Abstract
In the mean regression context, this study considers several frequently encountered heteroscedastic error models where the regression mean and variance functions are specified up to certain parameters. An important point we note through a series of analyses is that different assumptions on standardized regression errors yield quite different efficiency bounds for the corresponding estimators. Consequently, all aspects of the assumptions need to be specifically taken into account in constructing their corresponding efficient estimators. This study clarifies the relation between the regression error assumptions and their, respectively, efficiency bounds under the general regression framework with heteroscedastic errors. Our simulation results support our findings; we carry out a real data analysis using the proposed methods where the Cobb–Douglas cost model is the regression mean.
Similar content being viewed by others
References
Bement, T. R., Williams, J. S. (1969). Variance of weighted regression estimators when sampling errors are independent and heteroscedastic. Journal of the American Statistical Association, 64, 1369–1382.
Bickel, P. J., Klaassen, C. A., Ritov, Y., Wellner, J. A. (1998). Efficient and adaptive estimation for semiparametric models. New York: Springer.
Carroll, R. J. (1982). Adapting for heteroscedasticity in linear models. The Annals of Statistics, 10, 1224–1233.
Carroll, R. J., Ruppert, D. (1982). Robust estimation in heteroscedastic linear models. The Annals of Statistics, 10, 429–441.
Fang, Y., Lian, H., Liang, H., Ruppert, D. (2015). Variance function additive partial linear models. Electronic Journal of Statistics, 9, 2793–2827.
Fuller, W. A., Rao, J. (1978). Estimation for a linear regression model with unknown diagonal covariance matrix. The Annals of Statistics, 6, 1149–1158.
Hall, P., Carroll, R. (1989). Variance function estimation in regression: The effect of estimating the mean. Journal of the Royal Statistical Society: Series B (Methodological), 51, 3–14.
Jacquez, J. A., Mather, F. J., Crawford, C. R. (1968). Linear regression with non-constant, unknown error variances: Sampling experiments with least squares, weighted least squares and maximum likelihood estimators. Biometrics, 24, 607–626.
Joanes, D., Gill, C. (1998). Comparing measures of sample skewness and kurtosis. Journal of the Royal Statistical Society: Series D (The Statistician), 47, 183–189.
Kim, M., Ma, Y. (2012). The efficiency of the second-order nonlinear least squares estimator and its extension. Annals of the Institute of Statistical Mathematics, 64, 751–764.
Lian, H., Liang, H., Carroll, R. J. (2015). Variance function partially linear single-index models. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 77, 171–194.
Ma, Y., Chiou, J.-M., Wang, N. (2006). Efficient semiparametric estimator for heteroscedastic partially linear models. Biometrika, 93, 75–84.
Müller, H.-G., Zhao, P.-L. (1995). On a semiparametric variance function model and a test for heteroscedasticity. The Annals of Statistics, 23, 946–967.
Nadarajah, S. (2005). A generalized normal distribution. Journal of Applied Statistics, 32, 685–694.
Nerlove, M. (1963). Returns to scale in electricity supply. En “Measurement in economics-studies in mathematical economics and econometrics in memory of Yehuda Grunfeld” edited by Carl F. Christ. Palo Alto: Stanford University Press.
Tsiatis, A. (2006). Semiparametric theory and missing data. New York: Springer.
Acknowledgements
Mijeong Kim was supported by a Ewha Womans University Research Grant of 2015 and a National Research Foundation of Korea (NRF) grant funded by the Korean Government (NRF-2017R1C1B5015186). Yanyuan Ma was supported by National Science Foundation DMS-1608540.
Author information
Authors and Affiliations
Corresponding author
Electronic supplementary material
Below is the link to the electronic supplementary material.
Appendix
Appendix
1.1 Proof of Proposition 1
First, note that the construction of t ensures the property
which is crucial for the following proof.
