[Paper Review] Grace periods in comparative effectiveness studies of sustained treatments

Kerollos Nashat Wanis, Aaron L. Sarvet+5 more|arXiv (Cornell University)|Dec 21, 2022

Advanced Causal Inference Techniques32 references4 citations

TL;DR

This paper proposes stochastic grace period strategies—flexible, investigator-specified treatment protocols allowing temporary non-adherence—to improve causal inference in comparative effectiveness studies of sustained treatments. By enabling consistent estimation under realistic adherence patterns, the method identifies treatment effects independent of natural adherence differences, with data showing thiazide diuretics reduce 3-year cardiovascular risk more than ACE inhibitors, especially at 50% adherence during grace periods.

ABSTRACT

Researchers are often interested in estimating the effect of sustained use of a treatment on a health outcome. However, adherence to strict treatment protocols can be challenging for individuals in practice and, when non-adherence is expected, estimates of the effect of sustained use may not be useful for decision making. As an alternative, more relaxed treatment protocols which allow for periods of time off treatment (i.e. grace periods) have been considered in pragmatic randomized trials and observational studies. In this article, we consider the interpretation, identification, and estimation of treatment strategies which include grace periods. We contrast natural grace period strategies which allow individuals the flexibility to take treatment as they would naturally do, with stochastic grace period strategies in which the investigator specifies the distribution of treatment utilization. We estimate the effect of initiation of a thiazide diuretic or an angiotensin-converting enzyme inhibitor in hypertensive individuals under various strategies which include grace periods.

Motivation & Objective

• Address the limitations of strict deterministic treatment strategies that require continuous adherence, which are often unrealistic in clinical practice.
• Overcome the issue of positivity violations in causal inference due to non-adherence in real-world treatment settings.
• Develop a framework for estimating causal effects under grace period strategies that balance realism and identifiability.
• Compare natural grace period strategies (based on observed behavior) with stochastic grace period strategies (defined by investigators) to assess their impact on treatment effect interpretation.
• Provide a method for estimating treatment effects under various adherence levels during grace periods, enhancing the transportability and clinical relevance of results.

Proposed method

• Define stochastic grace period strategies as investigator-specified rules that allow temporary non-adherence, with adherence levels governed by a known distribution.
• Use Single World Intervention Graphs (SWIGs) to reason about identification assumptions under grace period strategies, distinguishing between natural and stochastic rules.
• Apply an estimator based on the efficient influence function to allow flexible machine learning for confounding adjustment in marginal structural models.
• Ensure identification by verifying exchangeability and positivity under the specified grace period strategy, with positivity guaranteed under natural grace periods.
• Implement estimation using data from Kaiser Permanente Colorado on antihypertensive monotherapies, comparing thiazide diuretics and ACE inhibitors under multiple grace period scenarios.
• Model grace periods with varying durations and adherence levels (e.g., 50% adherence) to assess sensitivity of risk estimates to adherence patterns.

Experimental results

Research questions

• RQ1 How can grace period strategies be defined to improve the realism and clinical relevance of comparative effectiveness studies?
• RQ2 What are the identification conditions for causal effects under stochastic grace period strategies, and how do they differ from those under natural grace period strategies?
• RQ3 To what extent do differences in observed adherence patterns during grace periods bias comparisons of treatment effectiveness between medications?
• RQ4 How do risk estimates for cardiovascular outcomes vary across different adherence levels during grace periods in antihypertensive treatment?
• RQ5 Can stochastic grace period strategies enable more reliable comparisons of 'pharmacological' effectiveness by decoupling outcome differences from adherence behavior?

Key findings

• Thiazide diuretics were associated with lower 3-year risk of acute myocardial infarction, heart failure, and stroke compared to ACE inhibitors across all grace period strategies.
• The greatest risk difference between thiazide diuretics and ACE inhibitors was observed when adherence during grace periods was set at 50%, indicating that risk differences are sensitive to adherence assumptions.
• Natural grace period strategies can produce misleading results if differences in effectiveness stem from variations in natural adherence rather than pharmacological efficacy.
• Stochastic grace period strategies ensure that adherence distributions are identical across treatment groups, enabling more reliable comparisons of pharmacological effectiveness.
• Positivity is guaranteed under natural grace period strategies for all observed histories within grace periods, but not necessarily under stochastic strategies.
• The proposed method using efficient influence function estimators with machine learning allows robust confounding adjustment and consistent estimation under complex grace period rules.