Latent Monthly Mortality Factor

• Latent Monthly Mortality Factor is a low-dimensional, dynamic process in stochastic mortality models that captures systematic monthly variations, including seasonality and epidemiological shocks.
• It is estimated using methods such as state-space modeling, factor analysis, and tensor decomposition, which link observed mortality data to underlying trends and transitory shocks.
• The factor facilitates accurate nowcasting, forecasting, and imputation of mortality data, enabling robust epidemiological interpretations and policy assessments.

A latent monthly mortality factor is a low-dimensional, typically unobservable, dynamic process that drives the bulk of temporal variation in mortality data at monthly resolution. In contemporary stochastic mortality modeling, such factors serve as core hidden states linking observed high-frequency deaths, exposures, and rates to underlying epidemiological and environmental mechanisms, seasonality, transitory shocks, and long-run trends. Depending on the modeling framework, the latent monthly factor may be identified via factor-analytic decompositions, dynamic state-space models, structured regression with seasonal, trend, and shock terms, or high-dimensional tensor decompositions. Estimation, interpretability, and forecasting depend crucially on the chosen statistical architecture, identifiability constraints, and inferential methodology.

1. Conceptual Foundation and Formal Definitions

The role of a latent monthly mortality factor is to capture, in an interpretable or at least low-dimensional fashion, the systematic variation of mortality over months, separating it from noise, age/sex/country structure, or cause-specific idiosyncrasies. In classical factor models (e.g., Lee-Carter, Li-Lee), the temporal index $\kappa_t$ or $k_t$ (with $t$ a monthly index) constitutes the primary latent factor, modulated by age-specific loadings (He et al., 2021). In array and tensor models, the relevant factor may emerge as one axis of a separable decomposition, as in separable factor analysis (SFA) (Fosdick et al., 2012) or Bayesian tensor-train structures (Zhang et al., 2023). State-space models introduce time-evolving latent factors $k_t$ governed by explicitly specified stochastic dynamics, such as seasonal ARIMA processes (Li et al., 9 Jan 2026). In all cases, the factor is defined via a hierarchical likelihood, either as a direct regressor or indirectly via probabilistic data-generating processes.

2. Statistical Models for Latent Monthly Mortality Factors

2.1. Separable Factor Analysis (SFA)

SFA provides a canonical route to decomposing multiway mortality arrays (age, sex, country, month) and extracting mode-specific latent factors (Fosdick et al., 2012). The monthly mode covariance $\Sigma_M$ is parameterized as

$\Sigma_M = \Lambda_M \Lambda_M^T + \Psi_M$

where $\Lambda_M$ is the $M \times k$ loading matrix. Conditional monthly factor scores $\widehat{z}_j$ give scalar or vector-valued indices that quantify how unusual mortality is in month $j$, after controlling for other dimensions, with

$$\widehat{z}_j = \bigl(\Lambda_M^T \Psi_M^{-1} \Lambda_M + I_k \bigr)^{-1} \Lambda_M^T \Psi_M^{-1} y_j$$

Projection of residuals onto loadings yields a scalar "latent monthly mortality factor" when $k=1$.

2.2. State-Space and Mixed-Frequency Models

In the mixed-frequency state-space model, the latent monthly mortality factor $k_t$ underlies both observed annual rates and monthly counts. Its dynamics are specified by a seasonal ARIMA (e.g., SARIMA$(2,0,0)\times(0,1,0)_{12}$), and it is inferred via Kalman smoothing and expectation-maximization: $$(1-\phi_1 L - \phi_2 L^2) (1-L^{12}) k_t = \mu + w_t$$ with direct linkages to both monthly and aggregated annual mortality via loading structures. Filtering, smoothing, and predictive recursion allow real-time tracking and forecasting of this factor as new monthly data accrue (Li et al., 9 Jan 2026).

2.3. Dynamic Regression and Seasonal Baseline Terms

Regression structures with explicit seasonal and trend terms can also define smooth latent monthly series, such as the baseline respiratory mortality factor $B_t$ in Russian mortality, which combines annual harmonics, calendar dummies, and time trends (Goldstein, 2020): $$B_t = \beta_6 \cos(2\pi t/12) + \beta_7 \sin(2\pi t/12) + \beta_8 \mathbf{1}_{\mathrm{Jan}(t)} + \beta_9 SE(t) + \beta_{10} u(t) + \beta_{11} u(t)^2$$ Monthly factors of this form capture the baseline component not directly associated with explicit exogenous drivers (e.g., influenza proxies).

