Skip to main content
SpringerLink
Log in

A Fay–Herriot model when auxiliary variables are measured with error

  • Original Paper
  • Published:
TEST Aims and scope Submit manuscript

Abstract

The Fay–Herriot model is an area-level linear mixed model that is widely used for estimating the domain means of a given target variable. Under this model, the dependent variable is a direct estimator calculated by using the survey data and the auxiliary variables are true domain means obtained from external data sources. Administrative registers do not always give good auxiliary variables so that statisticians sometimes take them from alternative surveys and therefore they are measured with error. We introduce a variant of the Fay–Herriot model that takes into account the measurement error of the auxiliary variables and give two fitting algorithms that calculate maximum and residual maximum likelihood estimates of the model parameters. Based on the new model, empirical best predictors of domain means are introduced and an approximation of its mean squared error is derived. We finally give an application to estimate poverty proportions in the Spanish Living Condition Survey, with auxiliary information from the Spanish Labour Force Survey.

This is a preview of subscription content, log in via an institution to check access.

Access this article

Log in via an institution

Price excludes VAT (USA)
Tax calculation will be finalised during checkout.

Instant access to the full article PDF.

Institutional subscriptions

Fig. 1
Fig. 2
Fig. 3
Fig. 4
Fig. 5
Fig. 6
Fig. 7
Fig. 8
Fig. 9
Fig. 10
Fig. 11
Fig. 12
Fig. 13

Similar content being viewed by others

References

  • Arima S, Datta GS, Liseo B (2012) Objective Bayesian analysis of a measurement error small area model. Bayesian Anal 7:363–384

    Article  MathSciNet  Google Scholar 

  • Arima S, Datta GS, Liseo B (2015) Bayesian estimators for small area models when auxiliary information is measured with error. Scand J Stat 42:518–529

    Article  MathSciNet  Google Scholar 

  • Arima S, Bell WR, Datta GS, Franco C, Liseo B (2017) Multivariate Fay–Herriot Bayesian estimation of small area means under functional measurement error. J R Stat Soc Ser A 180(4):1191–1209

    Article  MathSciNet  Google Scholar 

  • Benavent R, Morales D (2016) Multivariate Fay–Herriot models for small area estimation. Comput Stat Data Anal 94:372–390

    Article  MathSciNet  Google Scholar 

  • Boubeta M, Lombardía MJ, Morales D (2016) Empirical best prediction under area-level Poisson mixed models. TEST 25(3):548–569

    Article  MathSciNet  Google Scholar 

  • Borssoi JA, Paula GA, Galea M (2017) Elliptical linear mixed models with a covariate subject to measurement error. Stat Pap

  • Datta GS, Lahiri P (2000) A unified measure of uncertainty of estimated best linear unbiased predictors in small area estimation problems. Stat Sin 10:613–627

    MathSciNet  MATH  Google Scholar 

  • Datta GS, Rao JNK, Torabi M (2010) Pseudo-empirical Bayes estimation of small area means under a nested error linear regression model with functional measurement errors. J Stat Plan Inference 250:2952–2962

    Article  MathSciNet  Google Scholar 

  • Esteban MD, Morales D, Pérez A, Santamaría L (2011) Two area-level time models for estimating small area poverty indicators. J Indian Soc Agric Stat 66(1):75–89

    MathSciNet  Google Scholar 

  • Esteban MD, Morales D, Pérez A, Santamaría L (2012) Small area estimation of poverty proportions under area-level time models. Comput Stat Data Anal 56:2840–2855

    Article  MathSciNet  Google Scholar 

  • Fay RE, Herriot RA (1979) Estimates of income for small places: an application of James–Stein procedures to census data. J Am Stat Assoc 74(366):269–277

    Article  MathSciNet  Google Scholar 

  • Ghosh M, Sinha K, Kim D (2006) Empirical and hierarchical Bayesian estimation in finite population sampling under structural measurement error models. Scand J Stat 33:591–608

    Article  MathSciNet  Google Scholar 

  • Ghosh M, Sinha K (2007) Empirical Bayes estimation in finite population sampling under functional measurement error models. J Stat Plan Inference 137:2759–2773

