, Pearson regression). These findings suggested that participants were well engaged in the task and their overall performance was adequately tracked by the staircase procedure.

To investigate the component process of decision making, we fit the performance in the stop signal task with the accumulator model as described in the Methods section. Monte‐Carlo simulation was used to search for best fit parameters that minimized the Pearson chi‐square statistics for each participant. Across participants, the average chi‐square goodness of fit was 0.17 ± 0.19, suggesting an excellent model fit. For each parameter, the growth rate (μ _G) was 0.81 ± 0.48, SD of growth rate (σ _G) was 1.67 ± 0.49, reduction rate (μ _R) was 0.60 ± 0.55, SD of reduction rate (σ _R) was 2.00 ± 0.52, threshold (θ) was 23.79 ± 4.20, MT was 150.27 ± 44.32, leakage (k) was 0.02 ± 0.04, encoding time (Enc. Time) 1 was 30.18 ± 0.44, and Enc. Time 2 was 59.52 ± 16.97. The three primary parameters of interest, μ _G, θ, and MT, were not correlated (all P's > 0.3, pair‐wise linear regressions) across participants. Figure 1c,d show the model and its performance for one participant. The empirical RT profiles of both go and SE trials mirrored those of the model.

fMRI Results

To examine regional brain activations associated with behavioral performance and the component processes of decision making, we regressed the activations of Go+Stop trials against goRT and SSRT, separately, across participants. Activations of the right anterior insula, caudate head, SMA extending to pre‐SMA, and thalamus including the subthalamic nucleus (STN) showed a positive correlation with goRT (Fig. 2a). We mapped the activation in the thalamus and determined the inclusion of the STN as defined in the Talairach Daemon Brodmann atlas in WFU Pickatlas, as well as in Prodoehl et al. [2008]. Activations of the pre‐SMA showed a negative correlation with SSRT (Fig. 2b). As described in the Methods section, evidence accumulation, threshold, and MT jointly determined the goRT. Thus, we examined whether each of these model parameters involved distinct neural processes. In the regressions of Go+Stop trials against model parameters, the right anterior insula and thalamus including the STN were associated with a slower growth rate (Fig. 2c), the pre‐SMA was associated with a longer MT (Fig. 2d), and the right caudate was associated with a higher threshold (Fig. 2e). All results were significant at P < 0.05, corrected for family‐wise error (FWE) of multiple comparisons (Table 1).

Figure 2.

Regional activations in the regression of successful go and all stop (Go+Stop) trials with goRT, SSRT, and model parameters. (a) Regression of Go+Stop trials with goRT showed positive correlations in the right insula, right caudate head, SMA, and thalamus including the STN. (b) Regression of Go+Stop trials with SSRT showed negative correlation in the pre‐SMA. (c) Regression of Go+Stop trials with growth rate (μ _G) showed positive correlation in the cuneus and negative correlations in the thalamus including the STN, right insula, and SMA. (d) Regression of Go+Stop trials with MT showed positive correlation in the pre‐SMA. (e) Regression of Go+Stop trials with threshold (θ) showed positive correlation in the right caudate head. Z: MNI z‐coordinate. Red: positive correlation; Blue: negative correlation. All results are significant at P < 0.05, corrected for FWE of multiple comparisons. [Color figure can be viewed in the online issue, which is available at http://wileyonlinelibrary.com.]

Table 1.

Regional brain activations during successful go and all stop (Go+Stop) trials regressed against (a) goRT, (b) SSRT, (c) growth rate μ _G, (d) MT, and (e) threshold θ

Regression	Region	Cluster size (# voxels)	FWE P‐Value	Z‐Value	MNI coordinate
					X	Y	Z
(a) goRT
Positive	R Insula	193	0.001	4.60	33	23	10
	L/R SMA/pre‐SMA	357	0.000	4.36	−6	−1	55
	R Caudate	89	0.032	4.28	12	8	7
	L/R Thalamus/STN	153	0.004	4.05	3	−25	−2
Negative	None
(b) SSRT
Positive	None
Negative	R pre‐SMA	75	0.038	3.92	6	29	46
(c) μG
Positive	R Cuneus	71	0.003a	5.13	3	−82	31
Negative	L/R Thalamus/STN	219	0.001	5.12	6	−13	−5
	L SMA/pre‐SMA	198	0.001	4.08	−6	−7	49
	R Insula	85	0.037	4.03	39	17	10
(d) MT
Positive	R pre‐SMA	99	0.023	5.44	12	20	40
Negative	None
(e) θ
Positive	R Caudate	42	0.025a	4.66	21	17	19
Negative	None

^aPeak voxel P‐value; L: left; R: right.

