Distributional Change in Ordinal Data with Missing Observations: Minimal Mobility and Partial Identification

This paper proposes a robust framework for measuring and interpreting distributional changes in ordinal data with missing observations by utilizing optimal transport to define minimal-mobility configurations and employing partial identification to construct sharp bounds on discrepancy measures and mobility patterns.

The Gist

The Big Picture: The "Mystery of the Moving Crowd"

Imagine you are a detective trying to figure out how a crowd of people changed their minds over time.

The Problem:
You have two snapshots of a crowd, taken one year apart.

• Snapshot A (Year 1): You know exactly how many people are standing in the "Happy," "Okay," and "Sad" zones.
• Snapshot B (Year 2): You know the numbers for those same zones again.

The Mystery:
You don't have a video of the people moving. You don't know who moved from "Happy" to "Sad." Did a few people jump all the way across the room? Did everyone take one small step? Or did the crowd just shuffle slightly?

Because you only have the "before" and "after" counts (the marginal distributions) and not the individual paths, you can't know the exact truth. This is a common problem in surveys where people drop out or refuse to answer questions (missing data).

The Solution: The "Minimal Effort" Rule

The author, Rami Tabri, proposes a clever way to solve this mystery without needing the missing video footage. He asks a simple question:

"What is the least amount of walking required to turn the crowd in Snapshot A into the crowd in Snapshot B?"

He calls this "Minimal Mobility."

The Analogy: The Heavy Boxes

Imagine the people in the "Happy" zone are heavy boxes. You need to move them to match the arrangement in Snapshot B.

• Moving a box from "Happy" to "Okay" takes 1 unit of effort.
• Moving a box from "Happy" to "Sad" takes 2 units of effort.

The paper uses a mathematical tool called Optimal Transport (think of it as the world's most efficient logistics planner) to find the arrangement that uses the minimum total effort.

• The Result: This gives you a measure of movement.
  • If you have complete data: The math gives you a single, precise number (a point estimate) for the minimum movement.
  • If you have missing data: The math cannot give you one single number. Instead, it provides a range (an interval). It tells you the movement lies somewhere between a "best case" and a "worst case."

The Twist: The "Missing People" (Partial Identification)

In real life, surveys are messy. Some people didn't answer the question.

• In Snapshot A, maybe 10% of people said "I don't know."
• In Snapshot B, maybe 15% said "I don't know."

We don't know where those missing people would have stood if they had answered. This creates a range of possibilities.

The Paper's Fix:
Instead of guessing one single answer, the author draws a safety net (a range of bounds).

1. Best Case: We assume the missing people were positioned in a way that made the crowd look as similar as possible. This gives the lowest possible movement score.
2. Worst Case: We assume the missing people were positioned to make the crowd look as different as possible. This gives the highest possible movement score.

The result isn't always a single number, but a range (e.g., "The crowd moved between 4% and 10%"). This is called Partial Identification. It's honest: it admits we don't know the exact truth, but it tells us exactly how much we can know.

Note: These bounds characterize the extremal movement across categories (how much people shifted from one opinion to another), which is distinct from measuring extremal dependence between variables.

The "Blueprints" (Minimal-Mobility Configurations)

The paper doesn't just give you a number; it describes the structure of how the move must have happened.

Imagine the "Minimal Effort" plan looks like this:

• "To get from Snapshot A to B with the least walking, at least 50 people must move from 'Happy' to 'Okay'."
• "No one needs to jump from 'Happy' to 'Sad' directly."

This isn't a single, unique blueprint. Instead, it represents a set of feasible plans. Any explanation of the data that claims "minimal movement" must look like one of the plans in this set.

• If someone claims, "The crowd didn't move at all!" you can say, "No, the math says at least 50 people had to shift."
• If someone claims, "Everyone jumped from 'Happy' to 'Sad'!" you can say, "That's possible, but it would require way more effort than the minimum. The data doesn't force that to happen, but it does force the smaller shifts."

This serves as an interpretive benchmark. It logically implies the minimum movement required by the data, rather than imposing a normative rule on how people should behave.

The Real-World Test: Arab Barometer

The author tested this on real data about how people in Iraq and Morocco felt about the United States over two years.

• The Question: Did people's opinions change?
• The Missing Data: Some people refused to answer.

