Near transfer to an unrelated N-back task mediates the effect of N-back working memory training on matrix reasoning

Abstract

The extent to which working memory training improves performance on untrained tasks is highly controversial. Here we address this controversy by testing the hypothesis that far transfer may depend on near transfer using mediation models in three separate randomized controlled trials (RCTs). In all three RCTs, totalling 460 individuals, performance on untrained N-back tasks (near transfer) mediated transfer to Matrix Reasoning (representing far transfer) despite the lack of an intervention effect in RCTs 2 and 3. Untrained N-back performance also mediated transfer to a working memory composite, which showed a significant intervention effect (RCT 3). These findings support a model of N-back training in which transfer to untrained N-back tasks gates further transfer (at least in the case of working memory at the construct level) and Matrix Reasoning. This model can help adjudicate between the many studies and meta-analyses of working memory training that have provided mixed results but have not examined the relationship between near and far transfer on an individual-differences level.

Extended data Figure 1 |. Schematic of single mediator models with post-test Countermanding as the dependent variable.

Path coefficients are standardized regression coefficients with p values reported in brackets. Bold text indicates statistical significance. Covariates (not displayed): Pre-test Countermanding and Pre-test N-back. a, Retrospective study, whole sample (N = 271). b, Retrospective study, RCT 1 (N = 94). c, Retrospective study, RCT 2 (N = 177).

Extended data Figure 2 |. N-back training progress.

Mean n-level achieved per day (40 minutes of training) for active groups in RCT 1 (N = 72), RCT 2 (N = 109), and RCT 3 (N = 128), weighted by the number of trials. RCT differences may reflect the use of slightly different adaptive algorithms. Shaded areas represent standard error of the mean.

Extended data Figure 3 |. Scatter plots demonstrating relationships between main outcome measures at pre-test in RCT 1 (N = 94) and RCT 2 (N = 177).

Data is collapsed across training conditions and solid lines represent the linear fit of data. a, Near Transfer (untrained N-back) and Matrix Reasoning in RCT 1. b, Near Transfer and Countermanding in RCT 1. c, Matrix Reasoning and Countermanding in RCT 1. d, Near Transfer and Matrix Reasoning in RCT 2. e, Near Transfer and Countermanding in RCT 2. f, Matrix Reasoning and Countermanding in RCT 2.

Extended data Figure 4 |. Scatter plots demonstrating relationships between main outcome measures at post-test in RCT 1.

Solid lines represent the linear fit of data. a, Near Transfer (untrained N-back) and Matrix Reasoning performance for N-back training group (N = 72). b, Near Transfer and Countermanding performance for N-back training group (N = 72). c, Matrix Reasoning and Countermanding performance for N-back training group (N = 72). d, Near Transfer and Matrix Reasoning performance for No-contact control group (N = 22). e, Near Transfer and Countermanding performance for No-contact control group (N = 22). f, Matrix Reasoning and Countermanding performance for No-contact control group (N = 22).

Extended data Figure 5 |. Scatter plots demonstrating relationships between main outcome measures at post-test in RCT 2.

Solid lines represent the linear fit of data. a, Near Transfer (untrained N-back) and Matrix Reasoning performance for N-back training group (N = 109). b, Near Transfer and Countermanding performance for N-back training group (N = 109). c, Matrix Reasoning and Countermanding performance for N-back training group (N = 109). d, Near Transfer and Matrix Reasoning performance for No-contact control group (N = 68). e, Near Transfer and Countermanding performance for No-contact control group (N = 68). f, Matrix Reasoning and Countermanding performance for No-contact control group (N = 68).

Extended data Figure 6 |. Scatter plots demonstrating relationships between main outcome measures at pre-test and post-test in RCT 3 (N = 189).

Circles represent the N-back training group (N = 128) and cross marks represent the Active control group (N = 61). a, Matrix Reasoning and Near Transfer (untrained N-back) performance at pre-test b, Matrix Reasoning and Countermanding performance at pre-test. c, Near Transfer and Countermanding performance at pre-test. d, Matrix Reasoning and Near Transfer performance at post-test e, Matrix Reasoning and Countermanding performance at post-test. f, Near Transfer and Countermanding performance at post-test.

Figure 1 |. Parallel mediation models in the Retrospective study (RCT 1 and RCT 2).

In all models, the independent variable is Training group (N-back training or No-contact control), mediators are post-test N-back and Countermanding performance, and the dependent variable is post-test Matrix Reasoning performance. Path coefficients are standardized regression coefficients with p values reported in brackets. Bold text indicates statistical significance. No corrections for multiple comparisons were made. Covariates (not displayed): Pre-test N-back, Pre-test Countermanding and Pre-test Matrix Reasoning. a, Retrospective study; whole sample (N = 271). b, Retrospective study; RCT 1 (N = 94). c, Retrospective study; RCT 2 (N = 177).

Figure 2 |. Parallel mediation models in the Replication study (RCT 3).

In all models, the independent variable is Training group (N-back training or Active control) and the dependent variable is post-test performance on Matrix Reasoning, with pre-test data for all included measures entered as a covariates. Path coefficients are standardized regression coefficients with p values reported in brackets. Bold text indicates statistical significance. No corrections for multiple comparisons were made. a, RCT 3 (N = 189). Mediators are post-test N-back and Countermanding. b, RCT 3 - analytical subset (N = 170). Mediators are WM Composite 1, consisting of N-back, Sequencing, Simple Span, and Complex Span, and IC Composite, consisting of Countermanding and UCancellation tasks. c, RCT 3 - analytical subset (N = 170). Mediators are WM Composite 2, consisting of Sequencing, Simple Span, and Complex Span, and the IC Composite.

Figure 3 |. Alternative mediation models of transfer to Matrix Reasoning in the Replication study (RCT 3).

In all models, the independent variable is Training group (N-back training or Active control) and the dependent variable is post-test performance on Matrix Reasoning, with pre-test data for all included measures entered as a covariates. Path coefficients are standardized regression coefficients with p values reported in brackets. Bold text indicates statistical significance. No corrections for multiple comparisons were made. RCT 3 - analytical subset (N = 170). a, Parallel multiple-mediator model. Mediators are post-test N-back, Running Span, Sequencing, Simple Span, and Complex Span tasks. b, Serial mediator model in which the first mediator is post-test N-back and the second is WM Composite 2, consisting of Sequencing, Simple Span, and Complex Span tasks. c, Serial mediator model in which the first mediator is post-test N-back and the second is IC Composite, consisting of Countermanding and UCancellation tasks.

Figure 4 |. Alternative mediation models of transfer to composite measures in the Replication study (RCT 3).

In each model, the independent variable is Training group (N-back training or Active control) and the dependent is Gf Composite (a), WM Composite 2 (b), or IC Composite (c), with pre-test measures entered as a covariates. Path coefficients are standardized regression coefficients with p values reported in brackets. Bold text indicates statistical significance. No corrections for multiple comparisons were made. RCT 3 - analytical subset (N = 170).

Figure 5 |. Composite z-scores at pre-test relative to post-test in the Replication study (RCT 3).

Data are presented as mean values +/− SEM, plotted separately for N-back training (filled circle; N = 115) and Active Control (empty circle; N = 55) groups. a, WM Composite 1, consisting of N-back, Sequencing, Simple Span, and Complex Span. b, WM Composite 2, consisting of Sequencing, Simple Span and Complex Span. c, IC composite, consisting of Countermanding and UCancellation. d, Gf Composite, consisting of UCMRT, ETS Surface Development, ETS Space Relations, and DAT Form Board.