📊 statistics

Causal Inference for Case Studies in Behavioral Health

This paper introduces a causal inference framework for behavioral health case studies that utilizes $\Omega$ estimands based on outcome supports rather than distributions, enabling valid causal conclusions under unmeasured confounding with only a positivity assumption.

Original authors: Shane Sparkes

Published 2026-07-09

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Shane Sparkes

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Problem: The "Time Travel" Dilemma

Imagine you are a therapist trying to figure out if a specific treatment helped a client. In a perfect world, you could use a "time machine" to see what would have happened to that exact same person if they hadn't received the treatment. You could compare their "treated" self with their "untreated" self.

But, as the paper points out, time travel doesn't exist. You can't rewind the clock. Furthermore, in real life, you can't measure everything. There are always hidden factors (like a client's bad day at work, a family argument, or the weather) that might influence both the treatment they get and how they feel. In statistics, these hidden factors are called confounders. Usually, to prove cause and effect, you have to measure and control for all these hidden factors. But in behavioral health, it's often impossible or even unethical to measure everything without ruining the natural flow of care.

The Solution: Looking at the "Menu" Instead of the "Order"

The author, Shane Sparkes, proposes a new way to solve this. Instead of trying to calculate the average improvement (which requires knowing all the hidden probabilities), he suggests looking at the range of possibilities.

The Analogy: The Restaurant Menu
Imagine a client's anxiety level is like a menu of possible feelings, ranging from "Calm" (1) to "Panic" (10).

Before treatment: The client's "menu" of possible feelings includes everything from 1 to 10. They might feel a 10 today, a 7 tomorrow, or a 2 next week.
After treatment: If the treatment works, the client might still feel a 1, 2, or 3, but they might never feel a 9 or 10 again. The "9" and "10" have been removed from the menu.

The paper argues that if the treatment changes the menu (the set of possible outcomes), we can say it had an effect, even if we don't know the exact odds of feeling a 7 versus an 8.

The Core Idea: $\Omega$ -Estimands

The paper introduces a fancy term called $\Omega$ -estimands (pronounced "Omega-estimands").

Traditional methods look at the distribution (the probability of every number). They ask: "What is the average score?"
This paper's method looks at the support (the list of numbers that are actually possible). It asks: "What numbers are on the menu?"

The author claims that if you look at the menu (the support) rather than the odds (the distribution), you don't need to know the hidden confounders. As long as a basic condition called Positivity is met, the "menu" you see in the real world is the same "menu" you would have seen in a perfect experiment.

What is Positivity?
Think of Positivity as a rule of "open options." It means that no matter what hidden situation a client is in (bad day, good day, crisis), the treatment provider still has the possibility of offering the treatment. If a crisis is so severe that the provider cannot offer the usual treatment, Positivity is broken. But if the provider can always offer the treatment (even if they adjust how they do it), the "menu" remains valid for comparison.

The Three "Rules" for the Method

To make this work in a real clinic, the paper suggests three conditions:

Positivity (The Open Door Rule): As mentioned above, the treatment must be possible in all situations. The "menu" of treatments shouldn't disappear just because the client's life got complicated.
Epistemic Preservation (The "Same Shape" Rule): This is a bit technical, but it basically means that the person making the judgment (the therapist or client) should use a consistent way of thinking about uncertainty. If they guess the "average" feeling before treatment, they should guess the "average" after treatment using the same mental logic.
Support Recall (The "Memory Menu" Rule): Usually, we don't have data for the "before" period because the client wasn't being measured yet. The paper suggests asking the client to remember their baseline. The rule here is: Even if the client's memory isn't perfect, as long as they remember the range of possibilities (e.g., "I used to feel between 8 and 10"), that memory is good enough to compare against the current "menu."

The Case Study: A Simple Example

The paper uses a real example of a 23-year-old student with anxiety.

Before treatment: Their anxiety scores were high (around 10).
After 19 sessions: Their scores dropped, and the lowest score they ever hit was a 4.
The Calculation: The author uses a simple math trick (Bayesian reasoning) to say: "Given that the student started at 10 and dropped to a low of 4, there is a strong belief that the treatment worked."
The Result: They calculated a probability (about 59%) that the treatment was effective. This isn't a "proof" in the scientific sense, but it gives the therapist and client a coherent, logical reason to believe the treatment helped, based on the fact that the "menu" of possible feelings changed.

The Bottom Line

This paper claims that we can prove a treatment works in a single person's life (an N=1 study) without needing to measure every hidden factor in their life.

Old way: "We need to measure everything to be sure." (Often impossible).
New way: "If the treatment removes the worst possibilities from the client's life (changes the menu), and the treatment was always an option (Positivity), then we have evidence of a causal effect."

The author admits this method is a trade-off: it is very robust against hidden errors (confounders) but might miss subtle changes if the "menu" stays the same but the frequency of feelings changes. However, for behavioral health, where eliminating the worst symptoms (the "9s and 10s") is often the goal, this method offers a practical, logical path forward for therapists and clients to understand their progress.

