Predictors of Effective Working Memory Training in Individuals with Alcohol Use Disorders

Abstract

Background:

Low working memory (WM) capacity is associated with alcohol use disorders (AUDs). The importance of WM to adaptive functioning has led to a recent influx of studies attempting to improve individual WM capacity using various cognitive training methods. The present study aimed to examine the efficacy of complex WM training for improving WM capacity among individuals with AUD.

Methods:

Individuals were randomized to complete either adaptive WM training or active control training. We applied a methodologically rigorous and structured approach, including a battery of near and moderate transfer measures in those with AUDs and a control group. Additionally, we examined cognitive factors (at baseline) and other predictors of adherence, training task improvement, and transfer.

Results:

Results suggest improved WM in individuals with AUDs and controls, as evidenced by improved scores on several transfer measures, after adaptive WM training. However, individuals with AUDs showed poorer adherence and less improvement on the training tasks themselves. Neither IQ, WM, sex, nor condition predicted adherence. Level of training task performance, baseline WM, and IQ predicted transfer task improvement.

Conclusions:

This is the first study to rigorously examine both the efficacy of WM training in those with AUDs, and predictors of successful training program adherence and transfer in a large sample. Among study completers, results suggest that although AUD status does not predict training improvement and transfer. However, AUD status did predict lower program adherence. WM training was more effective in those with higher cognitive ability at baseline. This study provides direct translation to the development of cognitive interventions for treating AUD.

1. Introduction

Poor self-regulation and reduced behavioral control are hallmarks of alcohol use disorder (AUD) and are characterized by impulsive decision-making, difficulty inhibiting prepotent responses, and reward biases (Finn, 2002; Iacono et al., 2000). Several core executive cognitive functions have been linked to poor behavioral control and self-regulation (De Wit, 2009; Finn, Gunn, & Gerst, 2015; Hofmann, Friese, & Strack, 2009; Hofmann, Schmeichel, & Baddeley, 2012). In particular, working memory (WM) capacity is a mechanism integral to self-control and decision-making (Broadway, Redick, & Engle, 2010; Hofmann et al., 2009). Further, research suggests that WM capacity may partially explain the association between externalizing psychopathology, including alcohol problems, and impulsive decision-making and behavioral control (Finn et al., 2015; Gunn & Finn, 2013, 2015; Wiers, Ames, Hofmann, Krank, & Stacy, 2010). These findings suggest that those with AUDs serve to benefit from improved working memory and self-regulation.

The use of cognitive training methods to improve a variety of individual cognitive functions has received increased attention over the past decade of research. The methods and approaches used in WM training vary widely. Promising research is derived from “core” training, or programs thought to target domain-general mechanisms of WM. Characteristically, these programs include multiple tasks spanning modalities, use interference to test maintenance, enforce the use of speed for encoding and retrieval, adapt to individual abilities, minimize automatization, and demand high cognitive workloads or engagement (Morrison & Chein, 2011). Therefore, these programs seek to “train” the central executive (Baddeley, 2007), hypothesized to be the core component of WM, responsible for decision-making and inhibitory control (Baddeley & Hitch, 1974; Cowan, 1995). Specifically, recent complex span training programs require simultaneous storage and processing while performing distracting tasks, thus targeting domain-general attentional control mechanisms of WM (Foster et al., 2017; Harrison et al., 2013). However, a significant portion of the WM training literature utilizes n-back training programs (e.g. Jaeggi, Buschkuehl, Jonides, & Shah, 2011; Jaeggi et al., 2008), which have been repeatedly shown to measure a different construct from complex span tasks (Jaeggi, Buschkuehl, Perrig, & Meier, 2010; Kane, Conway, Miura, & Colflesh, 2007; Redick & Lindsey, 2013), making it difficult to draw conclusions about transfer to complex WM. Given that complex span measures of WM are related to self-regulation and decision-making, these methods provide an appropriate framework for examining WM training in those with AUDs.

Six studies have examined the effects of WM training in alcohol- or drug-use populations: stimulant addicts (Bickel, Yi, Landes, Hill, & Baxter, 2011), methamphetamine addicts (Brooks et al., 2017), methadone-maintenance patients (Rass et al., 2015), heavy alcohol users (Houben, Wiers, & Jansen, 2011), alcohol-dependent individuals (Snider et al., 2018), and a sample of inpatients diagnosed with alcohol, cannabis, or cocaine use disorder (Wanmaker et al., 2017). Although several studies provide evidence of far transfer by reduced impulsive decision-making on the delay discounting task (Bickel et al., 2011), reduced drinking (Houben et al., 2011), and improved self-regulation (Brooks et al., 2017), they also included a number of methodological limitations. Examples include unstandardized training programs (e.g., variable number of training sessions (Bickel et al., 2011), confounding control conditions (Houben et al., 2011), passive controls (Brooks et al., 2017), assessment of transfer with tasks similar to those trained on (Houben et al., 2011), and use of samples simultaneously enrolled in confounding active treatments (Brooks et al., 2017; Wanmaker et al., 2017). It is essential to first examine whether WM training results in expected improvements in cognitive outcomes (i.e., WM capacity itself), independent of expectancy effects, before it can be tested for clinical outcomes and intervention.

