📈 economics

Spatiotemporal Autoregressive Models for Areal Compositional Data

This paper introduces a spatiotemporal multivariate autoregressive model tailored for areal compositional panel data, establishing its theoretical properties and demonstrating its utility in capturing complex economic dynamics through applications to Berlin's housing market and Spain's regional sectoral compositions.

Original authors: Matthias Eckardt, Philipp Otto

Published 2026-03-16

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Matthias Eckardt, Philipp Otto

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 understand the economy of a city or a country, but instead of looking at total numbers (like "total sales"), you are looking at recipes.

In economics, many things are "compositional." This means they are parts of a whole that must always add up to 100%.

Example: In a city's housing market, every transaction is either a Condo, a Developed Plot, or Undeveloped Land. If condos go up, the other two must go down. They are locked in a dance where one step up forces the others to step down.

For a long time, statisticians have struggled to model these "recipes" when they change over time (temporal) and across space (geography). Standard math tools treat these numbers like independent ingredients, which breaks the "recipe" rule and leads to wrong answers.

This paper introduces a new, sophisticated "recipe tracker" called a Spatiotemporal Autoregressive Model for Compositional Data. Here is how it works, explained simply:

1. The Problem: The "Zero-Sum" Puzzle

Imagine a pie. You can't just add more chocolate to the pie without taking some vanilla away.

Old Models: Treated the chocolate and vanilla as if they were in separate bowls. They would predict that chocolate could go up and vanilla could go up at the same time, which is impossible for a single pie.
The Issue: Real-world data (like Berlin's housing market or Spain's business sectors) has two extra layers of complexity:
1. Time: What happened last month affects this month.
2. Space: What happens in Berlin's north affects Berlin's south (neighbors influence neighbors).

2. The Solution: The "Magic Translator"

The authors realized they couldn't do the math on the "pie" directly because the rules are too weird (the sum must always be 100%). So, they invented a Magic Translator.

The Analogy: Imagine the pie is a flat, round table. It's hard to do calculus on a round table. So, they use a special lens (called an Isometric Log-Ratio transformation) to project that round table onto a flat, infinite sheet of paper (Euclidean space).
Why do this? On the flat sheet, the math becomes normal and easy. They can use standard tools to track how the "ingredients" move around, knowing that when they project the results back onto the round table, the "100% rule" is automatically respected.

3. The Engine: The "Echo Chamber"

Once the data is on the flat sheet, the model uses a system of Echoes to predict the future.

Temporal Echo (Time): If the housing market was hot last month, it's likely to be hot this month. The model listens to the "echo" of the past to predict the future.
Spatial Echo (Space): If a neighborhood in Berlin starts selling more condos, does the neighborhood next door do the same? The model maps out a "neighborhood map" (a weight matrix) to see how much one area whispers to its neighbors.
The Autoregressive Part: This just means the model is a "self-remembering" machine. It looks at its own history and its neighbors' history to make a guess about what happens next.

4. The Real-World Tests

The authors tested their new "Recipe Tracker" on two very different scenarios:

Case A: Berlin's Housing Market (The "Fast" City)
- Data: Monthly data from 1995–2015 across 24 districts.
- What they found: The market is very sticky. If you see a rise in condo sales, it tends to keep rising for a while. Neighbors do influence each other, but the biggest driver is the city's own recent history. They saw a clear shift: people stopped buying raw land and started buying finished condos.
Case B: Spain's Economy (The "Slow" Country)
- Data: Annual data from 2012–2021 across 2,793 towns.
- What they found: Here, the "neighbor" effect was almost non-existent. A town's economy was mostly just repeating what it did last year. The "echo" of the past was very loud, but the "whisper" from neighbors was quiet.

5. Why This Matters

Before this paper, if you wanted to study how economic sectors shift over time and space, you had to use clumsy tools that often gave you impossible results (like predicting 110% market share).

This new model is like a GPS for economic recipes. It tells you:

Where the economy is going (Time).
How one region influences another (Space).
How the parts of the economy trade off against each other (Composition).

The Takeaway

The authors built a mathematical bridge that allows us to study complex, shifting economic landscapes without breaking the fundamental rules of how those landscapes work. It's a tool that helps policymakers and economists understand not just what is changing, but how the change ripples through time and across the map, all while keeping the math honest.

1. Problem Statement

The paper addresses a significant gap in spatial econometrics and compositional data analysis (CoDA). While regional economic systems (e.g., housing markets, industrial sectors) are often described by compositional data (vectors of proportions summing to a constant, such as market shares), existing models struggle to handle these data when they exhibit both spatial dependence (neighboring regions influence each other) and temporal dynamics (autocorrelation over time).

Key challenges identified include:

Simplex Constraints: Compositional data resides on a simplex (sum-to-one constraint), violating the assumptions of standard Euclidean statistical models.
Spatiotemporal Complexity: Conventional methods often treat spatial and temporal dependencies separately or fail to account for the intrinsic dependence structure of compositions (where an increase in one component necessitates a decrease in others).
Lack of Frameworks: There is a scarcity of models specifically designed for areal panel data where the response variable is a composition observed over time across multiple regions.

