Database of Nutrient-Outcome Associations

We identified a total of 34 eligible associations between nutrients and health outcomes (Table 1). Sixteen of 34 were based on meta-analyses of epidemiological studies and of RCTs that were reported in the same paper,^12,19–27 11 were based on “matched” meta-analyses that were reported in separate papers,^26,28–36 3 on a meta-analysis of observational studies matched with a large RCT (>1,000 participants),^35,37–39 and 4 on a meta-analysis of RCTs matched with a large epidemiological study (>5,000 participants).^40–45 Some of the included papers reported results from a meta-analyses of more than one nutrients and health outcome pair.

Table 1

Eligible nutrient-outcome associations.

In Table 1, 20 out of 34 associations pertained to vitamins, 7 pertained to minerals (such as calcium) and trace elements (such as selenium), and 7 to either macronutrients (fiber) or fatty acids. The examined endpoints span a wide range of clinical outcomes including mortality, stroke, and other cardiovascular outcomes, various cancers, acute macular degeneration, and fractures, and surrogate outcomes such as systolic and diastolic blood pressure. In two cases nutrient intake of the mother was examined with relation to a clinical outcome in the child (prenatal multivitamin supplementation in the mother and risk of neural tube defects in the fetus, and maternal calcium intake and blood pressure in the offspring).

Concordance or Discordance Between Epidemiological and RCT Data

Table 2 shows concordance between the summary effects in the two research designs using the three definitions. Using the first definition, 6 out of 34 topics (18%) were qualitatively concordant, and the remaining 28 were classified as “unclear” (there was no topic with qualitatively discordant epidemiological and RCT data). Twelve of 34 associations (35%) had a statistically significant z-score, and using the second definition, evidence from epidemiological studies and evidence from RCTs was statistically significantly discordant. The remaining 22 associations had no evidence of significant discordance. Of the 12 association with a statistically significant z-score, in 6 the meta-analysis of the RCTs and that of the epidemiological studies pointed to different directions. Thus, six examples were quantitatively discordant using the third definition.

Table 2

Qualitative and quantitative concordance of effects in epidemiological studies and RCTs.

Citation Graphs

We were able to construct citation graphs for 28 topics (Table 3). For 6 topics we were unable to form reliable citation graphs (for technical reasons related to changes in the format of the citation information obtained from the ISI Web site). Briefly, there was variability in the total number of vertices in the 28 citation graphs: The median number of vertices (articles) was 253 (25^th–75^th percentiles: 95, 356), and the median number of edges (citation relationships) was 1181 (25^th–75^th percentiles: 255, 1620). The citation graphs are relatively sparsely connected with median density 0.018 (low density or low connectivity is the norm for citation graphs). The table shows also the mean hub and authority scores across all papers in a graph, and across the 20 papers with the highest respective scores. Graphs with higher mean hub scores and higher mean authority scores have more papers that “integrate” and “provide” information (as conveyed by citaiton relationships), respectively, compared with graphs with lower scores. When all papers in a graph were considered, the median hub score was 0.09 (25^th–75^th percentiles: 0.07, 0.13) and the median authority score was 0.13 (25^th–75^th percentiles: 0.10, 0.21). Finally, similar variability across the 28 topics was observed in the mean number of citation received, citations made, or total citations made or received. When all papers in a graph were considered the median numbers and 25^th–75^th percentile ranges were 4.3 (3.2, 4.7), 4.3 (3.2, 4.7) and 8.5 (6.3, 9.4), respectively. (When all papers are considered in a citation graph, the mean number of citations made equals the mean number of citations received.) The corresponding mean scores or numbers over the 20 papers with the largest respective scores or numbers are also shown (Table 3).

Table 3

Summary of citation analysis of nutrient exposure-health outcome pairs studied by systematic reviews or large studies.

Associations Between Citation Graph Metrics and Qualitative and Quantitative Concordance

There was no association between the characteristics of the citation graphs and the observed qualitative or quantitative concordance in the 28 analyzable topics (Table 4). The table shows analyses in which the predictors (numbers or mean scores) were log transformed using the base 2. Thus, the odds ratios in the table represent the change in the odds of finding concordance per doubling of the value of the predictor. The p-values remain identical for any other logarithmic transformation of the predictors. Analyses using other transformations (square root, no transformation) were qualitatively similar. It appears that the examined connectivity metrics of the citations maps do not correlate with the likelihood that the summary findings from meta-analyses of epidemiological studies and of RCTs are in qualitative or quantitative agreement.

Table 4

Associations between concordance and quantitative characteristics of the citation graphs.