Abstract
Binary outcomes defined by logical (Boolean) "and" or "or" operations on original continuous and discrete outcomes arise commonly in medical diagnoses and epidemiological research. In this manuscript, we consider applying the “or” operator to two continuous variables above a threshold and a binary variable, a setting that occurs frequently in the modeling of hypertension. Rather than modeling the resulting composite outcome defined by the logical operator, we present a method that models the original outcomes thus utilizing all information in the data, yet continues to yield conclusions on the composite scale. A stratified propensity score adjustment is proposed to account for confounding variables. A Mantel-Haenszel style combination of strata-specific odds ratios is proposed to evaluate a risk factor. The benefits of the proposed approach include easy handling of missing data and the ability to estimate the correlations between the original outcomes. We emphasize that the model retains the ability to evaluate odds ratios on the simpler and more easily interpreted composite scale. The approach is evaluated by Monte Carlo simulations. An example of the analysis of the impact of sleep disordered breathing on a standard composite hypertension measure, based on blood pressure measurements and medication usage, is included.
Introduction
In clinical research and epidemiology, many primary outcomes are "composite" or "Boolean" in nature. A classic example is the diagnosis of hypertension, which is typically defined as having a systolic blood pressure (SBP) $\ge 140$ mmHg, OR a diastolic blood pressure (DBP) $\ge 90$ mmHg, OR currently taking anti-hypertensive medication.
The standard analytical approach involves collapsing these disparate pieces of information—two continuous measurements and one binary flag—into a single binary variable for use in logistic regression. However, this method is inherently "lossy." First, thresholding continuous data discards significant variance and nuance within the "normal" and "high" ranges. Second, it handles missing data poorly; if a single component is missing (e.g., a missing DBP record), the entire Boolean outcome becomes uncertain, often necessitating the exclusion of that subject from the analysis. This paper proposes a joint modeling framework that utilizes all component information while still providing results on the familiar composite scale.
Methodology
The proposed framework employs a Joint Modeling approach within a multivariate normal system.
The Multivariate Model
The model treats the two continuous variables (SBP and DBP) and a latent variable representing the propensity to take medication as a system of equations. These equations are linked by shared random intercepts, capturing the underlying correlations between outcomes for each individual. For the binary medication component, a latent probit structure is used, assuming an underlying continuous "urge" or "need" for medication that triggers a positive response once a specific threshold is crossed.
Adjustment and Weighting
To address confounding variables in observational data—such as age, BMI, or smoking status—the authors utilize Propensity Score Stratification. Subjects are grouped into five strata based on their propensity scores, balancing covariates between high-risk and low-risk groups. Finally, a Mantel-Haenszel weighting technique combines the results from these strata into a single, adjusted odds ratio for the Boolean outcome.
Data and Evaluation
The method was evaluated using both synthetic and real-world data:
- Monte Carlo Simulations: Synthetic datasets of 1,000 subjects were generated to compare the joint model's parameter recovery against standard logistic regression.
- Sleep Heart Health Study (SHHS): This real-world dataset included 5,015 subjects. The analysis focused on the relationship between the Respiratory Disturbance Index (RDI)—a marker for sleep apnea severity—and the composite hypertension outcome.
Results
Simulation Performance
Simulations demonstrated that the joint model provided superior 95% confidence interval coverage compared to the standard Boolean model. The standard approach tended to be overly conservative or biased when confounding was present, whereas the joint model remained robust.
Sleep Heart Health Study Findings
The unadjusted odds ratio for hypertension (comparing high RDI to low RDI) was 1.64. However, after adjusting for confounders using the joint model, the odds ratio decreased to 1.13 (95% CI: 0.96–1.33). This suggests that much of the apparent link between sleep apnea and hypertension may be mediated by other factors like age and BMI.
Furthermore, the joint model allowed for "deconstructing" the effect. It revealed that high RDI was significantly associated with medication usage but not strongly associated with elevated SBP or DBP at the time of the exam. The model also estimated the correlation between SBP and DBP at 0.41, while finding that DBP and medication usage were virtually uncorrelated ($5.7 \times 10^{-7}$).
Discussion
The joint modeling approach offers significant advantages in statistical efficiency and interpretability. By modeling components directly, the framework is resilient to missing data, using correlations between known variables to inform missing values. It also provides a more granular view of how risk factors influence specific disease components.
Limitations include the current assumption that data is "Missing Completely at Random" (MCAR), which may not hold in all clinical settings. Future work should explore "Missing at Random" (MAR) assumptions and extend the model to longitudinal data to track the evolution of these component relationships over time.
Summary updated