Random Coefficient Model

Definition

A random (or changing) coefficient (RC) model is a linear regression $y_t = x_t'\beta_t + \varepsilon_t$ in which the coefficient vector $\beta_t$ varies rather than being a fixed parameter. Two distinct traditions exist: (1) the Swamy (1970/1971) i.i.d. model, in which $\beta_t$ (or $\beta_i$ across units) are i.i.d. draws from a distribution with mean $\bar\beta$ and covariance matrix $\Delta$ , estimated by generalized least squares (GLS)/minimum norm quadratic unbiased estimator (MINQUE); and (2) the Chow (1984) state-space model, in which $\beta_t$ follows a latent Markov process $\beta_t = M\beta_{t-1} + \eta_t$ , estimated by the Kalman filter. The Swamy tradition treats coefficient heterogeneity as a cross-sectional or unconditional phenomenon; the Chow tradition treats it as dynamic and serially correlated.

Key Ideas

When $V = 0$ the model collapses to fixed-coefficient ordinary least squares (OLS); the Kalman gain goes to zero and no updating occurs.
When $M = I$ (random walk) the coefficients accumulate shocks permanently — the most common time-varying parameter (TVP) specification.
When $|\text{eigenvalues}(M)| < 1$ the coefficients are stationary with unconditional mean $\bar{\beta}$ and covariance satisfying the Lyapunov equation $\Gamma_0 = M\Gamma_0 M' + V$ .
The smoothed estimate $\hat{\beta}_{t|s}$ (conditioning on all data up to $s \geq t$ ) is the minimum mean squared error (MMSE) estimate of $\beta_t$ .
Two estimation routes — Kalman filter recursion (§2) and GLS regression (§3) — produce the same optimal estimates; the Kalman smoother is a Bayesian GLS estimator with the prior implied by the state equation.

How It Works

Swamy's i.i.d. Random Coefficient Model (1970/1975)

Model specification (Swamy-Mehta 1975, eq. 2.2):

$y_t = x_t'\beta_t + \varepsilon_t, \qquad \beta_t \overset{\text{i.i.d.}}{\sim} (\bar\beta_1, \Delta_1)$

where $\Delta_1$ is an unrestricted $K\times K$ positive semi-definite covariance matrix (relaxing the diagonal restriction of Swamy 1971). The induced observation model has heteroskedastic variance $\text{Var}(y_t) = x_t'\Delta_1 x_t + \sigma^2$ , so the GLS weight matrix is $\Sigma_1 = \text{diag}[x_t'\Delta_1 x_t + \sigma^2]$ .

Minimum average-risk estimator of $\bar\beta_1$ (eq. 2.10): With prior mean $r_1$ and precision $\psi_1^{-1}$ :

$\bar\beta_1^* = (X'\Sigma_1^{-1}X + \psi_1^{-1})^{-1}(X'\Sigma_1^{-1}y + \psi_1^{-1}r_1)$

Under normality this equals the posterior mean of $\bar\beta_1 \mid y$ . Under a diffuse prior ( $\psi_1^{-1} \to 0$ ) it collapses to the minimum variance linear unbiased estimator (MVLUE):

$\bar b_1 = (X'\Sigma_1^{-1}X)^{-1}X'\Sigma_1^{-1}y$

MINQUE of $\Delta_1$ (eqs. 2.24/2.26): The minimum norm quadratic unbiased estimator solves $\bar s_1 = \bar Z_1 \delta_2$ , where $\delta_2$ stacks the $K(K+1)/2$ unique elements of $\Delta_1$ and $\bar s_1$ contains sums of squares and cross-products of GLS residuals. Negative variance estimates are truncated to zero. A Bayesian analogue with prior $(\bar\delta_2, \Psi_2)$ adds regularization.

Conceptual distinction (Swamy-Mehta 1975, §1): the paper draws a sharp line between (a) $\beta_t$ random as part of the data-generating process — a feature of the world — and (b) $\beta$ random as a subjective Bayesian prior encoding the analyst's uncertainty. Both yield structurally identical estimators (eq. 2.10) but have different interpretations and different conditions for validity.