As several reviews have noted (Melby-Lervåg, Redick, & Hulme, 2016; Shipstead, Redick, & Engle, 2012), there is significant debate regarding the efficacy of WM training, and studies are often conducted with significant methodological concerns, including no or non-active control groups, narrow outcome measures, small sample sizes, and lack of delayed follow-up sessions. In order to appropriately test the efficacy of a WM training program, an active control group is essential. To create an appropriate control condition, participants should be randomly assigned to either the active WM training, or a control condition that is matched on duration of the intervention, involves similar training procedures (e.g., attending the same number of laboratory sessions), and is adaptive to control for motivation. Extensive batteries of outcome measures that utilize a variety of WM tasks across verbal and visuospatial domains to assess transfer effects further increase robustness of findings. In addition, measuring transfer at a delayed follow-up is essential to show that training effects are maintained after training is complete. Finally, recent work suggests the importance of examining baseline predictors of transfer. For instance, one study suggests that lower baseline cognitive ability (e.g., as may be expected in clinical populations such as those with an AUD) may leave more “room to improve” on WM capacity (Au et al., 2015). However, more recent studies using the same methods employed here suggest that individuals with higher baseline WM may show greater improvement (Foster et al., 2017). This examination of the demographic and cognitive factors associated with WM training-related gains is particularly important for ongoing implementation and dissemination of these methods.

With this in mind, the current study applies a number of important methodological considerations (outlined below in Methods) to examine: 1) the efficacy of a complex WM training program in those with AUDs; and 2) several predictors of training improvement, transfer, and adherence. We examined these aims by randomizing individuals with AUDs and controls (individuals without an AUD or any other externalizing disorder), recruited without expectation of improved cognitive capacity, to a complex span WM training program or an active control training condition.

2. Materials and Methods

The present study applied a number of important rigorous methodological considerations, including: (1) an adaptive (complex-span) WM training program (Harrison et al., 2013) that required significant time (15 sessions in less than 4 weeks), rigor, and both verbal and spatial WM tasks; (2) use of an active and adaptive control training condition (visual search); (3) inclusion of a control group; (4) a battery of near and moderate transfer measures; (5) immediate and delayed (1-month) follow-up assessment; (6) laboratory-controlled testing conditions; (7) examination of individual predictors of successful training; and (8) recruitment of participants without advertisement of potential cognitive benefits (in an effort to reduce expectancy biases; see Foroughi, Monfort, Paczynski, McKnight, & Greenwood, 2016). It should also be noted that the materials presented in the present analyses were a subset of a larger study (in which additional measures were collected) across the testing sessions. As an important first step, we present measures used to focus only on cognitive (i.e., WM) transfer effects. The study was approved by the university Institutional Review Board.

2.1. Participants and Recruitment

Participants (N=145) met criteria for one of the two groups; those with a current AUD (n=69) and control participants (n=76). AUD group status was confirmed using The Diagnostic and Statistical Manual of Mental Disorders’ (5^th ed.; DSM-5; American Psychiatric Association, 2013) criteria for current AUD. Given the high rates of AUD in the target sample (young adults), individuals were required to meet criteria for a current moderate or severe AUD (presence of 4 or more symptoms) in order to ensure a sample with significant clinical symptomatology. Control participants did not meet criteria for any AUD or other externalizing psychopathology (Conduct Disorder, Antisocial Personality Disorder, Attention Deficit Hyperactivity Disorder, Cannabis Use Disorder, or any other Substance Use Disorder). In order to recruit a representative sample, individuals in the AUD group were able to qualify for these other externalizing disorders. Participants were recruited from the community and campus in Bloomington, Indiana using advertisements and flyers calling for individuals similar to the target samples (e.g., “impulsive individuals”, “heavy drinkers”, and “individuals interested in psychological research”, “quiet, introspective, and reflective individuals”). Respondents were screened via telephone before being scheduled for the study. Participants in both groups met the following inclusion criteria: (a) between the ages of 18 and 30 years old; (b) read/speak English; (c) have at least a 6^th grade education; (d) have no history of psychosis or head trauma; and (e) agree to complete a urine screen and Breathalyzer test at the beginning of each session. Urine drug screens were used to corroborate self-report data and confirm group inclusion criteria.

2.2. Procedures

Participants completed a total of 19 testing sessions. Two baseline sessions were used to determine full inclusion and group (AUD or control) eligibility and assess baseline cognitive capacity. Following the completion of baseline sessions, participants were scheduled for training sessions. Each participant completed a total of 15 training sessions. Completion of training sessions was immediately followed by the first follow-up session (within 4 days of the final training session). In this session, participants completed the same battery of WM transfer measures administered at baseline. After one month (or 30 days) had elapsed from the date training was completed, participants completed a second follow-up session, in which the same battery of WM measures were administered. All sessions were completed in the laboratory, and participants were financially compensated for participation. Participants were financially compensated for their participation with $12 hourly for the baseline and follow-up sessions and $15 for each training session. They also had to opportunity to receive the following bonuses: $12 for being on-time to baseline and training sessions, training performance bonuses ranging from $0 - $46, and a $20 completion bonus. Total participant payments ranged from $520- $870.