The Findings:

1. Change Happened: Even in the "best case" scenario (where missing people are assumed to be similar), a significant chunk of the population (about 4% to 12%) had to change their answer to make the numbers match.
2. How they changed: The "structure" showed that people mostly took small steps (e.g., from "Very Favorable" to "Somewhat Favorable"). They didn't make giant jumps (e.g., from "Very Favorable" to "Very Unfavorable").
3. Robustness: Even when accounting for the missing people, the pattern of movement stayed the same. The missing data changed how much moved, but not how they moved.

Summary: Why This Matters

Think of this paper as a new kind of X-ray for social science.

• Old Way: Look at the before and after pictures and guess what happened. "Oh, the numbers changed, so people must have changed their minds!" (But maybe they didn't; maybe the missing data skewed it).
• New Way (This Paper): Use the "Minimal Effort" rule to calculate the absolute minimum change required.
  • It tells you the floor: "At least this much change happened" (either as a precise number or a safe range).
  • It tells you the structure: "It happened mostly through small steps, not giant leaps."
  • It handles missing data by giving you a safe range instead of a fake precise number.

It turns a vague comparison of numbers into a concrete, logical story about how a population must have shifted, even when we are missing pieces of the puzzle.

Technical Summary

Here is the revised technical summary of the paper "Distributional Change in Ordinal Data with Missing Observations: Minimal Mobility and Partial Identification" by Rami V. Tabri.

1. Problem Statement

The paper addresses a fundamental challenge in empirical economics and social science: measuring and interpreting distributional change in ordinal variables when only repeated cross-sectional data is available.

• The Identification Gap: In repeated cross-sections, researchers observe marginal distributions (e.g., the share of people in each category at time $t$ and $t+1$) but lack the joint distribution (the transition matrix showing how individuals moved between categories). Without joint data, the specific mechanism of change is unidentified.
• The Missing Data Problem: Furthermore, real-world datasets often suffer from pervasive missing observations (non-response), meaning the marginal distributions themselves are only partially observed.
• The Core Question: Given only marginal distributions (which may be fully observed or partially identified due to missing data), what is the minimal amount of probability mass reallocation required to transform one distribution into another, and what is the structure of this minimal change?

2. Methodology

The paper employs a framework combining Optimal Transport (OT) theory with Partial Identification methods.

A. Optimal Transport Representation (The Benchmark)

The author defines the measure of distributional change, $D(\mu, \nu)$, as the $L_1$ distance between the cumulative distribution functions (CDFs) of two distributions $\mu$ and $\nu$:
$$D(\mu, \nu) = \sum_{k=1}^{K-1} |F_\mu(k) - F_\nu(k)|$$
Using Proposition 1, the paper establishes that this scalar measure is equivalent to the Wasserstein-1 distance (Earth Mover's Distance) on the ordinal space $\{1, \dots, K\}$ with cost function $c(i,j) = |i-j|$.

• Interpretation: $D(\mu, \nu)$ represents the minimal cost (or minimal number of threshold crossings) required to reallocate probability mass from $\mu$ to $\nu$.
• Minimal-Mobility Configurations: The set of joint distributions (couplings) $\Pi^*(\mu, \nu)$ that achieve this minimum cost constitutes a feasible set of minimal-mobility configurations. These are not estimates of a unique true transition matrix, but rather the set of all joint distributions consistent with the observed marginals that minimize movement.
• Contribution Clarification: The optimal transport representation itself is a known mathematical result. The paper's primary contribution lies in utilizing this representation to characterize the set of feasible minimal reallocations consistent with marginal information, specifically extending this characterization to scenarios where marginals are partially identified due to missing data.

B. Partial Identification with Missing Data

To handle missing data, the paper adopts a worst-case partial identification approach (following Horowitz and Manski, 1995).

• Identified Sets: Let $p$ and $q$ be response probabilities. The true marginal distributions $\mu$ and $\nu$ lie within identified sets $\mathcal{M}_\mu$ and $\mathcal{M}_\nu$, defined by sharp bounds on the CDFs:
  $$p F^{obs}_\mu(k) \leq F_\mu(k) \leq p F^{obs}_\mu(k) + (1-p)$$
• Bounds on Discrepancy: Theorem 1 characterizes the identified set for the discrepancy measure $D(\mu, \nu)$.
  • Fully Observed Case: When data is complete, the identified set collapses to a point estimate.
  • Missing Data Case: When data is incomplete, the identified set expands to an interval $[\underline{D}, \overline{D}]$. These bounds are computed by minimizing and maximizing the distance over the identified sets of the marginals.
• Endpoint-Conditioned Couplings: The paper extends the analysis to the structure of change. It defines sets of optimal couplings ($\Pi^*_L$ and $\Pi^*_U$) corresponding to the lower and upper bounds of the discrepancy. Theorem 2 provides sharp bounds on the flow of mass between specific categories ($\pi_{ij}$) within these endpoint configurations.