Panel decomposition (eq. 3.2): In a cross-section–time-series panel, coefficients decompose as:

$\beta_{it} = \bar\beta + \alpha_i + \xi_{it}$

where $\alpha_i$ is an individual-specific permanent component and $\xi_{it}$ is time-varying and i.i.d. The mean coefficient $\bar\beta$ is estimated by a panel GLS/Bayes estimator (eq. 3.8); the cross-section dispersion matrix $\Delta$ is estimated by $(n-1)^{-1}S$ minus a precision-matrix correction (eq. 3.12).

Maximum Likelihood (ML) Pathology for Switching Regression (§2 intro): Before proposing the random coefficient model, the paper shows that maximum likelihood for Quandt's (1972) switching regression — a two-regime model with regime indicator $\lambda$ — is almost always unbounded: as $\sigma_i^2 \to 0$ with the regime mean held at an observed datum, the likelihood diverges. Bayesian methods with proper priors avoid this but yield posteriors with no closed form and no normal large-sample approximation (Johnson regularity conditions fail). The random coefficient model avoids both pathologies.

Simple Mean Estimator (Gumpertz-Pantula 1989)

The estimated generalized least squares (EGLS) estimator (eq. 2.4) requires $\hat{\boldsymbol{\Sigma}}_{\beta\beta}$ to be positive definite, but in small samples with large $\sigma^2$ this fails frequently (Table 1 of paper: up to 81% of replications). The simple mean $\bar{\mathbf{b}} = n^{-1}\sum_{i=1}^n \mathbf{b}_i$ avoids this:

Exact normality: $\bar{\mathbf{b}} \sim N\!\left(\boldsymbol{\beta},\; n^{-1}\boldsymbol{\Sigma}_{\beta\beta} + \sigma^2 n^{-2}\sum_i (\mathbf{X}_i'\mathbf{X}_i)^{-1}\right)$ for all $n, t$ .
Unbiased variance estimator: $n^{-1}\mathbf{S}_{bb}$ where $\mathbf{S}_{bb} = (n-1)^{-1}\sum_i (\mathbf{b}_i-\bar{\mathbf{b}})(\mathbf{b}_i-\bar{\mathbf{b}})'$ — no positive-definiteness constraint.
Test statistic: $\hat{T}_{BAR}^{*2} = [\mathbf{L}\bar{\mathbf{b}}-\lambda_0]'[\mathbf{L}n^{-1}\mathbf{S}_{bb}\mathbf{L}']^{-1}[\mathbf{L}\bar{\mathbf{b}}-\lambda_0]$ — Hotelling's $T^2$ on the $\mathbf{b}_i$ 's; $F(q,\nu^*)$ via Satterthwaite approximation where $\nu^* = g^{-1}[l'\boldsymbol{\Sigma}_{\beta\beta}l + n^{-1}\sigma^2\sum_i l'(\mathbf{X}_i'\mathbf{X}_i)^{-1}l]^2$ .
Asymptotic equivalence: $\bar{\mathbf{b}} \overset{a}{\sim} \hat{\boldsymbol{\beta}}_\text{GLS} \overset{a}{\sim} \hat{\boldsymbol{\beta}}_\text{EGLS}$ as $n\to\infty$ or $t\to\infty$ .
Trade-off: mean relative efficiency (MRE)( $\bar{\mathbf{b}}$ ) $\geq 90\%$ when model matrices are homogeneous ( $\delta=1$ ); drops to ~42% when highly heterogeneous ( $\delta=9$ ) and $\sigma^2$ is small. Prefer $\hat{T}_{BAR}^{*2}$ when $n$ is small or $\hat{\boldsymbol{\Sigma}}_{\beta\beta}$ is unreliable.

