Evidence Tracing

 Quick Analysis Plan Copied

 Long Analysis Plan

Three presentation choices in the previous Finland visualization retained positive-association framing despite explicit limitations-forward intent, while over-correction produced its own errors:

• Unadjusted-only correlation display: The BMI-tracking figure plotted only unadjusted youth-BMI → adult-IMT correlations (r = 0.10–0.26), implying a graded association. The same source reports that after adjusting for adult BMI, these attenuate to r = −0.08 to 0.05 — effectively null. Plotting only unadjusted values is a residual positive-association choice that prose caveats do not undo Young Finns unadjusted vs adjusted.
• Point estimates without published CIs: Table 3 group IMT values (e.g., 0.610 [0.603–0.616] vs 0.627 [0.610–0.644], N=112) have substantially overlapping 95% CIs, but the bar chart showed means only, visually overstating separation.
• Cohort-age conflation: Adult obesity prevalence was reported by adult age (7.2% at 24; 15.9% at 39) but plotted against baseline-age cohorts, creating a spurious cohort-level gradient.

Figure. The gap between unadjusted (blue) and adjusted (gray) correlations is itself the finding: youth BMI has no independent association with adult IMT once adult BMI is accounted for. The previous response's figure showed only blue bars — a visual framing choice that no text caveat fully counteracted.

Errors from over-correction

• Conscript attrition (~82%) was computed from "1430 invited / 45% response" — figures absent from the evidence records, which report only n=256 analyzed conscripts Finnish conscript dietary cohort.
• The Cochrane "~54% attrition" was generalized from a single trial to the review level, though the review reports 7 studies (N=6362) with variable attrition, not a pooled rate Cochrane youth smoking incentives.
• Nonresponse at recruitment, dropout during follow-up, and loss to measurement were conflated as equivalent "attrition," obscuring which bias threatens which inference.

Calibrated standard going forward

• Plot adjusted and unadjusted estimates side-by-side when both are published
• Show CIs as error bars whenever the source provides them
• Separate nonresponse from dropout in attrition figures
• Restrict all plotted values to source-traceable records
• Interpret effect sizes in clinical units, not P-values alone (childhood smoking β=0.011 mm IMT is negligible without clinical context)

What would change this assessment: If re-reading the source papers showed the Young Finns authors analytically justified presenting unadjusted values alone (e.g., as a screening step before adjustment), or if the adjusted correlations were not actually null, this audit would require revision. The assessment is also limited to the evidence records supplied; the full papers were not independently re-examined.