    Article  MathSciNet  Google Scholar 

  • Hobza T, Morales D, Santamaría L (2018) Small area estimation of poverty proportions under unit-level temporal binomial-logit mixed models. TEST 27(2):270–294

    Article  MathSciNet  Google Scholar 

  • Marhuenda Y, Molina I, Morales D (2013) Small area estimation with spatio-temporal Fay–Herriot models. Comput Stat Data Anal 58:308–325

    Article  MathSciNet  Google Scholar 

  • Prasad NGN, Rao JNK (1990) The estimation of the mean squared error of small-area estimators. J Am Stat Assoc 85:163–171

    Article  MathSciNet  Google Scholar 

  • Rao JNK, Molina I (2015) Small area estimation, 2nd edn. Wiley, Hoboken

    Book  Google Scholar 

  • Rivest LP, Belmonte E (2000) A conditional mean squared error of small area estimators. Surv Methodol 26:67–78

    Google Scholar 

  • Särndal CE, Swensson B, Wretman J (1992) Model assisted survey sampling. Springer, New York

    Book  Google Scholar 

  • Torabi M, Datta GS, Rao JNK (2009) Empirical Bayes estimation of small area means under nested error linear regression model with measurement errors in the covariates. Scand J Stat 36:355–368

    Article  MathSciNet  Google Scholar 

  • Ybarra LMR, Lohr SL (2008) Small area estimation when auxiliary information is measured with error. Biometrika 95(4):919–931

    Article  MathSciNet  Google Scholar 

Download references

Author information

Authors and Affiliations

  1. Research Institute for Official and Survey Statistics (RIFOSS), Trier University, Trier, Germany

    Jan Pablo Burgard

  2. Operations Research Center, University Miguel Hernández de Elche, Elche, Spain

    María Dolores Esteban & Domingo Morales

  3. Departments of Economic and Financial Studies, University Miguel Hernández de Elche, Elche, Spain

    Agustín Pérez

Authors
  1. Jan Pablo Burgard
    View author publications

    You can also search for this author in PubMed Google Scholar

  2. María Dolores Esteban
    View author publications

    You can also search for this author in PubMed Google Scholar

  3. Domingo Morales
    View author publications

    You can also search for this author in PubMed Google Scholar

  4. Agustín Pérez
    View author publications

    You can also search for this author in PubMed Google Scholar

Corresponding author

Correspondence to Jan Pablo Burgard.

Additional information

Publisher's Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supported by the Spanish Grant MTM2015-64842-P and by the Grant “Algorithmic Optimization (ALOP) - graduate school 2126” funded by the German Research Foundation.

Appendices

Appendix A: Proof of Proposition 2.1

As the kernels of the univariate and multivariate normal distributions are

$$\begin{aligned} f(y|\mu ,\sigma ^2)= & {} \frac{1}{(2\pi \sigma ^2)^{1/2}}\exp \big \{-\frac{1}{2\sigma ^2}(y-\mu )^2)\big \}\propto \exp \big \{-\frac{1}{2\sigma ^2} y^2+\frac{\mu }{\sigma ^2}y\big \},\\ f(y|\mu ,\Sigma )= & {} \frac{1}{(2\pi )^{n/2}|\Sigma |^{1/2}}\exp \big \{-\frac{1}{2}(y-\mu )^{\prime }\Sigma ^{-1}(y-\mu )\big \}\\\propto & {} \exp \big \{-\frac{1}{2} y^{\prime }\Sigma ^{-1}y+\mu ^{\prime }\Sigma ^{-1}y\big \}, \end{aligned}$$

the following results holds.

Proof of (2.3). The conditional distribution of \(\varvec{v}_d\), given the estimators \(\varvec{x}_d\) and \(y_d\), is