State-Space Form

$y_t = x_t'\beta_t + \varepsilon_t, \qquad \varepsilon_t \sim \mathcal{N}(0,\sigma^2)$ $\beta_t = M\beta_{t-1} + \eta_t, \qquad \eta_t \sim \mathcal{N}(0, V)$

Kalman filter (forward pass):

$\hat{\beta}_{t|t-1} = M\hat{\beta}_{t-1|t-1}$ $\Sigma_{t|t-1} = M\Sigma_{t-1|t-1}M' + V$ $K_t = \Sigma_{t|t-1} x_t \bigl(x_t'\Sigma_{t|t-1}x_t + \sigma^2\bigr)^{-1}$ $\hat{\beta}_{t|t} = \hat{\beta}_{t|t-1} + K_t\bigl(y_t - x_t'\hat{\beta}_{t|t-1}\bigr)$ $\Sigma_{t|t} = (I - K_t x_t')\Sigma_{t|t-1}$

Backward smoother (for $s > t$ , Chow 1984 eq. 2.13):

$\hat{\beta}_{t|s} = \hat{\beta}_{t|t} + H_t H_{t+1}\cdots H_{s-2} K_{s-1}(y_s - \hat{y}_{s|s-1})$

where $H_t = \Sigma_{t|t}M'\Sigma_{t+1|t}^{-1}$ .

GLS Equivalence (Chow 1984 §3)

The smoothed estimate $\hat{\beta}_{t|s}$ equals the GLS estimator obtained by regressing $(y_1,\ldots,y_s)$ on $(x_1,\ldots,x_s)$ with $\beta_t$ as the (unknown vector) coefficient, using the state equation as a set of pseudo-observations. The GLS stacked system (3.1–3.2) is:

$\begin{bmatrix} y_1 \\ \vdots \\ y_s \end{bmatrix} = \begin{bmatrix} x_1 M^0 \\ \vdots \\ x_s M^{s-1} \end{bmatrix}\beta_1 + \text{(state-equation terms)} + \text{errors}$

This equivalence provides initial conditions $\hat{\beta}_{k|k}$ and $\Sigma_{k|k}$ from the first $k$ observations without requiring a prior.

MLE of Hyperparameters (§4)

The parameters $\sigma^2$ , $P = \sigma^{-2}V$ , and $M$ are estimated by maximizing the concentrated Kalman innovation log-likelihood:

$\log L = \text{const} - \tfrac{1}{2}\log|Q| - \tfrac{1}{2}(y-Z\beta)'Q^{-1}(y-Z\beta)/\sigma^2$

where $Q = I_T + A(I_{T-1}\otimes P)A'$ . No closed form for $P$ and $M$ ; gradient-based numerical methods are used. Conditional on $P$ and $M$ , closed-form MLEs for $\sigma^2$ and the cross-equation covariance $\Sigma$ exist.

System Extensions

Seemingly Unrelated Regression (SUR)-type (§5): $m$ equations with equation-specific $\beta_{jt}$ ; joint likelihood involves the full cross-equation covariance $\Sigma$ . MLEs for $\sigma_{ij}$ given $Q$ have closed forms.

Simultaneous equations (§6): Structural equations with time-varying parameters; linearize around current estimates and apply the GLS step iteratively (analogous to two-stage least squares (2SLS) with TVP). The linearized model (6.11) replaces the nonlinear structural form with a linear approximation valid locally.

Nonlinear systems (§7): $f(y_t, x_t, \beta_t) = 0$ ; Taylor-expand around $\delta_{t|t-1}$ and apply the same Kalman/GLS machinery.

Stationary Coefficients (§8)

When $|\text{eigenvalues}(M)| < 1$ , $\beta_t$ is stationary. The initial covariance satisfies the Lyapunov equation $\Gamma_0 = M\Gamma_0 M' + V$ , which must be solved numerically for general $M$ . Hannan-Dunsmuir (1976) identification conditions for autoregressive moving average (ARMA) $(p,q)$ representations of $\beta_t$ apply.