Following the semiparametric consideration in (1), we estimate the parameters \({\varvec{\theta }}\), while the nuisance parameter is the distribution \(f_{\epsilon \mid \mathbf{X}}(\epsilon ,\mathbf{x})\), which is subject to the constraints \(E(\epsilon |\mathbf{X})=0\) and \(E(\epsilon ^2|\mathbf{X})=1\). Note that we do not assume any parametric distribution model on \(\epsilon \). Thus, (1) is a semiparametric model. We first write out the joint distribution of \((\epsilon ,\mathbf{X})\)
where \(\eta _1(\cdot )\) and \(\eta _2(\cdot )\) denote the infinite-dimensional nuisance parameter function. Since \(\eta _1(\cdot )\) and \(\eta _2(\cdot )\) are probability density functions, we have
In the Hilbert space \(\mathcal {H}\) formed by all the mean zero finite variance functions, the nuisance tangent space of a semiparametric model is the mean squared closure of all the parametric submodel nuisance tangent spaces. A parametric submodel is defined as a parametric model included by the original semiparametric model and includes the true density \(f(\epsilon ,\mathbf{X};{\varvec{\theta }}_0,{\varvec{\gamma }}_0)\). A nuisance tangent space of the parametric model \(f(\epsilon ,\mathbf{X};{\varvec{\theta }},{\varvec{\gamma }})\) is a linear space spanned by the nuisance score vector \(S_{{\varvec{\gamma }}}=\partial \mathrm{log}f(\epsilon ,\mathbf{X};{\varvec{\theta }}_0,{\varvec{\gamma }})/\partial {\varvec{\gamma }}|_{{\varvec{\gamma }}_0}\). From the above general concept, we first derive the nuisance tangent space \(\Lambda \). From Condition (2), we have
From (15) and (16), it follows that
In the above, we use
Combining \(\Lambda _1\) and \(\Lambda _2\), the nuisance tangent space \(\Lambda \) can be written as
In order to find \(\Lambda ^\perp \), we let \(\mathbf{K}=\{\mathbf{g}(\mathbf{x},\epsilon ):\mathbf{g}(\mathbf{x},\epsilon )=\mathbf{g}_1(\mathbf{x})\epsilon +\mathbf{g}_2(\mathbf{x})t\}\) and have \(\Lambda ^\perp =\mathbf{K}\) by showing that \(\mathbf{K}\subset \Lambda ^\perp \) and \(\Lambda ^\perp \subset \mathbf{K}\).
For arbitrary \(\mathbf{h}(\mathbf{x},\epsilon )\in \Lambda \) and \(\mathbf{g}(\mathbf{x},\epsilon )=\mathbf{g}_1(\mathbf{x})\epsilon +\mathbf{g}_2(\mathbf{x})t\in \mathbf{K}\),
Hence, \(\mathbf{g}(\mathbf{x},\epsilon )=\mathbf{g}_1(\mathbf{x})\epsilon +\mathbf{g}_2(\mathbf{x})t\in \Lambda ^\perp \). Thus, \(\mathbf{K}\subset \Lambda ^\perp \).
Now, we will show \(\Lambda ^\perp \subset \mathbf{K}\).
For arbitrary \(\mathbf{h}(\mathbf{x},\epsilon )\in \Lambda ^\perp \), we can decompose \(\mathbf{h}(\mathbf{x},\epsilon )\) into \( \mathbf{h}(\mathbf{x},\epsilon )=\mathbf{r}_1(\mathbf{x},\epsilon )+\mathbf{r}_2(\mathbf{x},\epsilon ), \) where
It is obvious that \(\mathbf{r}_2(\mathbf{x},\epsilon )\in \mathbf{K}\). Since \(\mathbf{r}_2(\mathbf{x},\epsilon )\in \mathbf{K}\subset \Lambda ^\perp \) and \(\mathbf{h}(\mathbf{x},\epsilon )\in \Lambda ^\perp \), we have \(\mathbf{r}_1(\mathbf{x},\epsilon )=\mathbf{h}(\mathbf{x},\epsilon )-\mathbf{r}_2(\mathbf{x},\epsilon )\in \Lambda ^\perp \). For \(\mathbf{r}_1\), we can easily verify that
This indicates that \(\mathbf{r}_1(\mathbf{x},\epsilon )\in \Lambda \). Since \(\mathbf{r}_1(\mathbf{x},\epsilon )\) is also an element of \(\Lambda ^\perp \), it follows that \(\mathbf{r}_1(\mathbf{x},\epsilon )=\mathbf{0}\), and we obtain \(\mathbf{h}(\mathbf{x},\epsilon )=\mathbf{r}_2(\mathbf{x},\epsilon )\in \mathbf{K}\). We next show that \(\mathbf{h}(\mathbf{x},\epsilon )\in \mathbf{K}\) for arbitrary \(\mathbf{h}(\mathbf{x},\epsilon )\in \Lambda ^\perp \). Thus, \(\Lambda ^\perp \subset \mathbf{K}\).