2. Methodology

The authors propose a Spatiotemporal Multivariate Simultaneous Autoregressive (MSAR) model tailored for compositional panel data. The framework integrates three core components:

A. Geometric Transformation (Isometric Log-Ratio)

To handle the simplex geometry ( $S^D$ ), the model employs the Isometric Log-Ratio (ilr) transformation.

This maps the $D$ -part composition $Z_t$ from the simplex to a $(D-1)$ -dimensional Euclidean space ( $\mathbb{R}^{D-1}$ ).
The transformation preserves the Aitchison geometry (distances and inner products) of the compositional space, allowing standard linear algebra and likelihood-based inference to be applied in the transformed space.
The model assumes strictly positive components, though the authors note extensions (like $\alpha$ -transformations) exist for zero-inflated data.

B. The Spatiotemporal MSAR Model

Let $Y_t$ be the vectorized ilr-transformed composition for $n$ spatial units at time $t$ . The model is defined as:
$Y_t = \sum_{i=1}^q X_{t,i}\beta_i + W Y_t \Psi + \sum_{\ell=1}^L Y_{t-\tau_\ell} \Pi_\ell + E_t$
Where:

$W$ : Known spatial weight matrix (e.g., contiguity or distance-based).
$\Psi$ : $p \times p$ matrix of spatial autoregressive coefficients (capturing contemporaneous spatial and cross-component dependence).
$\Pi_\ell$ : Coefficient matrices for temporal lags $\tau_\ell$ (capturing temporal and cross-component lagged effects).
$X_{t,i}$ : Exogenous regressors.
$E_t$ : Error term with i.i.d. structure.

The model is estimated using a Quasi-Maximum Likelihood (QML) approach. The authors derive the reduced form of the model using the Kronecker product to vectorize the system, facilitating the construction of the Gaussian quasi-log-likelihood function.

C. Theoretical Properties

The paper establishes rigorous asymptotic properties under a joint asymptotic framework where both the number of spatial units ( $n$ ) and time points ( $T$ ) grow.

Identifiability: Proven that parameters are uniquely identifiable.
Consistency: The QML estimator is consistent as $T \to \infty$ (with $n$ potentially fixed or growing, subject to $n/T \to 0$ ).
Asymptotic Normality: The estimator follows a normal distribution, allowing for valid hypothesis testing and confidence intervals. The asymptotic variance is derived using a "sandwich" form to account for potential misspecification of the error distribution.

3. Key Contributions

Novel Model Specification: Introduction of the first multivariate simultaneous autoregressive model specifically designed for spatiotemporal areal compositional data.
Theoretical Foundation: Proofs of identifiability, consistency, and asymptotic normality for the QML estimator in a high-dimensional panel setting ( $n, T \to \infty$ ).
Interpretation Framework: Development of a method to interpret linear parameters in the ilr-space back onto the original simplex. The authors define local marginal effects, direct effects (own-unit), and indirect/spillover effects (neighboring units) that respect the sum-to-one constraint.
Robust Inference: Demonstration of the necessity of robust variance estimators (HAC) when dealing with non-Gaussian innovations or misspecified temporal dynamics.

4. Empirical Results

The model is validated through two distinct case studies:

Case Study 1: Berlin Real Estate Market (1995–2015)

Data: Monthly transaction compositions (Condominiums, Developed Land, Undeveloped Land) across 24 postcode areas.
Findings:
- Temporal Dominance: Strong short-run persistence (lag-1) and significant seasonal effects (lags 6 and 12).
- Spatial Effects: Moderate but statistically significant spatial spillovers, primarily concentrated in the "developed vs. undeveloped" balance.
- Dynamics: Impulse response functions show that shocks lead to a reallocation from undeveloped land to condominiums, with effects decaying over time but exhibiting seasonal re-oscillation.
- Fit: The model captures the complex spatiotemporal structure better than conventional methods, with residuals showing no systematic spatial autocorrelation.

Case Study 2: Spanish Economic Sectors (2012–2021)

Data: Annual sectoral compositions (Services, Industry, Construction) across 2,793 municipalities.
Findings:
- High Persistence: Extremely strong temporal autoregression (coefficients $\approx 0.96$ ), indicating slow evolution of sectoral structures.
- Weak Spatial Dependence: Unlike the Berlin case, spatial spillovers were negligible and statistically insignificant, consistent with the lack of spatial clustering in the descriptive plots.
- Model Performance: High $R^2$ values in the transformed space and low prediction errors on the simplex.

5. Significance and Implications

Methodological Advancement: The paper provides a unified framework for analyzing constrained multivariate panel data, bridging the gap between spatial econometrics and compositional data analysis.
Policy and Economic Insight: The ability to decompose effects into direct and spillover components on the original compositional scale allows policymakers to understand how local shocks (e.g., a housing boom) propagate to neighboring regions and how they alter the structural composition of the economy.
Flexibility: The framework is adaptable to various fields beyond economics, including ecology (species composition), epidemiology (disease prevalence shares), and demography.
Future Directions: The authors suggest extensions for zero-inflated data, non-Gaussian error structures, and time-varying coefficients to capture structural breaks.

In conclusion, Eckardt and Otto successfully demonstrate that ignoring the compositional nature of areal panel data leads to misspecification. Their proposed MSAR-ilr framework offers a statistically rigorous and interpretable tool for modeling the complex interplay of space, time, and structural constraints in regional economic systems.