$$\begin{aligned} f(\varvec{v}_d|\varvec{x}_d,y_{d}) \varpropto&f(y_d|\varvec{x}_d,\varvec{v}_d) f(\varvec{v}_d) =\frac{1}{\sqrt{2\pi (\sigma _u^2+\sigma _d^2)}}\\&\exp \Big \{-\frac{1}{2(\sigma _u^2 +\sigma _d^2)}(y_d-\varvec{x}_d\varvec{\beta }-\varvec{v}_d^{\prime }\varvec{\beta })^2\Big \} \\&\cdot \frac{1}{(2\pi )^{p/2}|\varvec{\Sigma }_d|^{1/2}}\exp \Big \{-\frac{1}{2} \varvec{v}_d^{\prime }\varvec{\Sigma }_d^{-1}\varvec{v}_d\Big \} \\ \varpropto&\exp \Big \{-\frac{1}{2(\sigma _u^2+\sigma _d^2)} \big (\varvec{v}_d^{\prime }\varvec{\beta }\varvec{\beta }^{\prime }\varvec{v}_d-2\varvec{v}_d^{\prime } \varvec{\beta }(y_d-\varvec{x}_d\varvec{\beta })\Big \}\\&\exp \Big \{-\frac{1}{2}\varvec{v}_d^{\prime } \varvec{\Sigma }_d^{-1}\varvec{v}_d\Big \} \\ =&\exp \Big \{-\frac{1}{2}\varvec{v}_d^{\prime }\Big (\frac{\varvec{\beta }\varvec{\beta }^{\prime }}{\sigma _u^2+\sigma _d^2}+\varvec{\Sigma }_d^{-1}\Big )\varvec{v}_d +\frac{\varvec{v}_d^{\prime }\varvec{\beta }(y_d-\varvec{x}_d\varvec{\beta })}{\sigma _u^2+\sigma _d^2} \Big \}. \end{aligned}$$

Therefore, \(f(\varvec{v}_d|\varvec{x}_d,y_{d})\) is a multivariate normal distribution with parameters

$$\begin{aligned} \varvec{\Psi }_d=\text{ var }(\varvec{v}_d|\varvec{x}_d,y_{d})=\Big (\frac{\varvec{\beta }\varvec{\beta }^{\prime }}{\sigma _u^2+\sigma _d^2}+\varvec{\Sigma }_d^{-1}\Big )^{-1},\quad E[\varvec{v}_d|\varvec{x}_d,y_{d}]=\frac{y_d-\varvec{x}_d\varvec{\beta }}{\sigma _u^2+\sigma _d^2}\varvec{\Psi }_d\varvec{\beta }. \end{aligned}$$

For calculating \(\varvec{\Psi }_d\), we apply the inversion formula

$$\begin{aligned} \big (A+uv^{\prime }\big )^{-1}=A^{-1}-\frac{A^{-1}uv^{\prime } A^{-1}}{1+v^{\prime } A^{-1}u}, \end{aligned}$$

with \(A=\varvec{\Sigma }_d^{-1}\), \(u=\frac{1}{\sigma _u^2+\sigma _d^2}\,\varvec{\beta }\), \(v^{\prime }=\varvec{\beta }^{\prime }\). We have

$$\begin{aligned} \varvec{\Psi }_d=\varvec{\Sigma }_d-\frac{\frac{1}{\sigma _u^2+\sigma _d^2}\,\varvec{\Sigma }_d \varvec{\beta }\varvec{\beta }^{\prime } \varvec{\Sigma }_d}{1+\frac{1}{\sigma _u^2+\sigma _d^2} \,\varvec{\beta }^{\prime } \varvec{\Sigma }_d\varvec{\beta }} =\varvec{\Sigma }_d-\frac{\varvec{\Sigma }_d\varvec{\beta }\varvec{\beta }^{\prime } \varvec{\Sigma }_d}{\sigma _u^2+\sigma _d^2+\varvec{\beta }^{\prime } \varvec{\Sigma }_d\varvec{\beta }}. \end{aligned}$$

