In Epigenomic Studies, Including Cell-Type Adjustments in Regression Models Can Introduce Multicollinearity, Resulting in Apparent Reversal of Direction of Association

Background: Association studies of epigenome-wide DNA methylation and disease can inform biological mechanisms. DNA methylation is often measured in peripheral blood, with heterogeneous cell types with different methylation profiles. Influences such as adiposity-associated inflammation can change cell-type proportions, altering measured blood methylation levels. To determine whether associations between loci-specific methylation and outcomes result from cellular heterogeneity, many studies adjust for estimated blood cell proportions, but high correlations between methylation and cell-type proportions could violate the statistical assumption of no multicollinearity. We examined these assumptions in a population-based study. Methods: CDKN2A promoter CpG methylation was measured in peripheral blood from 812 adolescents aged 17 years (Western Australian Pregnancy Cohort Study). Log(e) adolescent BMI was used as the outcome in a regression analysis with DNA methylation as predictor, adjusting for age/sex. Further regression analyses additionally adjusted for estimated cell-type proportions using the reference-based Houseman method, and simulations modeled the effects of varying levels of correlation between cell proportions and methylation. Correlations between estimated cell proportions and CpG methylation from Illumina 450K were measured. Results: Lower DNA methylation was associated with higher BMI when cell-type adjustment was not included; for CpG4, β = −0.004 log(e)BMI/%methylation (95% CI −0.0065, −0.001; p = 0.003). The direction of association reversed when adjustment for six cell types was made; for CpG4, β = 0.004 log(e)BMI/%methylation (−0.0002, 0.0089; p = 0.06). Correlations between CpG methylation and cell-type proportions were high, and variance inflation factors (VIFs) were extremely high (25 to 113.7). Granulocyte count was correlated with BMI, and removing granulocytes from the regression model reduced all VIFs to <3.1, with persistence of a positive association between methylation and BMI [CpG4 β = 0.004 log(e)BMI/%methylation (−0.0002, 0.0088; p = 0.06)]. Simulations supported major effects of multicollinearity on regression results. Conclusions: Where cell types are highly correlated with other covariates in regression models, the statistical assumption of no multicollinearity may be violated. This can result in reversal of direction of association, particularly when examining associations with phenotypes related to inflammation, as CpG methylation may associate with changes in cell-type proportions. Removing predictors with high correlations from regression models may remove the multicollinearity. However, this might hinder biological interpretability.