(P11-012-24) Arguments for Analyzing Intake of Total Fat and Carbohydrates As a Log-Ratio Instead of As Separate Predictor Variables: An Empirical Methods Comparison
Postdoctoral Researcher Karolinska Institutet Enskede, Stockholms Lan, Sweden
Disclosure(s):
Jakob Norgren, PhD: No relevant financial relationship(s) with ineligible companies to disclose.
Objectives: Intake of carbohydrates (eCarb) and total fat (eFat) as proportions of total energy intake (E%) constitute compositional data—summing up to 100% with protein and other energy sources. There are restrictions on how compositional data may be handled in regression models—due to collinearity—and the use of log-ratios is an established method to overcome this. The objective was to compare the log-ratio eCarb/eFat (CFr) with eCarb and eFat separately as predictor variables of a masked continuous health outcome (Y).
Methods: Baseline data from a clinical trial (FINGER) was used (n=1259). Distributions of eCarb, eFat, protein, and alcohol were graphically compared by boxplots of the extreme quintiles (Q1/Q5) of eCarb, eFat, and CFr. The effect on Y of eCarb, eFat, and CFr respectively, was analyzed in linear regression with and without adjustment for protein and alcohol. The linear relations between CFr, eCarb, and eFat were investigated graphically and in crude regression.
Results: The distributions of eCarb and eFat were almost identical in the extreme quintiles of CFr, eCarb, and (inversely) eFat—implying that eCarb and eFat are in fact proxies for CFr. However, outliers with high/low protein and alcohol intake accumulated in Q1/Q5 for eCarb and eFat, while being relatively evenly distributed between quintiles for CFr. The relation between β-coefficients was eCarb< CFr< eFat for the effect on Y, but after adjustment for protein and alcohol, eCarb and eFat aligned close to CFr—which was unaffected. One SD increase in CFr was associated with a reciprocal change in eCarb (6.31,95% CI: 6.20, 6.41 E%) and eFat (−6.26, CI: −6.44, −6.09 E%).
Conclusions: Within this dataset, CFr, eCarb, and eFat practically represented the same (one-dimensional) compositional parameter—but CFr was least biased by other energy sources. Since the distribution of protein has a relatively low and narrow distribution in typical epidemiological datasets, generalizability is likely to be high. Accepting that eCarb and eFat are proxies for CFr may have implications for the interpretation of “substitution analyses” including those variables. I suggest that CFr should be considered as the standard variable to report—approximately representing iso-caloric exchange between eCarb and eFat.
Funding Sources: Alzheimerfonden, af Jochnick Foundation