Together, the results showed that activation of the pre‐SMA was positively correlated with goRT and negatively correlated with SSRT. To confirm this, we performed a conjunction analysis of the two regressions and identified a cluster in the pre‐SMA (x = 6, y = 23, y = 43, 77 voxels, Z = 4.20), at cluster P < 0.05 FWE. Furthermore, activations to the regression against goRT and model parameters appeared to overlap. To demonstrate the overlapping areas, we performed regressions against each parameter, masked for small volume correction (SVC) by an region of interest (ROI) of the clusters identified for the regression against goRT using MarsBar (Brett et al., 2002). The results showed activations of the thalamus including the STN, SMA, and right insula in negative correlation with μ _G, and activations of the pre‐SMA and right insula (x = 33, y = 20, z = −5, 178 voxels, Z = 3.66) in positive correlation with MT. The clusters identified here showed coordinates identical to those reported in Table 1, except when noted. All activations survived voxel P < 0.05 FWE (Fig. 3). In particular, a few contrasts revealed clusters that overlapped in the pre‐SMA, including the conjunction analysis of goRT and SSRT, and the SVC on MT and the threshold (sagittal panel in Fig. 3).

Figure 3.

Areas overlapped across regressions of Go+Stop trials with behavioral measurements and model parameters. Cyan: conjunction analysis between the regressions of Go+Stop trials against goRT and SSRT; Yellow: SVC on the regression of Go+Stop trials against MT with an ROI mask from the regression of Go+Stop trials against goRT; Green: SVC on the regression of Go+Stop trials against μ _G with an ROI mask from the regression of Go+Stop trials against goRT. [Color figure can be viewed in the online issue, which is available at http://wileyonlinelibrary.com.]

In an exploratory analysis, the regression of GS trials against goRT showed activations in the right insula, caudate, pre‐SMA, and thalamus including the STN. The regression of Stop trials against SSRT showed activation in the pre‐SMA (Supporting Information Fig. 1). These results were almost identical to those shown in the regressions of Go+Stop trials. The regression of stop trial activations against μ _R also revealed regions that were shown in the regression of Go+Stop trials with μ _G; that is, the thalamus and the pre‐SMA (x = 0, y = 8, y = 52, 147 voxels, Z = 3.96) (Supporting Information Fig. 2).

DISCUSSION

Component Neural Processes of Decision Making in the Stop Signal Task

Our results identify the neural bases of individual variation in threshold setting, information accumulation, and MT during decision making in the stop signal task. The accumulator model describes individuals' performance with parameters that determine the speed of their responses. In combination with the model, brain imaging reveals distinct neural processes mediating the decision to respond, with the right anterior insula, SMA extending to pre‐SMA, and thalamus including the STN increasing activations to a slower growth rate, the right caudate increasing activations to a higher threshold, and the pre‐SMA increasing activations to a longer MT. Therefore, in support of our hypothesis, individuals vary in decision time for different reasons and recruit different neural structures in their decisions to act.

The structures that activate in association with these decision parameters have previously been implicated in decision making. The anterior insula responds to task uncertainty and difficulty [Grinband et al., 2006; Kotani et al., 2009], rule switch and response selection [Paulus et al., 2005], as well as risk prediction [Preuschoff et al., 2008], perhaps by integrating physiological including autonomic signals [Critchley, 2005; Damasio, 1996]. Of direct relevance to the current findings, insular activation has been associated with a prolonged RT in other behavioral tasks. For instance, Binder et al. [2004] reported a positive correlation between bilateral insular activation and RT in auditory syllable identification. In a fear‐disgust, two‐choice discrimination task, activation of the insula was positively correlated with RT on a trial‐by‐trial base [Thielscher and Pessoa, 2007]. Therefore, our current results not only accord with these earlier findings but also specify growth rate as the mediator of this association.

The STN is part of the cortico‐subcortical circuit of cognitive motor control [Bogacz et al., 2010], playing a critical role in response selection and inhibition [Aron and Poldrack, 2006; Frank, 2006; Gurney et al., 2001; Neubert et al., 2010; Volkmann et al., 2010], as well as feedback‐based learning [Brown et al., 2006]. The STN receives input from the prefrontal cortex through a hyperdirect pathway and “buys” time for appropriate actions [Cavanagh et al., 2011]. Activation of the STN was associated with quitting a default action and selecting an alternative action with greater task difficulty [Fleming et al., 2010]. Disruption of STN functions by high frequency deep brain stimulation resulted in impulsive responses in various decision tasks [Cavanagh et al., 2011; Frank et al., 2007]. In rodents, lesions of the STN reduced goRT in the stop signal task [Eagle et al., 2008]. In unit recordings, STN neurons responded to a switch from automatic to controlled actions, inhibiting inappropriate responses [Isoda and Hikosaka, 2008]. Thus, along with the earlier work, the current findings support a role of the STN in inhibitory control and specify information accumulation rate as the mediator of this role in decision making.

The caudate head receives inputs from the prefrontal cortices and controls a multitude of cognitive processes. In imaging studies of the stop signal task, greater activation of caudate head was associated with inhibiting erroneous responses [Vink et al., 2005], potentially reflecting an increased threshold for motor response, in light of the current findings. Computational modeling showed that the caudate head is involved in threshold adjustment for optimal performance. For example, in a recurrent network model of neuronal behaviors in monkeys, cortico‐striatal synaptic efficacy determined the performance on perceptual discrimination, decreasing for easier tasks, resulting in a lower threshold, and vice versa for difficult tasks [Lo and Wang, 2006]. The current findings corroborate this theoretical prediction.