C. Inference

The paper proposes a nonparametric bootstrap procedure (adapted from Horowitz and Manski, 2000) to construct confidence sets. This method accounts for both:

1. Sampling variability (finite sample error).
2. Identification uncertainty (the width of the bounds due to missing data).

3. Key Contributions

1. Characterization of Feasible Minimal Reallocations: The paper shifts the focus from computing a single transport cost to characterizing the set of feasible minimal-mobility configurations. It provides a structured way to describe distributional change without assuming a specific joint distribution, distinguishing between "what must occur" (minimal mobility) and "what is possible" within the identified set.
2. Integration of Missing Data: It develops a rigorous framework for partial identification in ordinal settings, deriving sharp bounds on both the magnitude of change and the structure of the transition matrix (couplings) in the presence of non-response.
3. Extremal Benchmarks: The paper introduces extremal benchmarks (lower and upper bounds) for the magnitude of change. It relates the framework to Fréchet inequalities, noting that these bounds characterize extremal movement across categories (the minimum and maximum feasible flow between specific states) rather than extremal statistical dependence.
4. Empirical Applicability: It demonstrates how to implement these bounds and conduct inference using standard resampling techniques, making the approach accessible for applied researchers.

4. Key Results

• Point vs. Interval Estimates: The scalar measure $D(\mu, \nu)$ yields a point estimate when marginals are fully observed, but becomes an interval (partially identified) when missing data is present, reflecting the uncertainty in the true magnitude of change.
• Structural Bounds: Even when the exact transition matrix is unidentified, the paper derives sharp bounds on the probability of moving from category $i$ to $j$ under the minimal-mobility assumption.
  • If the lower bound of a flow $\underline{\pi}_{ij} > 0$, then every minimal-mobility configuration in the feasible set requires movement from $i$ to $j$.
  • If the upper bound $\overline{\pi}_{ij} = 0$, then no minimal-mobility configuration uses that transition.
• Robustness to Missing Data: The empirical application shows that while the magnitude of change is sensitive to missing data (widening the confidence interval), the structure of the minimal-mobility configuration (e.g., whether changes are local or involve long jumps) often remains robust.

5. Empirical Illustration

The framework is applied to Arab Barometer data (Waves 7 and 8) for Iraq and Morocco, analyzing attitudes toward the United States on a 4-point ordinal scale.

• Findings for Iraq: A non-trivial fraction (4–10%) of the population must change categories to reconcile the waves. The minimal-mobility configuration shows that this change is primarily local (adjacent categories), suggesting gradual shifts rather than polarization.
• Findings for Morocco: The required reallocation is larger (5.5–11.5%) and involves more substantial reshuffling across categories compared to Iraq, yet still remains concentrated in adjacent transitions.
• Sensitivity Analysis: Comparing the lower and upper endpoint couplings revealed that while the amount of change is uncertain due to missing data, the pattern (local vs. long-distance movement) is stable. This suggests that conclusions about the nature of distributional change are robust to item non-response.

6. Significance

This paper provides a critical tool for researchers working with repeated cross-sectional data where panel data is unavailable.

• Interpretability: It moves beyond simple descriptive statistics (like differences in means or shares) to provide a structural interpretation of how distributions change by defining the set of feasible minimal reallocations.
• Robustness: By explicitly modeling missing data through partial identification, it prevents over-interpretation of results and quantifies the uncertainty introduced by non-response, distinguishing between point estimates (complete data) and intervals (incomplete data).
• Policy Relevance: The distinction between "minimal required movement" and "actual movement" helps policymakers understand the lower bound of social mobility or attitude shifts necessary to explain observed trends, offering a conservative baseline for evaluating social dynamics.

In summary, Tabri's work offers a mathematically rigorous and empirically practical framework to quantify the necessity of distributional change in ordinal data, separating what is identified by the marginals from what remains unobserved.