Using drift diffusion modeling to understand inattentive behavior in preterm and term-born children.

Participants

Sample recruitment is described in detail in Retzler et al. (2019) and a full description of all children tested is presented in the online supplementary material . In brief, following identification from hospital records and tracing of all babies born VP (≤32 weeks’ gestation) and admitted for neonatal intensive care in Nottingham University Hospitals NHS Trust, 65 children were recruited (16% of eligible births) to the study. As a comparison group, 48 term-born children (≥37 weeks’ gestation) were then recruited from the same geographical area, using advertisements distributed via local schools and in the community, as well as the University of Nottingham volunteer database. This was a two-stage process that screened for inattentive symptoms using the parent-rated SWAN scale (Stage 1), before inviting families to participate in the full study (Stage 2). This process ensured that the seven points on the SWAN scoring scale were represented in the term-born children, reflecting a range of attentional abilities (far below average, below average, slightly below average, average, slightly above average, above average, and far above average).

The subsample for the current analysis comprised all children with available task data suitable for the DDM analysis (see online supplementary material for full explanation of why data for some children were unavailable). Ten children in each group achieved a 100% hit rate, which prevents calculation of DDM parameters and rendered their data unsuitable for the analysis. This resulted in a subsample of 32 term-born children and 33 children born very preterm aged 8–11 years. Comparisons of the children included in the DDM analysis with those not included, revealed no differences in gestational age at birth, sex, ethnicity, socioeconomic status (as measured using the English Indices of Multiple Deprivation; McLennan et al., 2011 ), IQ, or scores of parent-rated inattention and hyperactivity ( p > .1 in all cases; see online supplementary material ).

Moreover, the VP children in the subsample for this analysis did not differ significantly from the wider eligible VP children ( n = 374; excluded either due to nonrecruitment to the study, or due to unsuitable data for the DDM) with respect to gestational age ( p = .34), birth weight ( p = .46), sex ( p = .41), or socioeconomic status, ( p = .19).

Experimental Task

Children were asked to complete a CPT-AX programmed using PsychoPy software ( Peirce, 2007 ) while electroencephalography (EEG) measurements were recorded as the last part of a test battery (EEG data not reported here). Children were seated at a desk in a quiet, unlit room facing a computer screen while wearing the EEG recording cap. An experimenter remained with them in the testing room at all times.

At the start of the task written instructions appeared on the screen to familiarize the children with the stimuli that represented cues and targets. The stimuli consisted of black abstract shapes (chosen so that they did not have a verbal label) filled with different patterns presented on a gray background (see Figure 2 ). One stimulus was designated as the target stimulus (in CPT-AX nomenclature, this represents the X stimulus) and one stimulus as the cue stimulus (in CPT-AX nomenclature, this represents the A stimulus). The same shapes were designated as cue and target for all children. The instructions were read out by the experimenter who told each child that they were required to respond as quickly as possible when they saw a cue-target sequence. They were informed that the cue shapes and target shapes might also appear in isolation and it was reiterated that it was only when they saw a cue-target sequence in the specified order that they needed to respond.

A continuous stream of stimuli was presented in the center of the screen. Each stimulus was presented for 250 ms separated by an interstimulus interval of 1,400 ms, during which a central fixation cross was displayed (see Figure 2 ). A cue-target “go” (A-X) trial was defined as a trial-pair where the stimuli designated as the cue and target were presented consecutively. Each time the child saw the target stimulus sequentially following the cue stimulus, they were required to respond as quickly as possible pressing the left-most button on a Cedrus RB-730 button box with their right hand. Children were instructed to keep their finger over the response button so that they could respond as quickly as they could. No response was required to other trial types, including those where the cue and target were presented in isolation from one another.

The task consisted of four blocks of 100 trials, with the cue stimulus, target stimulus and 11 different distractor stimuli presented. Trials were presented in a pseudorandomized order, with different orders for each block, but identical orders across participants. “Go” (A-X) cue-target sequences were presented 10 times within each block, as were cue-without-target “no-go” trials (A-not-X), and uncued-target “no-go” (X-not-A) trials. On “go” trials, participants were required to respond within 1,650 ms of stimulus onset (prior to the presentation of the subsequent stimulus) to be considered “correct.”

Commission errors

The total number of responses made on “no-go” trials (any trial other than a cued target) was summed as a measure of commission errors. This was reported as a percentage of erroneous responses out of the 360 “no-go” trials (error rates were too low to permit differentiation between type of “no-go” trial). Higher scores represent less accurate performance and therefore greater impulsivity.

Response time

The mean response time on correct hit (A-X) trials was calculated as a measure of response speed. Higher values represent slower response speed.

Response time variability

Finally, the standard deviation of response time on correct hit trials was calculated as a measure of response speed variability. Higher values represent greater variability in response speed.