2.3. Materials

2.3.1. Diagnostic Assessment.

The SSAGA-IV interview (Semi-Structured Assessment for the Genetics of Alcoholism; Collaborative Study on the Genetics of Alcoholism – COGA, 2005) was used to assess group criteria and symptom level predictors of training efficacy. The following modules were administered: Alcohol Use Disorder, Drug and Marijuana Use Disorder, Attention-Deficit Hyperactivity Disorder, Antisocial Personality Disorder, and Conduct Disorder.

2.3.2. Active WM Training and Visual Search Control Training.

Active training (AT) consisted of the adaptive Operation Span (OS) and Symmetry Span (SS) tasks (Harrison et al., 2013). Participants were told they were performing these tasks to improve performance and that the tasks may get easier over time. However, participants were not told that their cognitive capacity was expected to improve in order to offset potential expectancy effects.

In the OS task, participants were presented with a series of trials in which they were asked to make an accuracy judgment on an arithmetic equation (e.g., (6/2) + 1 =5, YES or NO”) before being presented with a to-be-remembered letter (consonant). The task was presented in an adaptive manner, where memory set size (number of equations + letter strings) was increased as the participant improved on the task. All training began at Level 1 difficulty (set sizes 2–4), and set size increased by 1 if the participant was successful on 87.5% or more of the equations and letters across 3 trials (each training session is composed of 8 sets of three trials). If a participant’s accuracy on the equations or letters was 75% or less, the level of the next set decreased by 1. In each training session, the level began at the highest level achieved in the previous session. In the adaptive SS task, participants made symmetry judgments on matrix patterns and recalled matrix locations in the correct serial order. In each trial, participants were presented with a pattern of black and white squares and were asked to make a judgment of symmetry from the vertical axis. Then, participants were presented with a 4×4 matrix with one square highlighted in red, which they were instructed to remember. The number of judgments made and matrix positions per level were the same as the OS task, and the task had identical criteria for level progression.

In the control training, adaptive visual search (VS) tasks (hands and letters) were used. VS tasks are an appropriate control, as they have been shown to be unrelated to measures of WM (Kane, Poole, Tuholski, & Engle, 2006) and have been successfully utilized in a number of WM training studies (De Simoni & von Bastian, 2018; Foster et al., 2017; Guye & von Bastian, 2017; Harrison et al., 2013; Redick et al., 2013; von Bastian, Langer, Jäncke, & Oberauer, 2013). As evidenced by others (Foster et al., 2017; Harrison et al., 2013; Redick et al., 2013), these tasks result in practice-related gains, without significant transfer to other measures of individual WM capacity. In these tasks, participants were presented with an array of letters or hands that included a target stimulus, “F” in the letters version and right pointing hand in the hands version, facing either the right or left (mirror-reversed) and an array of distractor letters or hands (normally and mirror-oriented Es/hands) and/or inverted Ts/hands. Each trial began with a fixation dot in the center of the screen, followed by an array of letters/hands for 500ms, then a mask (16×16 array of black squares) for 2500ms. The size of the letter/hand array varied based on difficulty, beginning with a 2×2 array and advancing based on performance to a maximum of 16×16. Participants were instructed to indicate which direction the target stimulus was facing during the mask presentation. Each array size was presented for 24 trials (1 block), and each training session involved 16 blocks. As in the training tasks, trials increased in difficulty when the participant was accurate on 87.5% of the trials and decreased in difficulty if the participant was less than 75% accurate.

2.3.3. WM Transfer Measures.

Individual WM capacity was measured with a battery of tasks. These measures were used to assess both baseline WM capacity and training-related gains. Two categories of training effects were measured: near and moderate transfer. The near transfer battery was made of tasks structurally similar to those trained on, but with different stimuli (complex span tasks). The moderate transfer battery was made of WM tasks that are structurally different from those trained on (updating tasks). Performance on each of the following tasks was measured by the number of correct stimuli recalled; higher values reflected higher individual WM capacity.

Near transfer tests included the Rotation Span (RTS: Harrison et al., 2013), the Reading Span (RDS: Harrison et al., 2013), and the Auditory Consonant Trigram (ACT: Brown, 1958). In the RTS task, a letter is presented to the participant in either the correct direction or the mirror-reversed direction. The letter is then rotated to one of 8 different angles, and participants are asked to determine if the letter is facing the correct direction or the mirror-reversed direction when turned upright. After making the judgment, participants see a short or long arrow pointing in one of the eight directions. Then, after a set of judgments followed by arrows, participants are asked to click on the arrows that were presented in the correct order. There were a total of 15 trials, with 3–10 arrows each. In the RDS, participants are presented with a series of sets of trials (ranging from 3–7 trials) which involve reading a grammatically correct sentence that either does or does not make logical sense. Participants respond “TRUE” or “FALSE” and are immediately presented with a letter to keep in mind. At the end of each set, they recall the letters in the correct order. In the ACT, participants are presented with 3, 4, and 5 nonsense consonant strings, and then asked to count backwards by 3’s from a 3-digit number for 18 or 36 seconds. As in the other near transfer tasks, they are asked to recall the consonant strings after each set of trials. Each of these tasks (RTS, RDS, and ACT) tap the attentional-control mechanisms of the central executive of WM capacity by requiring the participant to retain information while completing secondary distracting tasks.