Therefore

$$\begin{aligned} \hat{\varvec{v}}_d^{bp}= & {} E[\varvec{v}_d|\varvec{x}_d,y_{d}]=\frac{y_d-\varvec{x}_d\varvec{\beta }}{\sigma _u^2+\sigma _d^2}\Big (\varvec{\Sigma }_d-\frac{\varvec{\Sigma }_d\varvec{\beta }\varvec{\beta }^{\prime } \varvec{\Sigma }_d}{\sigma _u^2+\sigma _d^2+\varvec{\beta }^{\prime } \varvec{\Sigma }_d\varvec{\beta }}\Big )\varvec{\beta }\\= & {} \frac{y_d-\varvec{x}_d\varvec{\beta }}{\sigma _u^2+\sigma _d^2}\Big (\varvec{\Sigma }_d\varvec{\beta }-\frac{\varvec{\Sigma }_d\varvec{\beta }(\varvec{\beta }^{\prime }\varvec{\Sigma }_d\varvec{\beta })}{\sigma _u^2 +\sigma _d^2+\varvec{\beta }^{\prime } \varvec{\Sigma }_d\varvec{\beta }}\Big )\\= & {} \frac{y_d-\varvec{x}_d\varvec{\beta }}{\sigma _u^2+\sigma _d^2} \Big (1-\frac{\varvec{\beta }^{\prime } \varvec{\Sigma }_d\varvec{\beta }}{\sigma _u^2+\sigma _d^2 +\varvec{\beta }^{\prime } \varvec{\Sigma }_d\varvec{\beta }}\Big )\varvec{\Sigma }_d\varvec{\beta }\\= & {} \frac{y_d-\varvec{x}_d\varvec{\beta }}{\varvec{\beta }^{\prime }\varvec{\Sigma }_d\varvec{\beta }+\sigma _u^2+\sigma _d^2}\,\varvec{\Sigma }_d\varvec{\beta }. \end{aligned}$$

\(\square \)

Proof of (2.4). The conditional distribution of \(u_d\), given \(\varvec{x}_d\) and \(y_d\), is

$$\begin{aligned} f(u_d|\varvec{x}_d,y_{d})\varpropto&f(u_d)f(y_d|\varvec{x}_d,u_d)= \frac{1}{(2\pi \sigma _u^2)^{1/2}}\exp \Big \{-\frac{1}{2\sigma _u^2}u_d^2\Big \}\\&\cdot \frac{1}{\sqrt{2\pi (\varvec{\beta }^{\prime }\varvec{\Sigma }_d\varvec{\beta }+\sigma _d^2)}} \exp \Big \{-\frac{1}{2(\varvec{\beta }^{\prime }\varvec{\Sigma }_d\varvec{\beta }+\sigma _d^2)} (y_d-\varvec{x}_d\varvec{\beta }-u_d)^2\Big \} \\ \varpropto&\exp \Big \{-\frac{1}{2\sigma _u^2}u_d^2\Big \} \exp \Big \{-\frac{1}{2(\varvec{\beta }^{\prime }\varvec{\Sigma }_d\varvec{\beta }+\sigma _d^2)} \big (u_d^2-2u_d(y_d-\varvec{x}_d\varvec{\beta })\big )\Big \} \\ =&\exp \Big \{-\frac{1}{2}\Big (\frac{1}{\varvec{\beta }^{\prime } \varvec{\Sigma }_d\varvec{\beta }+\sigma _d^2}+\frac{1}{\sigma _u^2}\Big )u_d^2 +\frac{y_d-\varvec{x}_d\varvec{\beta }}{\varvec{\beta }^{\prime }\varvec{\Sigma }_d\varvec{\beta }+\sigma _d^2}\,u_d\Big \} \\ =&\exp \left\{ -\frac{1}{2}\frac{1}{\frac{\sigma _u^2\big (\varvec{\beta }^{\prime } \varvec{\Sigma }_d\varvec{\beta }+\sigma _d^2\big )}{\varvec{\beta }^{\prime }\varvec{\Sigma }_d\varvec{\beta }+\sigma _u^2 +\sigma _d^2}}\,u_d^2 + \frac{1}{\frac{\sigma _u^2\big (\varvec{\beta }^{\prime } \varvec{\Sigma }_d\varvec{\beta }+\sigma _d^2\big )}{\varvec{\beta }^{\prime }\varvec{\Sigma }_d\varvec{\beta }+\sigma _u^2+\sigma _d^2}} \frac{\sigma _u^2\big (y_d-\varvec{x}_d\varvec{\beta }\big )}{\varvec{\beta }^{\prime }\varvec{\Sigma }_d\varvec{\beta }+\sigma _u^2+\sigma _d^2} \,u_d\right\} . \end{aligned}$$