Consequently, we have
\(\square \)
1.2 Proof of Theorem 1
In model (1), we obtain the score functions of \({\varvec{\theta }}=({\varvec{\alpha }}^\mathrm{T},{\varvec{\beta }}^\mathrm{T})^\mathrm{T}\) as
By projecting the above score vectors onto \(\Lambda ^\perp \), we can find the efficient score vectors of \({\varvec{\alpha }}\) and \({\varvec{\beta }}\).
First, \({\varvec{S}}_{\mathrm{eff},{\varvec{\alpha }}}\) and \({\varvec{S}}_{\mathrm{eff},{\varvec{\beta }}}\) are the form of \(\mathbf{g}_1(\mathbf{x})\epsilon +\mathbf{g}_2(\mathbf{x})t\). Thus, \({\varvec{S}}{_\mathrm{eff}}\in \Lambda ^\perp \).
Now, we verify that \({\varvec{S}}_{{\varvec{\theta }}}-{\varvec{S}}{_\mathrm{eff}}\in \Lambda \). \({\varvec{S}}_{{\varvec{\theta }}}-{\varvec{S}}{_\mathrm{eff}}\) is given by
We verify \({\varvec{S}}_{{\varvec{\theta }}}-{\varvec{S}}{_\mathrm{eff}}\in \Lambda \) by showing the following.
We can check the details as follows.
The above results imply that \(E({\varvec{S}}_{{\varvec{\theta }}}-{\varvec{S}}{_\mathrm{eff}})=\mathbf{0}\).
The above equations verify that \({\varvec{S}}_{{\varvec{\theta }}}-{\varvec{S}}{_\mathrm{eff}}\in \Lambda \).
Now, we find the optimal efficiency matrix. First, we calculate \({\varvec{S}}_\mathrm{eff}{\varvec{S}}^\mathrm{T}_\mathrm{eff}\).
where
Since \(E({\varvec{S}}_\mathrm{eff}{\varvec{S}}^\mathrm{T}_\mathrm{eff})=E\{E({\varvec{S}}{_\mathrm{eff}}{\varvec{S}}{_\mathrm{eff}}^\mathrm{T}|\mathbf{X})\}\), we have each block inside \(E({\varvec{S}}_\mathrm{eff}{\varvec{S}}^\mathrm{T}_\mathrm{eff})\) as follows.
-
1.