General RC Model with Concomitants (Swamy-Tavlas 1995)

Swamy-Tavlas (1995) survey and extend the RC framework to its most general form, nesting the Swamy (1970/71) simple model and all conventional fixed-coefficient specifications as special cases.

Direct and Indirect Effects Decomposition

Each observed coefficient $\beta_{tj}$ decomposes as:

$\beta_{tj} = \alpha_{tj} + \sum_l \delta_{tl}\psi_{tlj}$

where $\alpha_{tj}$ is the direct causal effect of regressor $x_{tj}$ on $y_t$ , and $\sum_l \delta_{tl}\psi_{tlj}$ are indirect proxy effects that arise because $x_{tj}$ is correlated with omitted variables $q_{tlj}$ (indirect through $\delta_{tl}$ , with loading $\psi_{tlj}$ ). Fixed-coefficient regression estimates conflate both components: a single OLS slope captures both the structural causal channel and any proxy information carried by the regressor. RC models with observable concomitants $z_{tl}$ can separate them, provided the stochastic law condition holds.

Stochastic law condition (Pratt-Schlaifer 1988): The model coincides with a causal law when $\tilde{\mathbf{x}}_t \perp \tilde{\mathbf{a}}_{xzt}$ conditional on $\mathbf{z}_t$ . This is the RC analogue of an exogeneity condition. Derivation via eqs. (13)–(14) shows the general RC model can satisfy it; conventional fixed-coefficient (CFC) models generally cannot, because they absorb both direct and indirect effects into a single fixed coefficient.

General Model Specification

$y_t = \mathbf{x}_t'\boldsymbol{\beta}_t$ $\boldsymbol{\beta}_t = \boldsymbol{\Pi}\mathbf{z}_t + \mathbf{L}\boldsymbol{\varepsilon}_t$ $\boldsymbol{\varepsilon}_t = \boldsymbol{\Phi}\boldsymbol{\varepsilon}_{t-1} + \mathbf{a}_t, \quad \mathbf{a}_t \overset{\text{i.i.d.}}{\sim} (0, \Delta_\mathbf{a})$

$\boldsymbol{\Pi}$ : $K \times m_1$ fixed matrix of mean coefficient responses to concomitants
$\mathbf{z}_t$ : $m_1$ -vector of observable concomitants — variables correlated with indirect effects but uncorrelated with direct effects conditional on $\mathbf{x}_t$
$\mathbf{L}$ : $K \times m_2$ known loading matrix
$\boldsymbol{\Phi}$ : $m_2 \times m_2$ stable coefficient persistence matrix
Total free parameters: $n = Km_1 + m_2^2 + m_2(m_2+1)/2 + 1$

The Swamy (1970/71) simple model is the special case $\boldsymbol{\Pi}=0$ , $\boldsymbol{\Phi}=0$ , $\mathbf{L}=I_K$ , diagonal $\Delta_\varepsilon$ — i.i.d. coefficients with no systematic mean or persistence.

19 Conventional Fixed-Coefficient Special Cases

The general model nests 19 CFC models by imposing parameter restrictions (§2.4):

Cases	Model class	Key restriction
1–11	CFC models	$\boldsymbol{\Pi}=0$ AND restrictions on $\boldsymbol{\Phi}$ , $\mathbf{L}$
6	GARCH	Swamy-Tavlas (1994) transformation; variance equation from coefficient dynamics
12–19	RC models of increasing generality	Delete $\mathbf{z}_t$ restrictions progressively

Cases 1–5 cover OLS, GLS, distributed-lag, and ARMA structures. Case 6 connects to the generalized autoregressive conditional heteroskedasticity (GARCH) family: Swamy-Tavlas (1994, Economics Letters 46: 7–10) showed a GARCH conditional variance equation emerges from the coefficient dynamics when concomitants and loading matrix are appropriately restricted. Cases 12–19 restore time-varying parameters with successively fewer constraints.