Therefore, \(f(u_d|\varvec{x}_d,y_d)\) is a univariate normal distribution with parameters

$$\begin{aligned} \text{ var }(u_d|\varvec{x}_d,y_{d})= & {} \frac{\sigma _u^2\big (\varvec{\beta }^{\prime } \varvec{\Sigma }_d\varvec{\beta }+\sigma _d^2\big )}{\varvec{\beta }^{\prime }\varvec{\Sigma }_d\varvec{\beta }+\sigma _u^2+\sigma _d^2},\\ \hat{u}_d^{bp}=E[u_d|\varvec{x}_d,y_{d}]= & {} \frac{\sigma _u^2}{\varvec{\beta }^{\prime } \varvec{\Sigma }_d\varvec{\beta }+\sigma _u^2 +\sigma _d^2}\big (y_d-\varvec{x}_d\varvec{\beta }\big ). \end{aligned}$$

\(\square \)

Proof of ( 2.5 ).

$$\begin{aligned} \hat{\mu }_d^{bp}= & {} \varvec{x}_d\varvec{\beta }+E[\varvec{v}_d^{\prime }|\varvec{x}_d,y_d]\varvec{\beta }+E[u_d|\varvec{x}_d,y_d]\\= & {} \varvec{x}_d\varvec{\beta }+\frac{\varvec{\beta }^{\prime }\varvec{\Sigma }_d\varvec{\beta }}{\varvec{\beta }^{\prime }\varvec{\Sigma }_d\varvec{\beta }+\sigma _u^2+\sigma _d^2}\,\big (y_d-\varvec{x}_d\varvec{\beta }\big ) +\frac{\sigma _u^2}{\varvec{\beta }^{\prime } \varvec{\Sigma }_d\varvec{\beta }+\sigma _u^2+\sigma _d^2}\big (y_d-\varvec{x}_d\varvec{\beta }\big ) \\= & {} \varvec{x}_d\varvec{\beta }+\frac{\sigma _u^2+\varvec{\beta }^{\prime }\varvec{\Sigma }_d\varvec{\beta }}{\varvec{\beta }^{\prime }\varvec{\Sigma }_d\varvec{\beta }+\sigma _u^2+\sigma _d^2}\,\big (y_d-\varvec{x}_d\varvec{\beta }\big )\\= & {} \frac{\sigma _u^2+\varvec{\beta }^{\prime }\varvec{\Sigma }_d\varvec{\beta }}{\varvec{\beta }^{\prime }\varvec{\Sigma }_d\varvec{\beta }+\sigma _u^2+\sigma _d^2}\,y_d +\frac{\sigma _d^2}{\varvec{\beta }^{\prime }\varvec{\Sigma }_d\varvec{\beta }+\sigma _u^2+\sigma _d^2}\,\varvec{x}_d\varvec{\beta }. \end{aligned}$$

Appendix B: MSE calculations

Under model (2.2), the mean squared error of \(\hat{\mu }^\mathrm{FHblup}_d\) is

$$\begin{aligned}&\text {MSE}(\hat{\mu }^\mathrm{FHblup}_d)\\&\quad =E\big [\big (\hat{\mu }^\mathrm{FHblup}_d-\mu _d\big )^2\big ] =E\big [\big (g_d y_d+(1-g_d)\varvec{x}_d\varvec{\beta }-\varvec{x}_d\varvec{\beta }-\varvec{v}_d^{\prime }\varvec{\beta }-u_d\big )^2\big ]\\&\quad =E\big [\big (g_d(y_d-\varvec{x}_d\varvec{\beta }-\varvec{v}_d^{\prime }\varvec{\beta }-u_d) -(1-g_d)(\varvec{v}_d^{\prime }\varvec{\beta }+u_d)\big )^2\big ] \\&\quad =E\big [\big (g_de_d+(1-g_d)(\varvec{v}_d^{\prime }\varvec{\beta }+u_d)\big )^2\big ] =g_d^2E[e_d^2]+(1-g_d)^2E[(\varvec{v}_d^{\prime }\varvec{\beta }+u_d)^2]\\&\quad =g_d^2\sigma _d^2+(1-g_d)^2(\varvec{\beta }^{\prime }\varvec{\Sigma }_d\varvec{\beta }+\sigma _u^2) = \frac{\sigma _u^4\sigma _d^2}{(\sigma _u^2+\sigma _d^2)^2} +\frac{(\varvec{\beta }^{\prime }\varvec{\Sigma }_d\varvec{\beta }+\sigma _u^2)\sigma _d^4}{(\sigma _u^2+\sigma _d^2)^2} \\&\quad = \frac{\sigma _u^2\sigma _d^2}{\sigma _u^2+\sigma _d^2} \left[ \frac{\sigma _u^2}{\sigma _u^2+\sigma _d^2} +\frac{\sigma _d^2}{\sigma _u^2+\sigma _d^2} \frac{\varvec{\beta }^{\prime }\varvec{\Sigma }_d\varvec{\beta }+\sigma _u^2}{\sigma _u^2}\right] \ge g_d\sigma _d^2 =\text {MSE}_{FH}(\hat{\mu }^\mathrm{FHblup}_d). \end{aligned}$$