2.4. Tensor-Train Factorizations

Bayesian Poisson+Tensor-Train models for high-dimensional monthly count data (across age, sex, cause, month) return low-rank sets of monthly factors from the time-core of the TT decomposition: $$R_{i,j,k,t} \approx \sum_{r_1, r_2, r_3} G^{(1)}_{i,r_1} G^{(2)}_{j,r_1, r_2} G^{(3)}_{k, r_2, r_3} G^{(4)}_{t, r_3}$$ with each column $f_\ell$ of $G^{(4)}$ forming a distinct latent monthly mortality factor, interpretable as baseline, seasonality, COVID, harvesting, or compensation effects (Zhang et al., 2023).

3. Estimation and Computational Methodologies

Estimation methods reflect the underlying statistical structure:

• Maximum Likelihood (ML) and Bayesian Inference: SFA and tensor-train decompositions employ block-coordinate ascent, EM, or Gibbs/Metropolis-within-Gibbs to optimize likelihoods or sample from joint posteriors given latent hidden states (Fosdick et al., 2012, Zhang et al., 2023).
• Kalman Filtering and Smoothing: State-space approaches leverage Kalman recursions for real-time filtering and trajectory smoothing, integrating both annual and monthly information (with missing-data handling for incomplete intra-annual periods) (Li et al., 9 Jan 2026).
• Principal Component Analysis / SVD: For classical and time-varying factor models, SVD and localized PCA extract principal monthly indices and resolve time-dependent loadings (He et al., 2021).
• GLM and Poisson Regression: Baseline models estimated via GLM or Poisson regression link explicit seasonal, trend, and event terms to observed monthly rates, extracting the latent baseline factor as a fitted component (Goldstein, 2020, Berkum et al., 2022).
• Post-processing for Latent Factor Extraction: In Bayesian GMRF models, posterior MCMC output across age and time yields factor scores via empirical PCA (Alexopoulos et al., 2018).

4. Model Selection, Identification, and Diagnostics

Model identification depends on appropriate constraints—sum-to-zero on factors, unit-norm loadings, Kronecker constraints in array models, or penalization in tensor decompositions (Fosdick et al., 2012, Berkum et al., 2022). Key model selection techniques include:

• Likelihood Ratio Tests: To determine appropriate factor dimensionality ($k$), sequential $\chi^2$ tests between nested models are employed (Fosdick et al., 2012).
• Cross-Validation: Holdout experiments select $k$ or regularization parameters based on out-of-sample mean squared error (Fosdick et al., 2012).
• Forecast Validation: Rolling-origin and split-sample validation clarify boundary points in time-varying loading models for switching between local regression extrapolation and steady-state (naive) forecast (He et al., 2021).
• Epidemiological Interpretation: Leading factors are interpreted via their posterior mean trajectories—seasonal patterns, event (e.g., pandemic) shocks, and long-term trends are mapped to distinct factor components (Zhang et al., 2023).

5. Interpretability and Epidemiological Significance

The latent monthly mortality factor can represent seasonality (winter/summer cycles), long-run annual drift, extraordinary shocks (influenza, pandemics, heat waves), or operational artifacts (reporting delays). For example:

• Seasonal Factor: The leading monthly factor in SFA or tensor-train models typically visualizes canonical winter mortality peaks and summer troughs (Fosdick et al., 2012, Zhang et al., 2023).
• Pandemic/Excess Factors: COVID-specific latent factors extracted from high-frequency death data capture within-year mortality shocks, with separation from regular seasonal structure via model architecture (Berkum et al., 2022, Zhang et al., 2023).
• Baseline Trend: In regression decomposition, decline in the latent baseline factor post-2015 in Russian mortality aligns with pneumococcal vaccine rollout, allowing interpretation of observed reduction as vaccine-induced or care-enhanced (Goldstein, 2020).

6. Applications and Comparison Across Frameworks

Latent monthly mortality factors are applied in:

• Nowcasting and Forecasting: Real-time state-space updates enable intra-year nowcasting and improved annual forecasts as new data arrive, outperforming annual-only or temporally reconciled approaches (Li et al., 9 Jan 2026).
• Mortality Imputation and Smoothing: SFA-based factors enable efficient and epidemiologically informed imputation of missing monthly rates, exploiting shared structure across age, sex, and country (Fosdick et al., 2012).
• Pandemic Impact Assessment: Decomposition into baseline and excess latent monthly factors separates routine mortality from transient epidemic or environmental shocks (Berkum et al., 2022, Zhang et al., 2023).
• Structural Change and Cohort Analysis: Time-varying factor-loading structures adapt to long-term structural change in age-mortality profiles, facilitating robust long-horizon mortality extrapolation (He et al., 2021).

In summary, latent monthly mortality factors occupy a central role in high-frequency mortality modeling, providing a parsimonious, interpretable, and dynamically-updatable basis for estimation, prediction, and epidemiological interpretation of complex, multivariate death data (Fosdick et al., 2012, Li et al., 9 Jan 2026, Zhang et al., 2023, He et al., 2021, Goldstein, 2020, Berkum et al., 2022, Alexopoulos et al., 2018).