The moderate transfer effects of training on WM capacity were assessed using the Running Letter Span (RLS: Harrison et al., 2013), the Running Spatial Span (RSS: Harrison et al., 2013), and the Keep Track (KT: Harrison et al., 2013) tasks. In the RLS task, a series of letters were presented one at a time (2 letters per second). Before the presentation, participants were told they are to remember a set size (number of letters to recall) and recall the most recent set size (n) in the correct serial position. For example, participants could be told to remember the last 5 letters shown, and then presented with 7 letters. Set sizes vary from 3 to 9, and two trials per set size were completed. The RSS task is identical to the RLS task, except that to-be- remembered stimuli are matrix locations on a 4×4 matrix. In the KT task, participants are presented sequentially with 16 words from 6 categories. They are told to remember the most recent instances of a certain number of categories. There were 15 trials, and the number of to-be-remembered categories per trial (set size) were 2, 3, 4, 5, or 6. Before each trial, participants are told the category of the most recent instance they should recall. Each of these tasks require the participant maintain a string of stimuli and respond based on a varying set size. It should be noted that in Harrison et al. (2013) these tasks were also referred to as near transfer measures. However, here we have re-classified them as moderate transfer measures given they are structurally different from the complex span tasks (e.g., no secondary processing task). Although both types of tasks measures WM, improvement on theses updating tasks would represent stronger evidence that individual core WM capacity has been altered, given their unique measurement characteristics and the lower likelihood that strategies practiced on the complex span training would as easily transfer to these tasks.

2.3.4. Intelligence.

The Wechsler Abbreviated Scale of Intelligence (WASI; Wechsler, 1999) was used to assess verbal and performance IQ at baseline. This assessment includes four subtests: block design, vocabulary, matrix reasoning, and similarities.

2.4. Data Analyses

R (R Core Team, 2013) was used to conduct all analyses. In order to examine training task improvement and transfer effects, lme4 (Bates, Mächler, Bolker, & Walker, 2015) was used to perform linear mixed effects analysis. Analysis of variance (ANOVA) and chi square tests were used to examine differences in demographic characteristics and baseline WM capacity by group and condition. Chi-square and logistic regression analyses were used to examine demographic (sex, group, condition) predictors of training adherence in the full sample of individuals who were assigned to training (n = 189). In order to account for shared variance across time points (baseline, follow-up 1, and follow-up 2), linear mixed models were used to examine the main effects of near and moderate transfer (change in individual WM scores from baseline to follow-up sessions), as well as predictors of these transfer effects in the final sample of individuals who completed training and follow-up sessions (n = 145). Model specifications are described in more detail below for each model.

3. Results

3.1. Final Sample Demographics and Training Program Adherence

189 individuals were eligible for the study (i.e., met full inclusion criteria after being interviewed using the SSAGA (COGA, 2005)). Of those 189, 97 were randomized into the AT, and 92 into the VS condition. Seventy-five (77%) of the participants assigned to the AT completed the full protocol (AUD = 35, Control = 40). Seventy (76%) of the participants assigned to the VS (control) training completed the full protocol (AUD = 34, Control = 36), resulting in a full sample of 145. Table 1 presents the final sample demographics (N=145), as well as split by group and condition. Table 1 also presents ANOVA and chi-square statistics comparing demographic variables by group and condition. There was no significant difference in participants’ age, sex, ethnicity, or student status as a function of condition (VS vs. AT), suggesting appropriate baseline comparisons. Additionally, there were no significant differences between group (AUD vs. Control) on age, ethnicity, or student status. Among individuals in the AUD group, there were a number of comorbid current externalizing diagnoses: 17% with attention deficit hyperactivity disorder, 26% with conduct disorder, 17% with antisocial personality disorder, 18% with cannabis use disorder, and 19% with another substance use disorder.

Table 1:

Demographic Characteristics of Final Sample by Group and Condition

		Group			Condition
	Full (n = 145)	Controls (n = 76)	AUDs (n = 69)		VS (n=70)	Active (n = 75)
Age (M, sd)	22.06 (2.42)	22.30 (2.64)	21.80 (2.14)	F(1,143)= 1.59, p = 0.21	22.04 (2.63)	22.08 (2.22)	F(1, 143) = 0.01, p = 0.93
% Female	60	67	52	X2(1) = 3.24, p=0.20	66	55	X2(1) = 1.73, p=0.42
% Caucasian	74	70	80	X2(1) = 1.40, p=0.24	76	73	X2(1) = 0.02, p=0.89
% Student	83	84	81	X2(1) = 0.07, p=0.79	86	80	X2(1) = 0.48, p=0.49
IQ (M, sd)	115.42 (12.61)	116.04 (8.58)	114.74 (15.92)	F(1,143)= 0.38, p = 0.54	114.67 (8.75)	116.11 (15.36)	F(1,143)= 0.47, p = 0.50

Note.

