Dynamics in Latent Variables

Reuning, Kenwick & Fariss — Political Analysis (2019)

panel data
code
analysis
Author

Robert W. Walker

Published

July 27, 2026

The PDF

The Core Problem

Political scientists increasingly use latent variable models to measure unobservable concepts—ideology, democracy, human rights, transparency.

When applied to time-series cross-sectional (TSCS) data, researchers face a fundamental tradeoff:

Static models

  • Minimize bias
  • Allow rapid change
  • ❌ Ignore temporal structure
  • ❌ Lose efficiency

Dynamic models

  • Model temporal dependence
  • Leverage time-series information
  • ❌ Over-smooth rapid changes
  • ❌ Bias when shocks occur

Goal: A model that handles both stability and sudden change.

Static Models

The simplest approach: treat every unit-period as independent.

\[\theta_{it} \sim \mathrm{N}(0, 1) \quad \forall i = 1,\ldots,N \;\; \& \;\; \forall t = 1,\ldots,T\]

How it works

  • Estimates for each unit-period depend only on observed manifest variables at that time point
  • Allows sudden changes in the latent trait between periods

Limitations

  • Treats observations as independent → violates temporal structure
  • Social science indicators are often coarse or missing
  • Results in wide, uninformative credible intervals when data are sparse

Dynamic Models

The standard solution: model the latent trait as a random walk.

\[\theta_{i1} \sim \mathrm{N}(0,1), \qquad \theta_{it} \sim \mathrm{N}(\theta_{i(t-1)},\, \sigma), \qquad \sigma \sim \mathrm{HN}(0,3)\]

Benefits

  • Incorporates temporal autocorrelation as prior information
  • Produces tighter (more efficient) credible intervals
  • Theoretically sensible when traits are slow-moving

Critical limitation

  • Smoothing induces bias near shocks
  • Rapid institutional collapse, elections, coups appear as gradual transitions
  • Dynamic model performs poorly in the periods surrounding a sudden change

The Robust Dynamic Model

Key insight: Replace the Normal transition prior with a Student’s t distribution.

\[\theta_{i1} \sim \mathrm{N}(0,1), \qquad \theta_{it} \sim \mathrm{T}_4(\theta_{i(t-1)},\, \sigma), \qquad \sigma \sim \mathrm{HN}(0,3)\]

The Student’s t with 4 degrees of freedom has heavier tails than the Normal:

  • During stability → behaves like the dynamic model (efficient, smooth)
  • During shocks → the heavy tails permit large jumps without over-smoothing

The model retains temporal structure while accommodating extreme values (“outliers” = shocks).

This is equivalent to the standard dynamic model in the absence of volatility, and outperforms it when volatility is present.

Simulation Design

To compare all three models under known data-generating conditions:

  1. Latent trait \(\theta_{i1}\) drawn from \(\mathrm{N}(0,1)\)
  2. Trait follows a random walk thereafter
  3. With probability p, a unit experiences a shock — \(\theta_{it}\) re-drawn from \(\mathrm{N}(0,1)\)
  4. Binary manifest indicators generated with error from the true latent trait

Models evaluated on:

  • 95% credible interval coverage around shocks
  • Within-unit rank correlations
  • Cross-validated accuracy
  • Differences between adjacent time periods

Benchmark parameters: shock probability = 0.01, innovation SD = 0.05.

Simulation Results

For a stable unit (no shock):

  • Static: wide, wandering intervals — inefficient
  • Dynamic & Robust: tight, well-centred intervals — equally good

For a unit experiencing a shock (period 15):

  • Static: captures the true value but with very wide intervals
  • Dynamic: over-smooths the jump; misses the true value for several periods
  • Robust: tight intervals that rapidly adapt to the new level

Key finding: The robust model is never worse than the dynamic model, and is substantially better when shocks occur.

Model accuracy in time surrounding shocks. The dynamic model (green) dips sharply at t = 0; the robust model (orange) is substantially better calibrated.

Application 1 — Judicial Ideology

Data: US Supreme Court votes, 1937–2015 (Martin & Quinn 2002)

  • Items = individual justice votes (affirm/overturn)
  • Original model used a dynamic prior

Posterior predictive accuracy

Model Correct Predictions
Dynamic 72.77%
Robust 72.95%

WAIC (lower = better fit)

Model WAIC
Dynamic 46,852
Robust 46,647 (Δ = 205, SE = 15.6)

Substantive finding: The robust model detects a sudden shift in Rehnquist’s voting in 1987— his first year as Chief Justice—consistent with strategic incentive changes. The dynamic model smooths this over and misses it entirely.

Application 2 — Democracy

Data: Country-year democracy estimates, 1950–2008 (Pemstein, Meserve & Melton 2010)

  • 10 ordinal indicators (Polity, Freedom House, Polyarchy, Bollen, Vanhanen, …)
  • Original model used a static prior

WAIC comparison

Model WAIC
Static 93,267
Dynamic 79,237
Robust 65,082

Substantive findings

  • Philippines: Robust model clearly identifies the 1972 imposition of martial law under Marcos and the subsequent recovery after 1981 and 1987 — the dynamic model anticipates these changes too early and underestimates the abruptness.
  • Afghanistan: Only the robust model captures the short-lived Saur Revolution in 1978 (quickly followed by the Soviet invasion in 1979) — invisible in the dynamic and static models.

Discussion & Takeaways

The robust dynamic model is a better default when:

  • The latent trait is subject to punctuated equilibria or sudden regime change
  • Researchers want efficiency and the ability to detect rapid shifts
  • The phenomenon of interest includes institutional collapse, coups, or electoral shocks

Practical recommendations

  1. Estimate both dynamic and robust models; compare with WAIC and posterior predictive checks
  2. Visually inspect well-known historical cases as a validity check
  3. Estimate \(\sigma^2\) directly from data when possible; sensitivity-test fixed values
  4. Be cautious when \(\sigma^2 \to 0\): the robust model may artificially detect shocks

Limitations & future work

  • Change-point models, mixture models, and continuous-time latent variable models are complementary alternatives worth exploring
  • Causal explanation of detected shocks remains outside the scope of measurement models

Summary

Feature Static Dynamic Robust
Models temporal structure ❌ ✅ ✅
Efficient (tight intervals) ❌ ✅ ✅
Handles sudden shocks ✅ ❌ ✅
Unbiased near shocks ✅ ❌ ✅
Fit (WAIC — democracy) worst middle best

Replacing the Normal transition prior with a Student’s t(4) distribution is a simple, theoretically motivated change that substantially improves latent variable estimates for politically volatile phenomena— at no cost when the world is stable.

Replication code & data: https://doi.org/10.7910/DVN/SSLCFF