📊 statistics

Bounding Causal Effects for Ordinal Outcomes Under Positive Dependence

This paper addresses the challenge of estimating causal effects for ordinal outcomes by demonstrating that standard positive dependence assumptions are insufficient to justify tighter independence-based bounds, leading the authors to propose a new, stronger condition called diagonal tail dominance (DTD) and to derive improved local bounds that are validated through theoretical analysis, simulations, and a clinical trial application.

Original authors: Micha Mandel, Daniel Rodan

Published 2026-06-16

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Micha Mandel, Daniel Rodan

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

Imagine you are trying to decide if a new medicine works better than a placebo. In many medical studies, the results aren't simple "yes" or "no" numbers. Instead, they are ordinal outcomes—ranked categories like "Terrible," "Poor," "Fair," "Good," and "Excellent."

The problem is that while we know "Excellent" is better than "Good," we don't know how much better. Is the gap between "Fair" and "Good" the same as the gap between "Good" and "Excellent"? We don't know. Because of this, standard math tools (like calculating an average improvement) break down. You can't just average "Terrible" and "Excellent" to get a meaningful number.

This paper tackles a specific challenge: How can we prove a treatment helps, without making up arbitrary numbers or assuming we know exactly how the patients' outcomes are linked?

Here is the story of the paper, broken down with simple analogies.

1. The "Black Box" Problem

Imagine you have two groups of people.

Group A gets the new treatment.
Group B gets the standard care.

You can see what happened to Group A and what happened to Group B. But you cannot see the "secret twin" of every person. You don't know what would have happened to a specific person in Group A if they had been in Group B instead.

In statistics, we call this the Joint Distribution. It's a giant grid (or matrix) showing every possible combination of outcomes.

Top Left: People who would have been "Terrible" with treatment AND "Terrible" with control.
Bottom Right: People who would have been "Excellent" with treatment AND "Excellent" with control.
The Diagonal: People who would have been "Good" in both worlds.

The Catch: We can only see the margins of this grid (the totals for Group A and the totals for Group B). The inside of the grid is a black box. We don't know how the rows and columns line up.

2. The "Worst-Case" Guess (Model-Free Bounds)

Since we can't see the inside of the box, statisticians usually draw a "safety fence" around the truth. They calculate the absolute widest possible range where the truth could hide, assuming anything could be happening inside the box.

The paper shows that these safety fences are often huge. It's like trying to guess the temperature inside a house by only looking at the outside. You might say, "It's between -20°F and 100°F." That's technically true, but it doesn't help you decide whether to wear a coat.

3. The "Magic Assumption" (Independence)

To make the fence tighter, researchers often try to make a guess about how the inside of the box is arranged. A common guess is Independence.

The Analogy: Imagine shuffling two decks of cards (one for Treatment, one for Control) and dealing them out. If the decks are truly independent, getting a "King" in the Treatment deck has nothing to do with getting a "King" in the Control deck.

If we assume the decks are shuffled independently, we can calculate a much tighter, more useful number. For example, we might say, "At least 50% of people improved."

The Paper's Big Discovery:
The authors asked: "Is it safe to use this 'shuffled deck' (independence) math if we believe the treatment is helpful?"

They tested this against the standard rules of "Positive Dependence" (the idea that if you do well in one scenario, you likely do well in the other).

The Result: They found that even strong, standard rules of "doing well together" are NOT enough to guarantee that the "shuffled deck" math is correct.
The Metaphor: Imagine you have two friends, Alice and Bob. You know they are both generally happy people (Positive Dependence). But that doesn't mean their happiness is perfectly synchronized in a way that lets you use the "independent" math. Sometimes, the math says "Alice is happy 60% of the time," but the reality is she is happy 40% of the time. The "independent" guess was too optimistic.

4. The New "Diagonal Tail" Rule

Since the old rules didn't work, the authors invented a new, stricter rule called Diagonal Tail Dominance (DTD).

The Analogy:
Imagine the grid of outcomes again. The "Diagonal" is where a person gets the exact same rank in both worlds (e.g., "Fair" in both).

DTD says: If a person is at a specific rank (say, "Fair") in the Control world, they must be strictly better off (or at least as good) in the Treatment world more often than they would be if the decks were just shuffled randomly.

It's a very specific, strong condition. It's like saying: "If you are standing on step 3 of the stairs without the medicine, the medicine guarantees you are more likely to be on step 4 or higher than if you were just guessing."

The Verdict:

If this new rule holds, the "shuffled deck" math is safe and gives a tight, useful lower bound.
However, the authors admit this rule is very strong. It might not be true in many real-world situations. It's like requiring a specific, perfect alignment of stars to make a prediction work.

5. The "Local" Compromise

Since the "Global" rule (DTD) is too strict, the authors proposed a Local version.

The Analogy:
Instead of demanding the rule works for every single step of the ladder, we only demand it works for the top steps (the best outcomes).

Maybe we can't be sure the medicine helps people who are "Terrible" (Step 1).
But we can be reasonably sure that if someone is "Good" (Step 4), the medicine makes them "Excellent" (Step 5) more often than random chance would.

By applying the math only to these "safe zones" (the local tail), they can still get a tighter, more useful answer than the huge "worst-case" fence, without needing the impossible global rule.

6. The Real-World Test (Stroke Trial)

The authors tested this on real data from a stroke trial.

The Old Way (No assumptions): The answer was "Between 15% and 79% of people improved." (Useless).
The "Shuffled Deck" Way (Assuming Independence): The answer was "At least 48% improved." (Better, but risky if the assumption is wrong).
The New "Local" Way: By applying their new rule only to the better outcomes, they found a safe lower bound of 22%.