Swamy-Tinsley (1980) Iterative GLS

A feasible GLS algorithm (§2.5) that avoids MCMC:

Initialize: $\boldsymbol{\Phi}=0$ , $\Delta_\mathbf{a} = I$
Compute $\boldsymbol{\beta}_t$ from current $(\boldsymbol{\Pi}, \boldsymbol{\Phi}, \Delta_\mathbf{a})$ via GLS
Update $\boldsymbol{\Pi}$ via OLS of $\hat{\boldsymbol{\beta}}_t$ on $\mathbf{z}_t$
Compute residuals $\hat{\boldsymbol{\varepsilon}}_t = \hat{\boldsymbol{\beta}}_t - \hat{\boldsymbol{\Pi}}\mathbf{z}_t$
Estimate $\boldsymbol{\Phi}$ via multivariate AR(1) on $\hat{\boldsymbol{\varepsilon}}_t$
Update $\Delta_\mathbf{a}$ from AR residuals
Recompute GLS weights
Repeat steps 2–7 until convergence

Convergence is typically achieved within a small number of cycles. The estimate of $\boldsymbol{\Pi}$ provides the direct-effects component; the residual $\mathbf{L}\hat{\boldsymbol{\varepsilon}}_t$ contains the indirect effects.

Empirical Record (Swamy-Tavlas 1995, Table 3)

Nine macroeconomic forecasting applications comparing RC models to their best CFC alternative:

Application	RC RMSE	Best CFC RMSE	Reduction
US M1 demand (missing money period)	4.6	19.8	77%
Dollar/DM exchange rate	beats RW	CFC above RW	—
Dollar/pound exchange rate	beats RW	CFC above RW	—
Dollar/yen exchange rate	beats RW	CFC above RW	—
US short-term interest rate	lowest	highest	11–60%

Schinasi-Swamy (1989) replicates the Meese-Rogoff (1983) finding for CFC models but shows RC models beat the random walk for all three dollar exchange rate pairs.

Testing Coefficient Constancy (Chow 1984 §10)

Test	$H_0$	Statistic	Reference
Chow F-test	Known break at $T_B$	Ratio of SSRs before/after	Chow (1960)
CUSUM	No structural change	Cumulated recursive residuals vs. straight-line bounds	Brown-Durbin-Evans (1975)
CUSUM $^2$	No variance change	Cumulated squared residuals	Brown-Durbin-Evans (1975)
Quandt max-LRT	Unknown break date	$\max_\lambda F(\lambda)$ over interior of sample	Quandt (1960)
LMP (locally most powerful) / Nyblom	$V = 0$	Score statistic against random-walk alternative	Nyblom (1983)
Pagan score	$V = 0$	Score statistic	Pagan (1980)

Why It Matters

Foundational framework for the Time-Varying Parameter VAR: the multivariate extension stacks $\beta_t = \text{vec}[(A_{1t},\ldots,A_{pt})']$ and applies the same Kalman filter / Carter-Kohn algorithm (Amisano-Federico 2004).
The GLS–Kalman equivalence shows that the Kalman smoother is a Bayesian GLS estimator; this clarifies the relationship between frequentist and Bayesian estimation of TVP models.
Completes Bayesian Multiperiod Prediction (Chow 1973): that paper showed the optimal predictor requires $E[A^k \mid \text{data}]$ ; this paper provides the estimation framework for the coefficient process ( $M$ , $V$ ) that feeds into that expectation.
The constancy testing survey (§10) synthesizes pre-1984 methods; post-1984 work (Zivot-Andrews 1992, Bai-Perron 1998) extends to multiple unknown breaks — see Structural Break Testing.

Open Questions

Consistent selection of the coefficient dynamics order $(p,q)$ in the ARMA representation of $\beta_t$ remains non-trivial.
High-dimensional systems: the $K \times K$ Lyapunov equation becomes infeasible for large $K$ ; sparsity or factor structure in $V$ is needed.
The Nyblom (1983) LMP test has low power against smooth gradual change; smooth transition autoregressive (STAR)-type alternatives are complementary.