Does Working Memory Training Transfer? A Meta-Analysis Including Training Conditions as Moderators

[XFMQ902SF]

A meta-analysis was undertaken to reexamine near- and far-transfer effects following working-memory training and to consider potential moderators more systematically. Forty-seven studies with 65 group comparisons were included in the meta-analysis. Results showed near-transfer effects to short-term and working-memory skills that were sustained at follow-up with effect sizes ranging from g = 0.37 to g = 0.72 for immediate transfer and g = 0.22 to g = 0.78 for long-term transfer. Far-transfer effects to other cognitive skills were small, limited to nonverbal (g = 0.14) and verbal (g = 0.16) ability and not sustained at follow-up. Several moderators (e.g., duration of training sessions, supervision during training) had an influence on transfer effects, including far-transfer effects. We present principles for how best to improve working memory through training in the narrow-task paradigm and conjecture how best to improve basic cognitive functions in complex activity contexts.

http://www.tandfonline.com/doi/full/10.1080/00461520.2015.1036274#.VW5mM_lVhBc

[Shoh]

Does anyone has access to the full text? I am especially interested in this part: "We present principles for how best to improve working memory through training in the narrow-task paradigm and conjecture how best to improve basic cognitive functions in complex activity contexts."                    

I would greatly appreciate if someone provides a link to the full text.

[Gwern Branwen]

"Does Working Memory Training Transfer? A Meta-Analysis Including
Training Conditions as Moderators", Schwaighofer et al 2015
fulltext: http://sci-hub.org/downloads/e224/10....@00461520.2015.1036274.pdf
https://www.dropbox.com/s/084fvteji1tyz8t/2015-schwaighofer.pdf
supplement: http://www.tandfonline.com/doi/suppl/10.1080/00461520.2015.1036274/suppl_file/hedp_a_1036274_sm0716.zip

I don't think there are any surprises here for anyone who has read
Melby-Lervag & Hulme 2013 or the other meta-analyses: messy literature
but clear near-transfer and little far-transfer.

My interest, as always, is in nonverbal ability and control groups, so
I'll just quote those bits:

> *Nonverbal ability*. Figure 5 shows forest plots for 45 immediate and 11 long-term (delayed) transfer effect sizes. Effect sizes varied significantly among studies, Q(44) D 95.44, p < .001, I 2 D 53.90%. After removal of the effect sizes of the studies by Klingberg et al. (2002) and Schmiedek et al. (2010), there was a nonsignificant small effect, g D 0.01, 95% CI [¡0.03, 0.23], p D .13. Effect sizes varied significantly among studies, Q(41) D 79.97, p < .001, I 2 D 48.73%. The additional exclusion of the effect sizes of the studies by Harrison et al. (2013) and Richmond et al. (2011) resulted in a small effect, g D 0.12, 95% CI [0.01, 0.22], p < .05. After removal of outliers, the mean effect size ranged from g D 0.08, 95% CI [¡0.04, 0.20], to g D 0.17, 95% CI [0.06, 0.28]. After removal of outliers and the effect sizes of the studies by Klingberg et al. (2002) and Schmiedek et al. (2010), there was a nonsignificant small mean effect, g D 0.08, 95% CI [¡0.01, 0.18], p D .08. The funnel plot and Egger’s test indicated no publication bias. Thus, in line with Hypothesis 1.4, there was a small immediate transfer effect of WM training to nonverbal ability. We conducted follow-up testing an average of 6.54 months after the posttest. After removal of outliers, the mean effect size was g D ¡0.12, 95% CI [¡0.32, 0.09]. The funnel plot and Egger’s test indicated no publication bias. Thus, in line with Hypothesis 1.4, there was no sustained transfer effect of WM training to nonverbal ability.

g=0.08 is trivially small and could be explained by minor
methodological issues like loss of g-loading, but is as expected from
Melby-Lervag & Hulme 2013.

> *Type of control group*. We assumed that the mean effect on nonverbal ability for the comparison of training groups with passive control groups is larger than for the comparison of training groups with active control groups (Hypothesis 2.12). As can be seen in Table 4, type of control group explained variability in transfer effects to mathematical abilities. Type of control group was the only significant moderator for transfer effects to mathematical abilities, Q(1) D 4.58, p < .05. Eta-squared was large (.24), and the residual variance was not significant (p D .32). The mean effect size for comparisons of training groups with passive control groups was significantly larger than the mean effect size for comparisons of training groups with active control groups. Contrary to Hypothesis 2.12, type of control group was not a significant moderator of transfer effects of WM training to nonverbal ability. Because type of control group was a significant moderator for transfer effects only on mathematical abilities, and no other moderator explained variability in these transfer effects, type of control group could not be considered in multiple regression analyses for other moderators.

A passive/active split in only one domain is a little bit surprising
given my n-back meta-analysis (even Au et al acknowledges that there
is a passive/active split, just they offer up a different
explanation), so I looked in more closely and at the supplementary
information. I think the problem is both that a lot of the studies
aren't n-back and so may not have the same problem and also that they
are far from comprehensive. No Qiu, no polar, no Kundu, no Zhong, no
Clouter, Oelhafen, Smith, Sprenger etc etc. Missing so much data, it's
unsurprising if they reach weaker results and fail to reject the null.

--
gwern