Technical Summary: Causal Inference for Case Studies in Behavioral Health

Problem Statement
The paper addresses the fundamental challenge of conducting causal inference in observational $N=1$ settings, specifically within behavioral health case studies. Standard causal inference methodologies typically rely on randomization, experimental control, or large population frames to adjust for confounding. However, individual care processes lack these features. Providers and clients operate in a context of "unmeasured confounding" ( $U$ ), where a web of latent variables influences both treatment selection ( $T$ ) and outcomes ( $Y$ ). Furthermore, strict experimental designs (e.g., A-B-A-B reversals) are often unethical or subversive to the therapeutic relationship. Consequently, providers cannot empirically measure counterfactuals (e.g., $Y|do(t=0)$ when $Y|do(t=1)$ is observed) or rely on population-based matching assumptions that presuppose an "essential identity" between distinct time points or individuals. The core problem is establishing a valid route to causal estimands without requiring the measurement or adjustment of confounders.

Methodology and Framework
The authors propose a framework based on Structural Causal Models (SCMs) and a subjectivist (de Finetti) interpretation of probability, where probability is viewed as a personal measure of uncertainty rather than an objective frequency.

$\Omega$ -Estimands: The paper shifts the focus from contrasts of probability distributions (e.g., differences in means) to contrasts of supports. The support of a random variable, denoted as $\Omega$ , is the set of values with non-zero probability. The authors define $\Omega$ -estimands as functions of these supports (e.g., $\max(\Omega)$ , $\min(\Omega)$ , range, or cardinality) that do not incorporate probability weights.
The Core Assumption (Positivity): The framework relies on a single, critical assumption: Positivity. Formally, for all treatment values $t \in \Omega_T$ and unmeasured confounders $u \in \Omega_U$ , the conditional probability $f(t|u) > 0$ . In the subjectivist context, this implies that no state of the unmeasured confounder $u$ renders a specific treatment $t$ impossible. The authors argue that if treatment is conceptualized at a sufficiently coarse level (e.g., "exposure to case management" rather than specific micro-interventions), positivity holds even in the presence of unmeasured confounders.
Support Coincidence: The central theoretical result (Proposition 1) proves that under the positivity assumption, the observational support of an outcome coincides with its interventional support: $\Omega_{Y|t} = \Omega_{Y|do(t)}$ . This means that the set of possible outcomes observed in the real world is identical to the set of possible outcomes that would occur under a hypothetical intervention, rendering the identification of $\Omega$ -effects immune to unmeasured confounding.

Key Contributions and Extensions
The paper introduces three conditions to facilitate inference:

Condition I (Positivity): Sufficient to identify that observational and interventional supports coincide.
Condition II (Epistemic Preservation): Assumes that the "structural motif" of the agent's uncertainty (specifically the relationship between the expected value and the functional average of the support) remains invariant across treatment and non-treatment states. This allows the translation of $\Omega$ -effects into differences in expected values via an expected-average identity: $E(Y) = Av(Y) - \kappa$ , where $\kappa$ is a skew-related term. If $\kappa$ is preserved, differences in means acquire causal meaning.
Condition III (Support Recall): Allows a client's recalled baseline estimate ( $\hat{Y}$ ) to serve as a proxy for the unobserved baseline support ( $\Omega_{Y|t=0}$ ). It assumes that while the probability distribution of the recall may differ, the set of possible values (the support) remains anchored to the same reality.

Results and Case Study Application
The authors demonstrate the framework using a clinical case study of a 23-year-old student undergoing Cognitive Behavioral Therapy (CBT) for anxiety.

Data: 19 weekly anxiety scores (0–10 scale) showing a decreasing trend from 10 to 4.
Application: The authors define an $\Omega$ -effect as the contrast between the minimum of the baseline support and the minimum of the treatment support ( $\min(\Omega_{Y_0}) - \min(\Omega_{Y^*})$ ).
Inference: Using a subjective Bayesian approach, they calculate the posterior probability of treatment efficacy. By positing a prior belief and updating it with the observed data (the shift in the minimum score), they derive a posterior probability of efficacy ( $\approx 0.59$ ).
Interpretation: The result is not a frequency-based guarantee but a "coherent, tempered credence" regarding the treatment's ability to eliminate high-intensity symptom possibilities (a shift in support).

Significance and Claims
The paper claims to provide a "route to causal inference" for individual care processes that bypasses the traditional requirement of closing backdoor paths or measuring confounders.

Robustness over Sensitivity: The framework trades sensitivity to distributional changes (which are confounded) for robustness by focusing on support changes (which are identified under positivity). It argues that effective treatment often manifests as the elimination of extreme possibilities (e.g., preventing a panic attack), which is a change in support detectable without knowing confounders.
Practical Utility: It offers a principled, elementary method for providers to reason about treatment efficacy using measurement-based care data without advanced statistical training or experimental designs.
Limitations: The authors acknowledge that positivity is untestable and must be defended via the validity of the underlying Directed Acyclic Graph (DAG). They also note that the framework is insensitive to treatments that only shift probability mass within an unchanged support (e.g., making a "5" more likely than a "4" without eliminating the "10"). Furthermore, the results are coherence statements for a reasoning agent, not frequency guarantees.

In summary, the paper argues that by redefining causal targets as contrasts of outcome supports rather than distributions, and by accepting a subjectivist view of probability, one can achieve valid causal identification in $N=1$ behavioral health studies under the sole condition of positivity.