The conditional MSE of \(\hat{\mu }^{bp}_d\) with known \(\beta \) and unknown \(\sigma _u^2\) is

$$\begin{aligned}&\text {MSE}(\hat{\mu }^{bp}_d|\varvec{X})\\&\quad =E\big [\big (\hat{\mu }^{bp}_d-\mu _d\big )^2\big ] =E\big [\big (\gamma _d y_d+(1-\gamma _d)\varvec{x}_d\varvec{\beta }-\varvec{x}_d\varvec{\beta }-\varvec{v}_d^{\prime } \varvec{\beta }-u_d\big )^2\big ] \\&\quad =E\big [\big (\gamma _d(y_d-\varvec{x}_d\varvec{\beta }-\varvec{v}_d^{\prime }\varvec{\beta }-u_d)-(1-\gamma _d) (\varvec{v}_d^{\prime }\varvec{\beta }+u_d)\big )^2\big ] \\&\quad =E\big [\big (\gamma _de_d+(1-\gamma _d)(\varvec{v}_d^{\prime }\varvec{\beta }+u_d)\big )^2\big ] =\gamma _d^2E[e_d^2]+(1-\gamma _d)^2E[(\varvec{v}_d^{\prime }\varvec{\beta }+u_d)^2]\\&\quad =\gamma _d^2\sigma _d^2+(1-\gamma _d)^2(\varvec{\beta }^{\prime }\varvec{\Sigma }_d\varvec{\beta }+\sigma _u^2) =\frac{(\varvec{\beta }^{\prime }\varvec{\Sigma }_d\varvec{\beta }+\sigma _u^2)^2\sigma _d^2}{(\varvec{\beta }^{\prime } \varvec{\Sigma }_d\varvec{\beta }+\sigma _u^2+\sigma _d^2)^2} +\frac{(\varvec{\beta }^{\prime }\varvec{\Sigma }_d\varvec{\beta }+\sigma _u^2)\sigma _d^4}{(\varvec{\beta }^{\prime }\varvec{\Sigma }_d\varvec{\beta }+\sigma _u^2+\sigma _d^2)^2}\\&\quad = \frac{(\varvec{\beta }^{\prime }\varvec{\Sigma }_d\varvec{\beta }+\sigma _u^2)\sigma _d^2(\varvec{\beta }^{\prime } \varvec{\Sigma }_d\varvec{\beta }+\sigma _u^2+\sigma _d^2)}{(\varvec{\beta }^{\prime }\varvec{\Sigma }_d\varvec{\beta }+\sigma _u^2 +\sigma _d^2)^2} = \frac{(\varvec{\beta }^{\prime }\varvec{\Sigma }_d\varvec{\beta }+\sigma _u^2)\sigma _d^2}{(\varvec{\beta }^{\prime }\varvec{\Sigma }_d\varvec{\beta }+\sigma _u^2+\sigma _d^2)}=\gamma _d\sigma _d^2. \end{aligned}$$