This is a win. It's not as high as the risky "48%" guess, but it is much better than the useless "15%" worst-case scenario. It gives doctors a concrete, defensible number to say, "We are confident at least 22% of patients benefited."

Summary

This paper is a cautionary tale for statisticians and a guide for researchers:

Don't trust standard "positive correlation" rules to justify using simple independence math for ranked data; they often fail.
We need a new, stricter rule (Diagonal Tail Dominance) to make that math safe.
If the strict rule is too hard to believe, use a "Local" version that only applies to the best outcomes to get a better answer than the worst-case scenario.

The paper provides the mathematical tools to draw a tighter, safer fence around the truth, provided you are willing to accept specific, well-defined conditions about how the treatment affects the "tails" (the extremes) of the outcome distribution.

Technical Summary: Bounding Causal Effects for Ordinal Outcomes Under Positive Dependence

Problem Statement
Causal inference for ordinal outcomes presents a unique challenge because standard average treatment effects (ATE), defined as $E\{Y(1) - Y(0)\}$ , are ill-defined when the outcome scale lacks meaningful metric distances. While alternative estimands based on the joint distribution of potential outcomes, such as $\tau = P\{Y(1) \ge Y(0)\}$ (the probability of not being worse off) and $\eta = P\{Y(1) > Y(0)\}$ (the probability of benefiting), are conceptually appealing, they are generally not identifiable from observational data or randomized trials without assumptions on the joint distribution. Existing model-free bounds, derived solely from marginal distributions, are often too wide to be practically informative. Previous work suggested that assuming positive dependence between potential outcomes could yield tighter, independence-based lower bounds; however, the specific dependence conditions required to guarantee the validity of these bounds were not rigorously established.

Methodology
The authors employ a potential outcomes framework where $(Y(0), Y(1))$ denotes the pair of potential outcomes with a joint distribution $P$ and known marginals $P_0$ and $P_1$ . The methodology proceeds through three main stages:

Geometric Visualization: The paper utilizes colored matrix representations of the joint probability table to visualize the relationship between the true estimands ( $\tau, \eta$ ) and their model-free bounds. This graphical approach clarifies which entries of the joint distribution are captured by the bounds and highlights the "gap" caused by unidentifiable off-diagonal probabilities.
Evaluation of Standard Dependence: The authors test whether standard notions of positive dependence—specifically Positive Quadrant Dependence (PQD) and Positive Regression Dependence (PRD)—are sufficient to justify using independence-based lower bounds ( $\tau_I, \eta_I$ ). Through numerical counterexamples, they demonstrate that even strong forms of PRD do not guarantee that $\tau_I \le \tau$ or $\eta_I \le \eta$ .
Introduction of New Dependence Conditions: To resolve the insufficiency of standard measures, the authors propose a new condition termed Diagonal Tail Dominance (DTD). They define:
- Open-tail DTD: $P(Y(1) > j \mid Y(0) = j) \ge P(Y(1) > j)$ for all $j$ .
- Closed-tail DTD: $P(Y(1) \ge j \mid Y(0) = j) \ge P(Y(1) \ge j)$ for all $j$ .
  They prove that Open-tail DTD is sufficient for $\eta_I$ to be a valid lower bound for $\eta$ , and Closed-tail DTD is sufficient for $\tau_I$ to be a valid lower bound for $\tau$ .
Local DTD and Improved Bounds: Recognizing that global DTD is a strong and potentially unrealistic assumption, the authors derive "local" versions of these conditions. They show that if DTD holds only for a subset of the support (local DTD), one can construct improved lower bounds that are tighter than model-free bounds but do not require the strong global assumption.

Key Contributions and Results

Refutation of Standard Assumptions: The paper provides counterexamples demonstrating that PQD and PRD are insufficient to justify the use of independence-based lower bounds for ordinal causal effects. This corrects a potential misconception in the literature regarding the validity of such bounds under general positive dependence.
New Sufficient Conditions: The authors introduce Diagonal Tail Dominance (DTD) as a necessary and sufficient condition (in the sense of sufficiency for the bounds) to guarantee that independence-based bounds are valid. They clarify that DTD is distinct from and not implied by standard correlation measures like Kendall's tau or Spearman's correlation.
Practical Bounds via Local DTD: The paper derives improved lower bounds (Equations 8 and 9) that exploit local DTD. These bounds allow researchers to utilize independence assumptions on specific parts of the probability table where the condition is plausible, yielding tighter intervals than model-free bounds without requiring the strong global DTD assumption.
Empirical Application: The methodology is applied to data from a clinical trial on acute ischemic stroke. The analysis shows that while model-free bounds for the proportion of patients benefiting from treatment ( $\eta$ ) are wide (0.155 to 0.790), assuming independence yields a much sharper lower bound (0.486). The authors argue that while global DTD is questionable for this data, local DTD for higher outcome levels is plausible, leading to improved bounds ( $\tilde{\eta} = 0.224$ ) that are still significantly tighter than the model-free lower bound.

Significance and Claims
The paper claims to provide both theoretical insight and practical tools for causal inference with ordinal outcomes. Its primary significance lies in clarifying the precise dependence structure required to tighten causal bounds. The authors emphasize that while standard "weak" dependence notions are insufficient, the proposed DTD conditions offer a rigorous justification for using independence-based bounds. However, they maintain a modest stance, acknowledging that DTD is a strong assumption that may not hold in many settings and is untestable from data alone. Consequently, the paper advocates for the use of local DTD as a more plausible compromise, offering improved bounds that are less restrictive than global independence assumptions but still stronger than model-free limits. The work also notes that these techniques extend naturally to continuous outcomes and can be adapted to settings with confounders by calculating covariate-specific bounds.