Bayesian Seasonal Adjustment for Survey Time Series

Technical Summary: Bayesian Seasonal Adjustment for Survey Time Series

Problem Statement
Current seasonal adjustment procedures employed by national statistical offices (NSOs), specifically X-11 and X-12-ARIMA, treat survey estimates as exact observations. This approach discards the accompanying standard errors ($SE^d_t$) routinely computed by survey methodologists. Consequently, months with high sampling variance (high coefficient of variation, CV) receive the same weight in seasonal filters as months with low variance, despite the former having significantly larger measurement noise. This structural limitation prevents these methods from distinguishing between genuine seasonal drift and fluctuations caused by sampling error, and they fail to provide uncertainty measures (e.g., credible intervals) for trend movements or directional changes.

Methodology
The paper proposes the Dynamic Mini-Max Bayesian Structural Model (DMM-BSM), a framework that integrates time-varying sampling error variances directly into a Basic Structural Model (BSM) within a fully Bayesian inference scheme.

1. Model Hierarchy:

   • Sampling Model: Relates the direct survey estimate $Y^d_t$ to the latent BSM state $\alpha^d_t$ via a measurement equation: $Y^d_t = Z^\top \alpha^d_t + e^d_t$, where the error variance $V^d_t = (SE^d_t)^2$ is treated as known and time-varying. This contrasts with the standard BSM assumption of zero measurement error.
   • Linking Model: Inherited from Tam (2026), this component allows for borrowing strength across strata via hierarchical regression, though the empirical illustration operates at the domain level where stratum-level data is unavailable.
   • Propagation Model: Replaces the AR(1) model of previous work with a 12-dimensional BSM state vector comprising a local level trend ($L_t$) and a stochastic seasonal component ($\gamma_t$) for monthly data ($s=12$). The system matrices $T$ and $R$ are fixed, containing no free parameters.

2. Theoretical Equivalence:
   The paper leverages the Harvey–Todd (1983) equivalence, which establishes that a BSM with zero measurement error variance ($V_t=0$) reduces to the Airline ARIMA(0,1,1)(0,1,1)$_{12}$ model. Combined with Cleveland and Tiao (1976), this confirms that the BSM with $V_t=0$ is asymptotically equivalent to the X-11 filter. Thus, the DMM-BSM is presented not as a departure from X-11, but as a principled Bayesian generalization that incorporates sampling error.

3. Inference Algorithm:
   A two-block Gibbs sampler is employed to deliver the full joint smoothing posterior $\pi(\alpha_{1:T} | y_{1:T})$:

   • Block 1 (State Trajectory): Uses the Forward Filter Backward Sampler (FFBS) (Carter and Kohn, 1994; Frühwirth-Schnatter, 1994). The forward pass utilizes the Kalman filter with time-varying measurement variance $V^d_t$ to update the state. The backward pass draws the full state trajectory jointly from the smoothing posterior.
   • Block 2 (Hyperparameters): Updates variance hyperparameters ($\sigma^2_\xi, \sigma^2_\omega, \sigma^2_u$) and regression coefficients ($\beta$) using conjugate inverse-gamma and normal full conditionals, conditional on the state trajectory drawn in Block 1.
   • Because the system matrices are fixed, the sampler is "pure Gibbs" with no Metropolis steps.

Key Contributions

1. BSM Propagation: Replaces scalar AR(1) dynamics with a 12-dimensional BSM, enabling the separation of trend and seasonal dynamics within a Bayesian framework.
2. Sampling Error Incorporation: Explicitly models time-varying sampling variance $V^d_t$ in the Kalman filter measurement equation. The Kalman gain automatically down-weights high-CV observations relative to the structural prediction, a mechanism absent in X-11.
3. Exact Movement Inference: The FFBS Gibbs sampler enables the direct calculation of Calibrated Bayes Intervals (CBIs) for $k$-step trend movements ($\Delta^{(k)}_t = L_t - L_{t-k}$) and posterior probabilities of directional change ($P(\Delta_t > 0 | y_{1:T})$) for any $k \ge 1$, without requiring closed-form derivations.
4. Verified Coverage Superiority: Simulation studies demonstrate that DMM-BSM achieves substantially higher empirical coverage than the X-11-equivalent model (BSM with $V_t=0$), particularly in small domains with high CV.

Empirical Results
The framework was applied to 120 months of Australian Bureau of Statistics (ABS) Labour Force Survey data (Full-time employment) for two domains: Australia (AU, large domain, mean CV 0.38%) and the Australian Capital Territory (ACT, small domain, mean CV 1.62%).

• Trend Estimates: In the large domain (AU), DMM-BSM trends closely track X-11 trends, confirming the Harvey–Todd equivalence. In the small domain (ACT), DMM-BSM produces wider credible intervals during periods of high CV (e.g., the COVID-19 disruption), correctly reflecting increased uncertainty, whereas X-11 provides no uncertainty measure.
• Seasonal Variance: Maximum Likelihood Estimation (MLE) of the stochastic seasonal variance ($\sigma^2_\omega$) in the DMM-BSM collapses to zero (a boundary solution). This suggests that apparent seasonal fluctuations in the published series are largely attributable to measurement noise rather than genuine stochastic seasonal drift. In contrast, the X-11-equivalent model inflates $\sigma^2_\omega$ (to 573 persons) to compensate for the lack of a measurement noise channel.
• Coverage Performance:
  • AU (Large Domain): DMM-BSM achieved 94.6% coverage for trend levels and 94.5% for trend changes (near the nominal 95%). The X-11-equivalent achieved only 78.8% and 64.3%, respectively.
  • ACT (Small Domain): DMM-BSM achieved 86.7% (trend level) and 82.1% (trend change). While below nominal, this is a significant improvement over the X-11-equivalent's 68.2% and 50.8%. The paper attributes the remaining gap in small domains to the difficulty of estimating variance parameters from short, low signal-to-noise series, noting that the bias is conservative (intervals are valid but potentially wider than necessary).

Significance
The paper claims that DMM-BSM offers a "principled Bayesian generalisation" of existing practice. Its primary significance lies in correcting the structural flaw of treating survey estimates as exact. By incorporating sampling variances, the method prevents the "chasing of sampling noise" by the trend filter, which leads to overconfident and inaccurate intervals in standard X-11 approaches. The framework provides NSOs with model-conditional uncertainty quantification for trend movements and directional probabilities—outputs that are currently unavailable from standard seasonal adjustment tools. The empirical finding that stochastic seasonal variance collapses to zero when sampling error is properly modeled offers a new perspective on the nature of seasonal fluctuations in survey data, distinguishing between genuine drift and measurement noise.