^*.05

^**.01;

^***.001

A chi-square test examining program adherence (successful completion of all 19 sessions including baseline, training, and follow-up sessions) revealed neither gender, χ² (1) = 0.06, p = 0.81) nor condition, χ² (1) = 0.001, p = 0.98) predicted training completion. Additionally, linear regression revealed neither IQ (B = −0.004, p = 0.07), nor baseline WM (B = 0.0069, p = .76), predicted program adherence. However, logistic regression revealed that group did predict adherence (B = 0.25, p < .001), in that AUDs had a higher likelihood of dropping from the study before completion.

3.2. Efficacy of WM Training: Training Task Improvement and Transfer

3.2.1. Training Task Improvement.

A random intercepts mixed effects model was applied, using normalized (z-score) session averages, to examine level of improvement on each task. In order to examine effects of condition and group in the same model, normalized z-scores were computed for a combined score of training tasks received (AT or VS) for each participant. Fixed effects of session (1–15), condition (AT and VS), group (AUD and control), and interactions of session by group, session by condition, and group by condition were entered into the model, along with random intercepts of participant and task. Full results from this analysis are reported in Table 2. This model revealed main effects of session, indicating improvement on training tasks as session increased; and session by condition interaction, suggesting those in VS improved more than those in the AT (Figure 1A). This model also revealed a session by group interaction, suggesting controls improved more across training sessions compared to AUDs (Figure 1B). However, a group by condition interaction was not significant, suggesting that the groups performed similarly across conditions.

Table 2:

Random Intercept Mixed Effects Model of Training Task Improvement

Predictors
	B (95% CI)
(Intercept)	−1.23 (−1.43, −1.04), p < .001
Session	0.17 (0.16, 0.17), p < .001
Group	−0.10 (−0.36, 0.17), p = .488
Condition	0.53 (0.26, 0.79), p < .001
Session × Group	−0.01 (−0.02, −0.00), p < .01
Group × Condition	−0.03 (−0.33, 0.39), p = .867
Session × Condition	−0.07 (−0.08, −0.06), p < .001

Note. σ² = .328, Subject ICC = .522; Group = AUD v. controls; Condition = active WM training v. visual search (control) training; Session = Training sessions 1-15.

Figure 1:

Note. Average training task (z-score) for each training task (Panel A) and averaged across group (Panel B) at each training session. Control training tasks are represented by VS- H = Visual Search- Hands, VS-L= Visual Search- Letters, and active training tasks are represented by OS = Operation Span, SS = Symmetry Span.

3.2.2. Transfer Effects.

ANOVA was used to examine baseline differences in transfer measures between group and condition. Importantly, there were no significant differences between transfer measures by randomized conditions for five out of six measures at baseline, suggesting appropriate comparisons for those five tasks at follow-up (RDS: F(1,140) = 1.69, p = .19; RTS: F(1,139) = .66, p = .42; ACT: F(1,142) = .01, p = .91; RLS: F(1,140) = .12, p = .73; RSS: F(1,140) = 1.21, p = .27; KT: F(1,140) = 5.02, p = .03. Additionally, there were no baseline differences on any WM transfer measures between groups (RDS: F(1,140) = .96, p = .33; RTS: F(1,142) = .06, p = .81; ACT: F(1,142) = .001, p = .97; RLS: F(1,140) = 1.94, p = .16; RSS: F(1,140) = .05, p = .83; KT: F(1,140) = .55, p = .46). Linear mixed models were run for each transfer measure separately to examine effect of session (baseline, follow-up 1, and follow-up 2), condition (AT and VS), group (AUD and control), and group by condition, session by condition, and session by group interactions, with random intercepts of subject and session. Full models are presented in Table 3.

Table 3:

Linear Mixed Model of Near and Moderate Transfer Effects at Follow-up 1 and 2

	RDS	RTS	ACT	RLS	RSS	KT
	B (95% CI)
Group	−4.56 (−12.23, 3.11)	−2.33 (−8.31, 3.65)	−1.28 (−5.06, 2.49)	−5.82 (−11.50, −0.15)*	−2.43 (−7.58, 2.73)	−1.94 (−5.40, 1.52)
Condition	−6.05 (−13.67, 1.57)	−4.14 (−10.06, 1.77)	−1.13 (−4.88, 2.62)	−4.52 (−10.15, 1.12)	−3.93 (−9.04, 1.18)	−3.67 (−7.11, −0.23)*
Group × Condition	5.59 (−4.20, 15.37)	5.05 (−2.63, 12.72)	2.76 (−2.26, 7.78)	7.12 (−0.40, 14.63)*	4.28 (−2.32, 10.87)	2.58 (−1.85, 7.00)
F1 × Condition	−4.19 (−10.31, 1.92)	−7.79 (−12.37, −3.20)***	−2.85 (−5.33, −0.37)*	−2.69 (−6.49, 1.12)	−4.47 (−8.54, −0.40)*	1.51 (−1.22, 4.24)
F2 × Condition	−4.23 (−10.30, 1.84)	−5.69 (−10.26, −1.11)*	−4.05 (−6.53, −1.58)**	1.62 (−2.18, 5.43)	−2.25 (−6.30, 1.81)	1.08 (−1.65, 3.80)
F1 × Group	−0.16 (−6.28, 5.96)	1.03 (−5.63, 3.56)	−1.88 (−4.36, 0.60)	−0.35 (−4.16, 3.46)	−3.06 (−7.14, 1.01)	−2.30 (−5.03, 0.43)
F2 × Group	−0.54 (−6.62, 5.53)	−3.33 (−7.91, 1.62)	−3.73 (−3.73, −1.26)**	−2.93 (−6.73, 0.88)	−4.74 (−8.81, −0.68)*	−5.16 (−7.89, −2.44)***

Note. Effects of interest displayed, full model description provided in text;

^*.05

^**.01;

^***.001;

ACT = Auditory Consonant Trigram, RDS = Reading Span, RTS = Rotation Span, RLS = Running Letter Span, RSS = Running Spatial Span, KT = Keep Track Task, F1 = Follow-up 1, F2 = Follow-up 2; Group = AUD v. controls; Condition = active WM training v. visual search (control) training; Session = Baseline, follow-up 1, follow-up 2.

Near Transfer.

Models testing transfer effects revealed significant session by condition interactions on two near transfer measures, RTS and ACT, at initial follow-up, reflected by improved scores for those subjects randomized into AT. At follow-up 2 (one month), both effects remained significant, suggesting maintenance of transfer effects one month after training completion. These results are presented in Figure 2B and2C. ACT also revealed a group by session interaction at follow-up 2, indicating controls had higher scores at follow-up 2, compared to the AUD group. However, this effect is non-significant at follow-up 1. The third near transfer task (RDS) revealed no evidence of transfer.

Figure 2:

Note. Figure displays average total score for each near (A-C) and moderate (D-F) transfer measures at each session by condition across both groups. ACT = Auditory Consonant Trigram, RDS = Reading Span, RTS = Rotation Span, RLS = Running Letter Span, RSS = Running Spatial Span, KT = Keep Track Task; AT = Active Training Condition; VS= Visual Search Condition. Error bars represent 95% CI.

Moderate Transfer.

Additionally, a significant session by condition interaction was observed on the RSS task, suggesting transfer by improved scores for those subjects randomized into AT (Figure 2E). This effect was not maintained at follow-up session 2. There was no evidence of transfer for the other two moderate transfer measures (RLS or KT). Table 3 and Figures 2D-F present all moderate transfer results, including null findings.

3.3. Predictors of Training Improvement and Transfer Task Performance

3.3.1. Predictors of Training Task Improvement.

In order to more closely examine improvement on the training tasks themselves, cognitive predictors of training improvement were considered. A change score was used for each training task to examine whether cognitive (baseline WM and IQ) factors predicted improvement on any of the four training tasks (OS, SS, VS-H, or VS-L). Each change score was calculated as the difference between average level reached on session fifteen and average level reached on session one. As presented in Table 4, linear regression models (controlling for group) revealed that baseline WM (measured by the OS task) did not predict improvement on any of the four tasks. However, baseline IQ (measured by the WASI) did predict change in both AT tasks (OS and SS), but neither of the VS (control) training tasks. In other words, higher IQ was associated with more training gains on the AT tasks, but not the VS tasks. Group effects were non-significant in all models.

Table 4:

Linear Regression, Cognitive Predictors of Training Task Improvement

	O-Span Change	S-Span Change	VS-H Change	VS-L Change
Predictor	B (95% CI)
WM	−0.00 (−0.05, 0.05)	−0.03 (−0.08, 0.02)	−0.00 (−0.04, 0.03)	0.01 (−0.01, 0.02)
IQ	0.07 (0.03, 0.11)**	0.09 (0.05, 0.13)***	−0.02 (−0.03, 0.07)	−0.00 (−0.03, 0.03)

Note. Effects of interest displayed, full model description provided in text;

^*.05

^**.01;

^***.001;

ACT = Auditory Consonant Trigram, RDS = Reading Span, RTS = Rotation Span, RLS = Running Letter Span, RSS = Running Spatial Span, KT = Keep Track Task