Participant Characteristics and Clinical Symptoms

An age-standardized estimate of full scale IQ (FSIQ-2) was calculated from the Wechsler Abbreviated Scale for Intelligence ( Wechsler, 1999 ) using the vocabulary and matrices subtests. Inattentive and hyperactive-impulsive behaviors were measured using the raw scores on the SWAN rating scale ( Swanson et al., 2012 ) a parent-report measure of a child’s ADHD symptoms. This scale has been considered appropriate for use in community populations ( Swanson et al., 2012 ) as it allows measurement of variation in above average attention in addition to below average attention (more severe inattention; Arnett et al., 2013 ). To characterize inattention and hyperactivity in more clinical terms, measures of symptoms and risk of ADHD were assessed using the Conners 3-P ( Conners, 2008 ), with higher scores indicating greater symptoms. Children with scores above the predefined clinical cut-off were classified as “at risk” of diagnosis.

Diffusion Model Fitting

In light of the infrequent target stimuli in this task and consequent low trial numbers ( n = 40), as well as the fact that error rates on “go” trials were low, the EZ-DDM ( Wagenmakers et al., 2007 ) was used. The EZ-DDM is a simplified version of the full drift diffusion model and is ideal for use both with small numbers of trials, and with low numbers of error rates ( Wagenmakers, van der Maas, Dolan, & Grasman, 2008 ).

The EZ-DDM ( Wagenmakers et al., 2007 ) was fitted to the “go” RT and accuracy data from our task using custom R scripts (Wagenmakers provides customizable scripts for R here: https://www.ejwagenmakers.com/2007/EZ.R ). The EZ-DDM transforms a participant’s overall accuracy, mean RT, and variance in RT using three equations derived from the full DDM (see Wagenmakers et al., 2007 for a more detailed description of the procedure). This simplification allows calculation of the most cognitively salient DDM measures without complex parameter fitting procedures often requiring large numbers of trials. This comes at the cost of a more detailed account of behavior provided by the full DDM which includes measures of cross-trial variability and starting point. For the purposes of EZ-DDM, such parameters are kept constant.

The EZ-DDM provides estimates of drift rate, boundary separation, and nondecision time parameters for each participant. For drift rate, a higher value indicates that more information is processed per unit of time (higher drift rate indicates more efficient information processing); for boundary separation, a higher value indicates greater separation between decision boundaries and thus a more conservative approach to decision making based on greater evidence accumulation, conversely a lower value suggests an impulsive decision. Finally, for nondecision time, a higher value indicates greater time spent encoding the stimuli and preparing and executing responses.

Prior to fitting, response times below 200 ms were rejected as these are likely to reflect anticipatory responses from participants prior to cognitive processing of the current stimuli ( Ratcliff & McKoon, 2008 ).

To test the goodness of fit of the DDM model, 1,000 trials for each group were simulated using the “multisimul” function in the DMAT toolbox for MATLAB. The simulated data were created using model parameters (drift rate, boundary separate, nondecision time) averaged across individuals within each group to create mean parameters. Using these parameters, DMAT was used to create two simulated supersubjects. The mean RTs and accuracy from the observed data were then compared with those of our simulated data.

Statistical Analysis

As children in both groups presented with a range of levels of parent-rated inattention, group differences in cognitive performance were not expected, but were analyzed to provide context. Group differences in CPT-AX performance were examined using a MANCOVA with group (term-born or VP) as the between-subjects factor and age entered as a covariate as this differed between groups.

To assess the pattern of association between parent-rated inattention and task performance and DDM measures, first partial correlations controlling for age were performed, both across groups and split by group. Next, in order to assess the independent contribution of these variables for explaining the variance in parent-rated inattention, any task performance or DDM measure that showed a significant correlation with parent-rated inattention in either term-born or VP children was entered into a hierarchical multiple regression, with parent-rated inattention as the dependent variable. Group and age were entered into the first step, and hit rate, RT variability, and drift rate were entered in the second step. Due to high intercorrelations between task performance and DDM measures (see online supplementary material ), when these were entered in the second step, a data-driven stepwise-entry selection technique was used so that only those variables that added significant variance above and beyond that accounted for in the preceding steps were entered. This approach has been used previously ( Aarnoudse-Moens, Weisglas-Kuperus, Duivenvoorden, van Goudoever, & Oosterlaan, 2013 ) to better separate out effects among variables that are interrelated. In order to investigate any group-specific effects, group interaction terms were included as predictor variables in a regression analysis. However, to assess whether the interaction terms explained additional variance while accounting for the loss of any task performance or DDM measures in the second step of the analysis due to the use of stepwise-entry variable selection, a separate regression analysis was conducted using forced entry technique at all steps, in which group interaction terms were added in a final step.