Docosahexaenoic acid for reading, cognition and behavior in children aged 7-9 years: a randomized, controlled trial (the DOLAB Study)

362 children randomized from 1,376 invited (675 consented and screened), across 74 mainstream primary schools in Oxfordshire, UK; 359 of 362 reassessed at 16 weeks

Healthy UK children aged 7-9 initially underperforming in reading (at or below the 33rd centile). Well-nourished by any global standard; no deficiency was documented as an inclusion criterion. Mean baseline word reading standard score 84.7 (sd 6.3), about 1.5 sd below the norm; 20.2% eligible for free school meals; 91.2% white.

Parallel-group, fixed-dose, randomized, double-blind, placebo-controlled trial. 600 mg/day DHA from algal oil versus taste- and colour-matched corn/soybean oil placebo, 16 weeks. Outcomes are AGE-STANDARDIZED published instruments - BAS II Word Reading, BAS II Recall of Digits Forward/Backward, and the long-form Conners Parent and Teacher Rating Scales - not researcher-designed measures, which is a genuine strength. Randomization by an independent statistician (Centre for Statistics in Medicine, Oxford) with minimization on school and sex; blinding verified post hoc and intact. Registered NCT01066182 / ISRCTN99771026. THE CRITICAL DESIGN FACT - the intention-to-treat analysis on the full sample was NULL for reading (p = 0.279); the positive result is confined to a subgroup of 224 children at or below the 20th reading centile (p = 0.041) and a further 105 at or below the 10th centile (p = 0.011). The pre-planned claim CHECKS OUT on the papers own account: the original protocol set inclusion at the 20th centile and was relaxed to the 33rd before first randomization only because recruitment was falling short, so the 20th- and 10th-centile analyses were specified in the original protocol rather than chosen after seeing the data, and a treatment-by-subgroup interaction test was run and was significant (t = 2.152, p <= 0.05). It is still a subgroup, the interaction test is itself underpowered, and the abstract conclusion ("safe and effective way to improve reading and behavior") is stated far more strongly than the ITT result supports. Two further discounts the record must carry. (1) The reading-AGE version of the same data does NOT reach significance in the 20th-centile subgroup (Active +4.9 months v Placebo +4.1, p = 0.239) - so the widely quoted "0.8 months extra reading age" is a non-significant contrast derived from the significant standardized-score contrast. (2) Behaviour ratings had >15% missing data (change scores n = 246 of 362 for parents, 273 for teachers) and missing values were imputed with treatment-group medians, so the parent-rated behaviour finding rests on imputation over roughly a third of the sample. Industry ties: funded by Martek Biosciences, which also supplied product and placebo (stated to have had no role in design, analysis or publication); the first author declares paid consultancy for companies producing or promoting omega-3 - hence independence: developer-involved rather than independent.

No effect of DHA on reading in the full randomized ITT sample (Active +1.5, sd 4.4; Placebo +1.2, sd 4.3; p = 0.279; observed d approximately 0.07). A significant effect in the pre-planned subgroup of 224 children at or below the 20th centile (+2.0 v +0.9, p = 0.041, d approximately 0.27) and a larger one in the 105 at or below the 10th centile (+3.1 v +0.9, p = 0.011, d approximately 0.51). The same subgroup contrast expressed as reading age is NOT significant at the 20th centile (p = 0.239). Parent-rated ADHD-type behaviour improved on 8 of 14 Conners scales in ITT, but only 4 of 14 in the per-protocol analysis, and NO parent-rated effect survived within the reading subgroups. Little or no effect on TEACHER-rated behaviour (1 of 14 scales, Perfectionism, p = 0.033, on which both groups were below the population mean) or on working memory change scores - the divergence between parent and teacher ratings on a supposedly blinded trial is itself a warning sign. The authors call for replication; they later ran it and it failed.

None internally - double-blind placebo-controlled randomization. The relevant threat is subgroup-selection and outcome-multiplicity, not heredity.

DIRECTLY REPLICATED AND FAILED. DOLAB II (Montgomery et al. 2018, PLoS ONE 13:e0192909, slug nutr-montgomery-2018-dolab2-dha-replication) was run by the same team, pre-registered, with the same dose, duration, placebo and instruments, in MORE schools (84 v 74) and MORE children (376 v 362), and recruited directly into the sub-population where DOLAB I found its effect (baseline reading below the 20th centile). It found nothing: whole-sample reading change Active 0.64 (sd 3.7) v Placebo 0.83 (sd 3.6), p = 0.616, and in the pre-planned sub-10th-centile subgroup 1.4 (3.6) v 1.4 (3.7), p = 0.938. The post-intervention group mean difference on BAS II reading was 0.594 points (95% CI -1.937 to 0.749, reported as placebo minus active, so numerically FAVOURING active but far from significant), and the authors report the observed effect size on the primary outcome as d = 0.05. Under the replication rule the DOLAB I subgroup result cannot support a verdict above mixed, and the honest reading is that it was noise. Caveat in both directions: DOLAB II was a SELF-replication by the same team and funder, so it is not an independent replication - but a self-replication has every incentive to succeed and did not, which makes the failure harder to dismiss, not easier.