3.3.2. Training Tasks Performance as Predictor of Transfer.

In order to consider level of improvement on trained tasks as a predictor of transfer task improvement, average score at the final session (training session 15) was entered into a random intercepts linear mixed models as a fixed effect for each training task. In each model, session, group, and IQ were also considered fixed effects, and session and subject variables were random. Results are presented in Table 5. These analyses revealed a significant effect of final training score on transfer on all near and moderate transfer measures in both the OS and SS training tasks. However, analyses of performance on the VS tasks did not predict transfer task performance on most tasks. The final session performance in the VS-H task only predicted transfer task improvement on one task (RDS). IQ was non-significant in all models except the ACT for both the VS-L (B = 0.28, p = 0.01) and VS-H (B = 0.30, p = 0.003) models. Similar to main effect transfer models, group effects remained non-significant in all near transfer models. Regarding moderate transfer models, the effect of group was significant on the RLS task in the OS model (B = −5.84, p = .02) only. Overall, these results suggest that actual performance on the AT tasks was an important predictor of transfer task performance, regardless of baseline IQ, and suggests that completion of the AT had a larger impact on transfer measures than the VS (control) training.

Table 5:

Linear Mixed Model, Final Active Training Performance Predicting Transfer Effects

	RDS	RTS	ACT	RLS	RSS	KT
Session 15	B (95% CI)
OS	2.10 (1.16, 3.04)***	1.00 (0.12, 1.81)*	0.64 (0.12, 1.16)*	1.71 (0.97, 2.44)***	1.48 (0.71, 2.26)***	0.95 (0.50, 1.40)***
SS	2.27 (1.18, 3.35)***	1.75 (0.87, 2.64)***	0.66 (0.06, 1.26)*	1.58 (0.70, 2.47)***	2.26 (1.49, 3.04)***	0.88 (0.34, 1.41)**
VS-L	1.57 (−2.54, 5.68)	1.84 (−1.38, 5.05)	1.13 (−0.45, 2.70)	0.68 (−2.36, 3.72)	−0.15 (−2.74, 2.43)	0.40 (−1.49, 2.28)
VS-H	2.39 (0.38, 4.40)*	1.07 (−0.58, 2.71)	0.40 (−0.42, 1.22)	0.31 (−1.24, 1.87)	1.18 (−0.11, 2.46)	−0.05 (−1.02, 0.92)

Note. Effects of interest displayed, full model description provided in text; bolded p values < .05;

^*.05

^**.01;

^***.001;

ACT = Auditory Consonant Trigram, RDS = Reading Span, RTS = Rotation Span, RLS = Running Letter Span, RSS = Running Spatial Span, KT = Keep Track Task; VS- H = Visual Search- Hands, VS-L= Visual Search- Letters, OS = Operation Span, SS = Symmetry Span.

3.3.3. Cognitive Predictors of Transfer Task Improvement.

Linear mixed models were used again to examine whether cognitive ability (WM and IQ) at baseline predicted greater improvement from baseline to follow-up1 and follow-up 2 on transfer tasks. Again, session and subject were set as random intercepts. Other fixed effects included session, condition, baseline WM/IQ, and group. Results are presented in Table 6. Baseline WM (OS) predicted greater improvement on all three near transfer tasks and two moderate transfer tasks (RLS and KT). Additionally, baseline IQ (WASI) predicted greater improvement on two near transfer measures (RTS and ACT) and all three moderate transfer measures. Group effects were non-significant in all near transfer models. However, group effects were significant in both RLS models examining WM (B = −4.24, p = .05) and IQ (B = −4.45, p = .04), and the KT model examining WM (B = −3.07, p = .02), indicating those with AUD were less likely to show improvement when considering cognitive predictors.

Table 6:

Linear Mixed Models, Cognitive Predictors of Transfer Tasks

	RDS	RTS	ACT	RLS	RSS	KT
Predictors	B (95% CI)
WM	0.34 (0.17, 0.50)***	0.28 (0.14, 0.41)***	0.14 (0.06, 0.23)***	0.25 (0.12, 0.38)***	0.08 (−0.04, 0.19)	0.13 (0.05, 0.20)**
IQ	0.19 (−0.00, 0.38)	0.18 (0.03, 0.33)*	0.18 (0.08, 0.28)***	0.19 (0.04, 0.33)*	0.21 (0.08, 0.34)**	0.17 (0.09, 0.25)***

Note. Effects of interest displayed, full model description provided in text; bolded p values < .05;

^*.05

^**.01;

^***.001;

ACT = Auditory Consonant Trigram, RDS = Reading Span, RTS = Rotation Span, RLS = Running Letter Span, RSS = Running Spatial Span, KT = Keep Track Task; VS- H = Visual Search- Hands, VS-L= Visual Search- Letters, OS = Operation Span, SS = Symmetry Span.

