Gumpertz-Pantula (1989) A Simple Approach to Inference in Random Coefficient Models

Summary

Gumpertz and Pantula propose using the simple arithmetic mean $\bar{\mathbf{b}} = n^{-1}\sum_{i=1}^n \mathbf{b}_i$ of individual unit ordinary least squares (OLS) estimates as an alternative to estimated generalized least squares (EGLS) for inference about the population mean $\boldsymbol{\beta}$ in Swamy's random coefficient regression (RCR) model. The key advantage is that $\bar{\mathbf{b}}$ is exactly normally distributed for all sample sizes and its variance is unbiasedly estimated by $n^{-1}\mathbf{S}_{bb}$ without ever computing $\hat{\boldsymbol{\Sigma}}_{\beta\beta}$, which is frequently non-positive-definite in practice. The corresponding test statistic $\hat{T}_{BAR}^{*2}$ (Hotelling's $T^2$ on the $\mathbf{b}_i$'s) maintains empirical size close to 5% even when EGLS-based tests are severely oversized, and is asymptotically equivalent to EGLS as $n \to \infty$ or $t \to \infty$.

Key Claims

• Model: $\mathbf{y}_i = \mathbf{X}_i\boldsymbol{\beta}_i + \mathbf{e}_i$ with $\boldsymbol{\beta}_i \overset{\text{i.i.d.}}{\sim} N(\boldsymbol{\beta}, \boldsymbol{\Sigma}_{\beta\beta})$, $\mathbf{e}_i \sim N(\mathbf{0}, \sigma^2\mathbf{I})$, $i=1,\ldots,n$ units, $t$ observations each (Swamy 1971 setup).
• Individual OLS: $\mathbf{b}_i = (\mathbf{X}_i'\mathbf{X}_i)^{-1}\mathbf{X}_i'\mathbf{y}_i \sim N(\boldsymbol{\beta}, \mathbf{W}_i^{-1})$ where $\mathbf{W}_i^{-1} = \boldsymbol{\Sigma}_{\beta\beta} + \sigma^2(\mathbf{X}_i'\mathbf{X}_i)^{-1}$.
• Simple mean estimator: $\bar{\mathbf{b}} = n^{-1}\sum_i \mathbf{b}_i$ is exactly $N\!\left(\boldsymbol{\beta},\; n^{-1}\boldsymbol{\Sigma}_{\beta\beta} + \sigma^2 n^{-2}\sum_i (\mathbf{X}_i'\mathbf{X}_i)^{-1}\right)$ for all $n, t$.
• Variance estimator: $n^{-1}\mathbf{S}_{bb}$ where $\mathbf{S}_{bb} = (n-1)^{-1}\sum_i (\mathbf{b}_i - \bar{\mathbf{b}})(\mathbf{b}_i - \bar{\mathbf{b}})'$ is an unbiased estimator of $\text{var}(\bar{\mathbf{b}})$ — no positive-definiteness requirement.
• 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.
• Asymptotic equivalence: $\bar{\mathbf{b}}$ is asymptotically equivalent to $\hat{\boldsymbol{\beta}}_\text{GLS}$ and $\hat{\boldsymbol{\beta}}_\text{EGLS}$ as $t\to\infty$ (fixed $n$) and as $n\to\infty$ (fixed $t$).
• Monte Carlo (Table 1): When $\sigma^2=64$ and $n$ small, $\hat{\boldsymbol{\Sigma}}_{\beta\beta}$ is non-positive-definite in 51–81% of replications → EGLS test badly sized; $\hat{T}_{BAR}^{*2}$ empirical level within 0.02 of 5% in all configurations.
• Power (Table 3): For small $n$, $\hat{T}_{BAR}^{*2}$ can outperform $T^2_{EGLS}$; for large $n$, EGLS is more powerful. Both dominate OLS-based tests except when model matrices are nearly identical across units.
• Efficiency (Table 4): the mean relative efficiency (MRE) of $\bar{\mathbf{b}}$ is $\geq 90\%$ when $\delta=1$ (homogeneous $\mathbf{X}_i$); as low as 42% when $\delta=9$ (highly heterogeneous $\mathbf{X}_i$).

Concepts Introduced or Extended

• Random Coefficient Model — simple mean estimator $\bar{\mathbf{b}}$ as alternative to Swamy's EGLS; Hotelling $T^2$ test on individual OLS estimates; exact finite-sample normality.

Entities Mentioned

• Marcia Gumpertz
• Sastry G. Pantula

Quotes

"The main contribution of our article is to show that inferences can be made, without much difficulty, using the simple estimator $\bar{\mathbf{b}}$."

My Take

A clean, pedagogical note (The American Statistician) that makes one genuinely useful point: the sample variance of individual OLS slopes is an unbiased estimator of $\text{var}(\bar{\mathbf{b}})$ without any positive-definiteness constraint on $\hat{\boldsymbol{\Sigma}}_{\beta\beta}$, which is a real practical problem with the EGLS estimator in small samples. The Hotelling $T^2$ test on the $\mathbf{b}_i$'s is essentially treating them as a random sample from a multivariate normal — a multivariate analysis of variance (MANOVA) perspective that is both computationally trivial and well-calibrated. The efficiency loss is the trade-off: when units have very different design matrices ($\delta=9$), the equal-weight average wastes information that EGLS would exploit. Suitable guidance: use $\hat{T}_{BAR}^{*2}$ when $n$ is small, $\sigma^2/\|\boldsymbol{\Sigma}_{\beta\beta}\|$ is potentially large, or when EGLS is giving non-positive-definite estimates; switch to EGLS when $n$ is large enough for reliable variance component estimation.