As described earlier, previous studies examined how brain responds to speed accuracy trade‐off during perceptual or eye movement decisions [Domenech and Dreher, 2010; Forstmann et al., 2008; Kayser et al., 2009; Mansfield et al., 2011; Philiastides et al., 2011; Watanabe and Munoz, 2010]. In some of these studies, investigators addressed the within‐subject influence of speed versus accuracy emphasis on their performance, assuming a higher threshold in the accuracy scenario or comparing the threshold across different conditions [Domenech and Dreher, 2010; Forstmann et al., 2008; Mansfield et al., 2011]. Thus, while these previous studies specifically manipulated response speed by instructions, we emphasized both speed and accuracy in the stop signal task, a scenario that is perhaps most similar to the “neutral” condition in Forstmann et al. [2008]. In others, the RT models did not allow threshold and information accumulation to vary simultaneously [Watanabe and Munoz, 2010]. On the contrary, in the current work, individuals performed an identical behavioral task and our model of RT allowed both threshold and growth rate to vary without any additional assumptions. These differences in the details of models and behavioral paradigms may account for the discrepancy in the results. Furthermore, a distinct feature of our model is that information accumulation in the stop signal task is not contingent on external inputs [Hu and Li, 2012], as in studies of perceptual decision. After the go signal is presented, evidence for a response starts to accrue, as each passing moment without a stop signal suggesting a go. Thus, our paradigm draws on a process of self control and should be carefully contrasted with studies that require continuous integration of environmental inputs.

A Broader Role of the pre‐SMA in Response Selection

The focus on the pre‐SMA was motivated from our earlier work on the neural correlates of stop signal inhibition [Chao et al., 2009; Duann et al., 2009; Hu et al., 2012; Li et al., 2006; Zhang and Li, 2012b]. Although other brain regions, such as the right inferior frontal cortex [Aron and Poldrack, 2006] and the basal ganglia [Aron and Poldrack, 2006; Li et al., 2008b] were involved in the component processes of inhibitory control, the pre‐SMA was the only region that showed significant activations in the regression of Go+Stop trials with SSRT (Table 1 and Fig. 2). Importantly, the pre‐SMA activation showed a positive correlation with goRT and a negative correlation with SSRT.

The opposite correlation of pre‐SMA with goRT and SSRT suggests that while a decision to respond during go trials and to not respond during stop trials involves different neural activities, they both engage the pre‐SMA. Previous research reported that pre‐SMA is involved in response inhibition in various cognitive tasks [Buch et al., 2010; Chao et al., 2009; Hsu et al., 2011; Kenner et al., 2010; Li et al., 2006; Neubert et al., 2010]. On the other hand, many other studies support a role of the pre‐SMA in response selection [Kuehn and Brass, 2009; Mueller et al., 2007; Rushworth et al., 2002, 2004]. In a task‐switching paradigm, Rushworth et al. [2002] reported higher pre‐SMA activation when participants had to switch between rules in response to visual cues and motor response modalities. They also found that repetitive transcranial magnetic stimulation (TMS) of the pre‐SMA compromised motor performance during switching in the motor response modality. Using paired‐pulse TMS, Neubert et al. [2011] showed that pre‐SMA facilitates as well as inhibits movement execution in response selection through interaction with the primary motor cortex. The latter study suggests that response inhibition is a result of response selection, a process that is mainly controlled by pre‐SMA. The correlation of pre‐SMA, in opposite directions, with both goRT and SSRT in this study supports the hypothesis that pre‐SMA plays a broader role of response selection in the stop signal task.

In addition, we observed a positive correlation between pre‐SMA activity and MT. Pre‐SMA is part of the medial motor cortical structures responsible for motor preparation and planning, especially during internally timed movements [Cunnington et al., 2002; Deiber et al., 1999; Krieghoff et al., 2009; Lee et al., 1999]. In this study, participants determined their own timing of action because of the conflicting needs to respond quickly and monitor for the stop signal. Hence, pre‐SMA activation in association with MT is consistent with its role in preparation for internally timed movements. Taken together, these results suggest that this medial cortical structure is involved in aspects of stop signal performance that goes beyond response inhibition, in accord with its functional connectivity with a wide array of cortical and subcortical structures [Zhang et al., 2012a].

CONCLUSIONS

In conclusion, we characterized the cerebral processes of threshold setting and information accumulation during decision making in the stop signal task. In addition, we identified a broader role of the pre‐SMA in response selection in the stop signal task. These results not only elucidate the neural bases of individual variation in cognitive control but may also help us understand how these component processes of cognitive control are compromised in individuals with neurological or mental illnesses. It is possible that impulsive behaviors as observed in children with attention deficit hyperactivity disorder, chronically medicated patients with Parkinson's disease, and individuals addicted to illicit substances can be fruitfully examined in light of these findings.

Supporting information

Supplementary Information