4. Discussion

This study investigated the effects of complex WM training on individuals with and without AUDs. Consistent with previous studies examining complex-span training (Foster et al., 2017; Harrison et al., 2013), results revealed significant transfer on two near WM transfer measures (RTS and ACT) at immediate follow-up only for individuals who completed the active training. Importantly, these effects were also maintained at the second follow-up session and were not moderated by group in main effect models. Further, and novel to our study, we also found evidence of transfer on one moderate transfer task (RSS), also independent of group. However, this transfer effect was not maintained at one-month follow-up. We also found that baseline cognitive ability appeared to be important predictors of WM training and transfer task improvement. These results suggest that among individuals who completed training, training gains are at least as likely in individual with AUDs, compared to control subjects. Further, these improvements in WM occurred in a sample of individuals with AUD and rates of comorbid externalizing disorders that represent of the population at large (Compton, Conway, Stinson, Colliver, & Grant, 2005; Kessler et al., 2006; Stinson et al., 2005), increasing the generalizability of our findings.

In addition to examining the efficacy of WM training in individuals with AUDs, we also sought to identify predictors of program adherence, training task improvement, and transfer. To our knowledge, no studies have investigated the factors that may be associated with cognitive training program adherence among clinical samples, and very few have examined these factors in healthy populations (Jaeggi, Buschkuehl, Shah, & Jonides, 2014). Given the large time commitment and cognitive challenges associated with these protocols, this is an important variable to consider in the potential application of these programs. Importantly, the rate of drop-out was higher among individuals with AUDs. There are several factors that may have made it more difficult for those with AUDs to complete the study. For instance, the training schedule required subjects to come in at least four times per week, to provide a breath alcohol level of 0.00, and not be experiencing any hangover symptoms. For some heavy problematic drinkers, it may have been difficult to maintain these study requirements. Importantly, neither gender, baseline cognitive ability, nor condition predicted training program adherence.

The analyses presented here included all subjects who completed training, regardless of how much they improved on training tasks. However, unknown factors such as poor motivation, strategy, or additional cognitive skills may have impacted how much individuals improved on the training. Analyses examining the effect of training task performance on transfer tasks supports this notion. Notably, final level on the AT tasks predicted improvement on transfer tasks on all six (near and moderate) measures. This pattern was only observed in one transfer measure for one control training task. In other words, those who performed better on the AT tasks themselves were much more likely to see general improvements on individual WM capacity. This is a very important finding and may be a more sensitive way to assess actual transfer to core WM resources. These findings are particularly important when communicating the efficacy of WM training, and discussion of these findings should reflect that the more gains in training, the more likely individuals are to see transfer to core WM capacity.

It is also important to note that baseline intelligence predicted level of training improvement on both of the AT tasks, in that higher intelligence was associated with more training success. Similarly, and consistent with Foster et al. (2017), higher baseline WM capacity and IQ predicted improved WM capacity on several near and moderate transfer measures, despite having statistically more “room to improve”, as suggested by Au and colleagues (2016). As recently discussed by Redick, Shipstead, Wiemers, Melby-Lervåg, & Hulme, 2015, these findings may predict some barriers for further implementation of WM training methods among those with lower WM and related executive functioning, in this case for individuals with AUDs. Additionally, in these models an effect of group was revealed for two moderate transfer tasks, suggesting that when considering baseline cognitive ability, individuals with AUD are less likely to show improvement on transfer measures. However, these effects should be understood in the context of some limited generalizability, given it was not consistent across all tasks (including no significant effects on near transfer tasks) and other models (in which baseline cognitive ability was not accounted for).

It is important to understand the findings of this research within the context of its limitations. In particular, characteristics of the sample that limit the generalizability of the results. First, our sample is largely comprised of college students, with a mean IQ of 116, which reflects a higher average than the population at large and represents less variability in cognitive skills. Although our sample size was sufficient for the analyses presented here, we were limited in our ability to examine the role of comorbid externalizing problems on WM improvements. Future studies should recruit larger samples in order to examine the influence of comorbidities on training effects. Additionally, the study was conducted in an on-campus laboratory, which makes access to facilities convenient and familiar and may have affected retention in the positive direction. If these methods were to be implemented in community or medical settings, more barriers to adherence may exist. Finally, it is important to note that there was significant monetary incentive in this study. Individuals who completed the study made an average of $699.55. The study design did not direct recruitment to individuals looking to improve cognitive or memory skills, as in many cognitive training studies. In fact, a recent study indicates that using suggestive recruitment based on potential benefits of cognitive training leads to placebo effects on measures of fluid intelligence (Foroughi et al., 2016). In order to complete a rigorous scientific design, we found it important to minimize expectancy effects that may exist in a sample who is actively trying to improve their cognitive skills. However, without this internal motivation, monetary incentives were necessary to motivate improvement on the training and maintain retention of subjects. Future studies should recruit individuals who were specifically motivated to improve cognitive skills and compare them to a “blind” sample in order to understand if motivation plays an important role in training completion and transfer.

In summary, results suggest that those with AUDs who were able to complete a rigorous protocol improve on WM training programs and show evidence of near and some moderate transfer. However, individuals with AUDs are also less likely to adhere to training program requirements. Further, those who showed more training improvement on the active WM training program were also more likely to show improvement on transfer effects. Results also suggest that individuals who begin the training with higher cognitive ability are also more likely to benefit from training. This and future studies provide direct translation to the development of cognitive interventions for treating AUDs and related externalizing psychopathology.