\(E({\varvec{S}}_{\mathrm{eff},{\varvec{\alpha }}}{\varvec{S}}_{\mathrm{eff},{\varvec{\alpha }}}^\mathrm{T})\) is equivalent to
$$\begin{aligned} E\left( E\left[ \left\{ \epsilon - \frac{E(\epsilon ^3|\mathbf{X})}{E(t^2|\mathbf{X})}t \right\} ^2|\mathbf{X}\right] e^{-2\sigma (\mathbf{X},{\varvec{\beta }})}\mathbf{m}'_{{\varvec{\alpha }}}(\mathbf{X},{\varvec{\alpha }}){\mathbf{m}'}^\mathrm{T}_{{\varvec{\alpha }}}(\mathbf{X},{\varvec{\alpha }})\right) . \end{aligned}$$From the above expression, we can rewrite the part \(E\left[ \left\{ \epsilon - \frac{E(\epsilon ^3|\mathbf{X})}{E(t^2|\mathbf{X})}t \right\} ^2|\mathbf{X}\right] \) as
$$\begin{aligned} { E\left[ \left\{ \epsilon - \frac{E(\epsilon ^3|\mathbf{X})}{E(t^2|\mathbf{X})}t \right\} ^2|\mathbf{X}\right] =1+\frac{\{E(\epsilon ^3|\mathbf{X})\}^2}{E(t^2|\mathbf{X})}}. \end{aligned}$$Thus, \(E({\varvec{S}}_{\mathrm{eff},{\varvec{\alpha }}}{\varvec{S}}_{\mathrm{eff},{\varvec{\alpha }}}^\mathrm{T})\) becomes
$$\begin{aligned} E\left( \left[ 1+\frac{\{E(\epsilon ^3|\mathbf{X})\}^2}{E(t^2|\mathbf{X})}\right] e^{-2\sigma (\mathbf{X},{\varvec{\beta }})}\mathbf{m}'_{{\varvec{\alpha }}}(\mathbf{X},{\varvec{\alpha }}){\mathbf{m}'}^\mathrm{T}_{{\varvec{\alpha }}}(\mathbf{X},{\varvec{\alpha }})\right) . \end{aligned}$$ -
2.
\(E({\varvec{S}}_{\mathrm{eff},{\varvec{\alpha }}}{\varvec{S}}_{\mathrm{eff},{\varvec{\beta }}}^\mathrm{T})\) can be rewritten as
$$\begin{aligned}&E\left[ E\left\{ \epsilon t- \frac{E(\epsilon ^3|\mathbf{X})}{E(t^2|\mathbf{X})}t^2 |\mathbf{X}\right\} \frac{2}{E(t^2|\mathbf{X})}e^{-\sigma (\mathbf{X},{\varvec{\beta }})} \mathbf{m}'_{{\varvec{\alpha }}} (\mathbf{X},{\varvec{\alpha }}){{\varvec{\sigma }}'}^\mathrm{T}_{\varvec{\beta }}(\mathbf{X},{\varvec{\beta }}) \right] \\&\quad =E\left\{ -\frac{2E(\epsilon ^3|\mathbf{X})}{E(t^2|\mathbf{X})}e^{-\sigma (\mathbf{X},{\varvec{\beta }})} \mathbf{m}'_{{\varvec{\alpha }}} (\mathbf{X},{\varvec{\alpha }}){{\varvec{\sigma }}'}^\mathrm{T}_{\varvec{\beta }}(\mathbf{X},{\varvec{\beta }}) \right\} . \end{aligned}$$ -
3.
\(E({\varvec{S}}_{\mathrm{eff},{\varvec{\beta }}}{\varvec{S}}_{\mathrm{eff},{\varvec{\beta }}}^\mathrm{T})\) is equivalent to
$$\begin{aligned} E\left( \frac{4E(t^2|\mathbf{X})}{\{E(t^2|\mathbf{X})\}^2}{\varvec{\sigma }}'_{\varvec{\beta }}(\mathbf{X},{\varvec{\beta }}){{\varvec{\sigma }}'}^\mathrm{T}_{\varvec{\beta }}(\mathbf{X},{\varvec{\beta }})\right) =E\left\{ \frac{4}{E(t^2|\mathbf{X})}{\varvec{\sigma }}'_{\varvec{\beta }}(\mathbf{X},{\varvec{\beta }}){{\varvec{\sigma }}'}^\mathrm{T}_{\varvec{\beta }}(\mathbf{X},{\varvec{\beta }})\right\} . \end{aligned}$$
Theorem 1 is the immediate consequence of the above calculations. \(\square \)
About this article
Cite this article
Kim, M., Ma, Y. Semiparametric efficient estimators in heteroscedastic error models. Ann Inst Stat Math 71, 1–28 (2019). https://doi.org/10.1007/s10463-017-0622-0
Received:
Revised:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s10463-017-0622-0