Dietary exposure-outcome associations do not systematically differ by risk of bias in observational studies: a...

To investigate the influence of bias from methodological characteristics on exposure-outcome associations in nutrition observational studies using the ROBINS-E tool.

Prospective observational studies published between 1987 and 2018 were selected from a representative sample of 183 nutrition meta-analyses. Two reviewers conducted data extraction and risk of bias (RoB) assessments independently. The average difference in effect estimates attributable to bias was quantified using the ratio of risk ratio (RRR), comparing observational studies rated as "very high" or "high risk" to those rated as "some concerns" Differences across RoB domains, types of exposure, and outcomes were explored in subgroup analyses. The between-study variance was quantified using the heterogeneity estimator τˆ² and measured with I² statistic and 95% prediction intervals (PIs).

We included 77 prospective observational studies with 155 outcome-specific RoB assessments: 52.9% were rated as "high risk" 44.5% as "some concerns" and 2.6% as "very high risk" of bias. Overall RoB did not affect diet-disease associations (RRR 1.00, 95% confidence interval [CI] 0.94-1.06; I² = 36%; τˆ² = 0.004; PI 0.86-1.16). Individual RoB domains showed no differences, except for bias arising from measurement of the exposure (RRR 1.14, 95% CI, 1.01-1.28; I² = 45%; τˆ² = 0.02; PI 0.82-1.58) and bias due to missing data (RRR 0.87, 95% CI 0.77-0.98; I² = 64%; τˆ² = 0.04; PI 0.56-1.35). Findings were generally supported by subgroup analyses. However, the finding for missing data was not robust in meta-regression analyses accounting for non-independence.

Diet-disease associations showed no substantial differences across RoB ratings, despite most included studies being rated as "high risk," primarily due to confounding. However, since none of the studies was rated as "low risk" with ROBINS-E, these findings should be interpreted with caution and confirmed in larger samples including studies with lower RoB.

In nutrition research, observational studies are commonly used to explore how different foods, dietary pattern, or supplements may be linked to health. However, limitations in how studies are planned, implemented, or carried out can make these relationships appear stronger or weaker than they really are. This problem is known as bias. The ROBINS-E ("Risk Of Bias In Non-randomized Studies - of Exposures") tool provides researchers with a structured way to assess whether bias is present and how much it may affect the associations between exposures and health. In this meta-epidemiological study, we looked at 77 observational studies of dietary exposures published up to 2018. We assessed 155 outcome-specific risk of bias assessments to quantify the potential for bias, categorizing them into four levels: low risk, some concerns, high risk, and very high risk. Then, we compared the results of studies with higher versus lower risk of bias levels to investigate whether bias influenced the reported associations between dietary exposures and health. Most studies were rated as high risk of bias or with some concerns, mainly because of confounding. This means that other factors-such as education, physical activity, or smoking habits-can influence both what people eat and their health outcomes, making it difficult to know whether the observed effect is truly due to diet. None of the included studies were rated as low risk of bias. Based on our analysis, bias had no substantial impact on the results within studies rated as high risk or with some concerns. However, two specific sources of bias showed differences: studies with concerns about how diet was measured tended to report larger effects and so did studies in which many participants did not provide complete data on their diet and health outcomes. After we accounted for overlap between studies, the finding for missing data was less certain. Our findings suggest that many common sources of bias in nutrition observational studies may not substantially affect the observed associations between diet and health. However, because none of the studies were rated as low risk of bias, these findings should be interpreted cautiously. More research is needed to confirm these findings.