Abstract
A population-based study of a quantitative trait, such as Blood Pressure (BP), may be seriously compromised when the trait is subject to the effects of a treatment. Without appropriate corrections, this can lead to a considerable reduction of statistical power. Here we demonstrate this in the scenario of QTL mapping through Single-Marker Analysis. The data are simulated from a normal mixture for different values of allele frequencies, separation between normal populations, and Linkage Disequilibrium (LD). Several methods of correction are compared to determine which can best compensate for the loss of power if treatment effects are ignored. In one of these methods, underlying BPs are approximated by subtracting an estimate of the mean value of the medicine effect from observed BPs in treated subjects. We demonstrate the efficacy of this method across different choices of parameters. Finally, to account for quantitative traits that follow non-normal distributions, data are simulated from lognormal mixtures, and the Kruskal-Wallis test is used to obtain estimates of power for different methods of analysis.
Introduction
In genetic research, identifying Quantitative Trait Loci (QTLs) is essential for understanding the hereditary basis of continuous traits like blood pressure. However, medical interventions—such as antihypertensive treatments—often obscure these genetic signals. When a patient's BP is lowered by medication, the observed value no longer reflects their underlying genetic predisposition. This "phenotypic shrinkage" masks extreme values, significantly reducing the statistical power to detect genetic variants. This paper evaluates various statistical adjustment methods to determine which most effectively recovers this lost information.
Methods
The researchers employed a simulation-based framework to compare seven strategies for handling treatment effects across various genetic scenarios, adjusting parameters such as allele frequency ($p$), effect size ($d$), and Linkage Disequilibrium ($\delta'$):
- All underlying: The benchmark using true, unobserved genetic values.
- All observed: A naive analysis that ignores treatment status.
- Omit affected subjects: Excluding individuals with high BP (>140 mmHg).
- Omit treated subjects: Excluding all individuals currently on medication.
- Treatment as covariate: Including treatment status as a binary variable in the model.
- Constant adjustment: Adding the estimated average medicine effect back to the scores of treated individuals.
- Non-parametric adjustment: Using an iterative ranking method (the Levy algorithm).
Simulations were performed on both normal mixtures (standard traits) and lognormal mixtures (skewed traits) to test robustness.
Key Results and Comparative Analysis
The study reveals that common practices, such as excluding treated individuals or using treatment as a covariate, are often suboptimal. Removing treated subjects discards the most informative "extreme" data points, while covariate adjustment suffers from endogeneity—since treatment is a consequence of the trait, not an independent cause.
As shown in Figure 1, the Constant Adjustment method consistently maintains high statistical power, nearly matching the "Gold Standard" of underlying values. For instance, at a moderate genetic effect ($d=20$) and allele frequency $p=0.1$, the constant adjustment method preserves 97.3% power, whereas omitting treated subjects causes a drop to 89.9%.
Robustness to Non-Normal Distributions
To ensure applicability to traits that do not follow a standard bell curve, the authors tested lognormal mixtures using the Kruskal-Wallis test.
Even under conditions of reduced linkage ($\delta'=0.5$) and skewed distributions, the constant adjustment method (utilizing medians) remains the most effective strategy for preserving power (Figure 3). While overall power is lower due to the weaker genetic linkage, the relative superiority of constant adjustment over omission or naive analysis remains stable.
Conclusion
This research demonstrates that simple constant adjustment is a powerful tool for correcting treatment effects in genetic studies. By "undoing" the average effect of medication, researchers can reconstruct the phenotypic distribution and successfully map underlying QTLs. While real-world treatment effects are more variable than these simulations, this method provides a robust baseline for improving the accuracy of population-based genetic studies.
Summary updated