Testing the carbohydrate-insulin model of obesity in a 5-month feeding study: the perils of post-hoc participant exclusions

Abstract

A large feeding study reported that total energy expenditure (TEE) was greater on a low- versus high-carbohydrate diet, supporting the carbohydrate-insulin model of obesity. Recently, the validity of this finding was challenged in a post-hoc analysis excluding participants with putative non-adherence to the study diets. Here, we show why that analysis, based on a post-randomization variable linked to the outcome, introduced severe confounding bias. With control for confounding, the diet effect on TEE remained strong in a reanalysis. Together with sensitivity analyses demonstrating robustness to plausible levels of non-adherence, these data provide experimental support for a potentially novel metabolic effect of macronutrients that might inform the design of more effective obesity treatment.

Conflict of interest statement

DSL and CBE have conducted research studies examining the carbohydrate-insulin model of obesity funded by the National Institutes of Health and philanthropic organizations unaffiliated with the food industry; DSL received royalties for books on obesity and nutrition that recommend a low-glycemic load diet. No other author has relevant disclosures.

Figures

Fig. 1. Confounding arising from post-randomization participant exclusion on TEE effect size estimates.

TEE data from Ebbeling et al. [4] comparing Low- (20%) vs High- (60%) carbohydrate diets in Intention-to-Treat analyses. a Sequential elimination of 50% of participants based on “Unaccounted Energy” (UE, right to left, 1st to 18th of 36 quantiles) suggests a 42% attenuation in effect size. Figure modified from Hall and Guo in accordance with license CC0 https://www.biorxiv.org/content/10.1101/476655v5. The regression statistics were deleted because conditions for regression are not satisfied. (Individual points in this exclusion analysis are not independent of each other. Dashed line should be disregarded for the same reason.) b Exclusion analysis performed with final dietary intake and body composition data [7], with adjustment for potential baseline confounders as described in Methods. The results, expressed as a proportion of the diet effect present in the full cohort, indicate a lesser degree of effect attenuation (9%, or approximately one fifth of that observed by Hall and Guo). Qualitatively similar findings (i.e., substantially reduced effect attenuation) were obtained in a Per Protocol analysis (n = 104, data not shown).