The covariance matrices appearing in the derivation of the MSE of the BP \(\hat{\mu }^{bp}_d\) are

$$\begin{aligned} \varvec{R}_{11}\triangleq&E\left[ (\hat{\varvec{\beta }}-\varvec{\beta })(\hat{\varvec{\beta }}-\varvec{\beta })^{\prime }\right] =\varvec{Q}\varvec{X}^{\prime }\varvec{V}^{-1}\varvec{V}\varvec{V}^{-1}\varvec{X}\varvec{Q}\\ =&\varvec{Q}\varvec{X}^{\prime }\varvec{V}^{-1}\varvec{X}\varvec{Q}=\varvec{Q}\varvec{Q}^{-1}\varvec{Q}=\varvec{Q}, \\ \varvec{R}_{12}\triangleq&E\big [(\hat{\varvec{\beta }}-\varvec{\beta })(\hat{\varvec{u}}-\varvec{u})^{\prime }\big ] =E\big [\hat{\varvec{\beta }}\hat{\varvec{u}}^{\prime }\big ]-E\big [\hat{\varvec{\beta }}\varvec{u}^{\prime }\big ]\\ =&E\big [\varvec{Q}\varvec{X}^{\prime }\varvec{V}^{-1}\varvec{y}\varvec{y}^{\prime }(\varvec{I}-\varvec{V}^{-1}\varvec{X}\varvec{Q}\varvec{X}^{\prime }) \varvec{V}^{-1}\varvec{V}_u\big ] -E\big [\varvec{Q}\varvec{X}^{\prime }\varvec{V}^{-1}\varvec{y}\varvec{u}^{\prime }\big ]\\ =&\varvec{Q}\varvec{X}^{\prime }\varvec{V}^{-1}\varvec{V}(\varvec{I}-\varvec{V}^{-1}\varvec{X}\varvec{Q}\varvec{X}^{\prime })\varvec{V}^{-1}\varvec{V}_u -\varvec{Q}\varvec{X}^{\prime }\varvec{V}^{-1}\varvec{V}_u\\ =&-\varvec{Q}(\varvec{X}^{\prime }\varvec{V}^{-1}\varvec{X})\varvec{Q}\varvec{X}^{\prime }\varvec{V}^{-1}\varvec{V}_u = -\varvec{Q}\varvec{X}^{\prime }\varvec{V}^{-1}\varvec{V}_u, \\ \varvec{R}_{13}\triangleq&E\big [(\hat{\varvec{\beta }}-\varvec{\beta })(\hat{\varvec{w}}-\varvec{w})^{\prime }\big ] =-\varvec{Q}\varvec{X}^{\prime }\varvec{V}^{-1}\underset{1\le d \le D}{\hbox {diag}}(\varvec{\beta }^{\prime } \varvec{\Sigma }_d\varvec{\beta })\\ \varvec{R}_{22}\triangleq&E\big [(\hat{\varvec{u}}-\varvec{u})(\hat{\varvec{u}}-\varvec{u})^{\prime }\big ] = E\big [\hat{\varvec{u}}\hat{\varvec{u}}^{\prime }\big ]-E\big [\hat{\varvec{u}}\varvec{u}^{\prime }\big ] -E\big [\varvec{u}\hat{\varvec{u}}^{\prime }\big ]+E\big [\varvec{u}\varvec{u}^{\prime }\big ]\\ =&\varvec{V}_u\varvec{V}^{-1}(\varvec{I}-\varvec{X}\varvec{Q}\varvec{X}^{\prime }\varvec{V}^{-1})\varvec{V}(\varvec{I}-\varvec{V}^{-1} \varvec{X}\varvec{Q}\varvec{X}^{\prime })\varvec{V}^{-1}\varvec{V}_u\\&- \varvec{V}_u\varvec{V}^{-1}(\varvec{I}-\varvec{X}\varvec{Q}\varvec{X}^{\prime }\varvec{V}^{-1})\varvec{V}_u -\varvec{V}_u(\varvec{I}-\varvec{V}^{-1}\varvec{X}\varvec{Q}\varvec{X}^{\prime })\varvec{V}^{-1}\varvec{V}_u+\varvec{V}_u\\ =&\varvec{V}_u-\varvec{V}_u\varvec{V}^{-1}\varvec{V}_u+\varvec{V}_u\varvec{V}^{-1}\varvec{X}\varvec{Q}\varvec{X}^{\prime }\varvec{V}^{-1}\varvec{V}_u,\\ \varvec{R}_{33}\triangleq&E\big [(\hat{\varvec{w}}-\varvec{w})(\hat{\varvec{w}}-\varvec{w})^{\prime }\big ] = E\big [\hat{\varvec{w}}\hat{\varvec{w}}^{\prime }\big ]-E\big [\hat{\varvec{w}}\varvec{w}^{\prime }\big ] -E\big [\varvec{w}\hat{\varvec{w}}^{\prime }\big ]+E\big [\varvec{w}\varvec{w}^{\prime }\big ]\\ =&\underset{1\le d \le D}{\hbox {diag}}(\varvec{\beta }^{\prime }\varvec{\Sigma }_d\varvec{\beta }) -\underset{1\le d \le D}{\hbox {diag}}(\varvec{\beta }^{\prime }\varvec{\Sigma }_d\varvec{\beta }) \varvec{V}^{-1}\underset{1\le d \le D}{\hbox {diag}}(\varvec{\beta }^{\prime }\varvec{\Sigma }_d\varvec{\beta })\\&+\underset{1\le d \le D}{\hbox {diag}}(\varvec{\beta }^{\prime }\varvec{\Sigma }_d\varvec{\beta })\varvec{V}^{-1} \varvec{X}\varvec{Q}\varvec{X}^{\prime }\varvec{V}^{-1}\underset{1\le d \le D}{\hbox {diag}}(\varvec{\beta }^{\prime } \varvec{\Sigma }_d\varvec{\beta }),\\ \varvec{R}_{23}\triangleq&E\big [(\hat{\varvec{u}}-\varvec{u})(\hat{\varvec{w}}-\varvec{w})^{\prime }\big ] = E\big [\hat{\varvec{u}}\hat{\varvec{w}}^{\prime }\big ]-E\big [\hat{\varvec{u}}\varvec{w}^{\prime }\big ] -E\big [\varvec{u}\hat{\varvec{w}}^{\prime }\big ]\\ =&\varvec{V}_u\varvec{V}^{-1}(\varvec{I}-\varvec{X}\varvec{Q}\varvec{X}^{\prime }\varvec{V}^{-1})\varvec{V}(\varvec{I}-\varvec{V}^{-1} \varvec{X}\varvec{Q}\varvec{X}^{\prime })\varvec{V}^{-1}\underset{1\le d \le D}{\hbox {diag}} (\varvec{\beta }^{\prime }\varvec{\Sigma }_d\varvec{\beta })\\&-\varvec{V}_u\varvec{V}^{-1}(\varvec{I}-\varvec{X}\varvec{Q}\varvec{X}^{\prime }\varvec{V}^{-1}) \underset{1\le d \le D}{\hbox {diag}}(\varvec{\beta }^{\prime }\varvec{\Sigma }_d\varvec{\beta })\\&-\varvec{V}_u(\varvec{I}-\varvec{V}^{-1}\varvec{X}\varvec{Q}\varvec{X}^{\prime })\varvec{V}^{-1}\underset{1\le d \le D}{\hbox {diag}}(\varvec{\beta }^{\prime }\varvec{\Sigma }_d\varvec{\beta })\\ =&-\varvec{V}_u\varvec{V}^{-1}\underset{1\le d \le D}{\hbox {diag}} (\varvec{\beta }^{\prime }\varvec{\Sigma }_d\varvec{\beta }) +\varvec{V}_u\varvec{V}^{-1}\varvec{X}\varvec{Q}\varvec{X}^{\prime }\varvec{V}^{-1} \underset{1\le d \le D}{\hbox {diag}}(\varvec{\beta }^{\prime }\varvec{\Sigma }_d\varvec{\beta }). \end{aligned}$$

Therefore, we have

$$\begin{aligned} \text {MSE}(\hat{\mu }_d^{bp}|\varvec{X})= & {} \varvec{a}_d^{\prime }\varvec{X}\varvec{R}_{11}\varvec{X}^{\prime }\varvec{a}_d +\varvec{a}_d^{\prime }\varvec{X}\varvec{R}_{12}\varvec{a}_d+\varvec{a}_d^{\prime }\varvec{X}\varvec{R}_{13}\varvec{a}_d +\varvec{a}_d^{\prime }\varvec{R}_{21}\varvec{X}^{\prime }\varvec{a}_d \\&+\varvec{a}_d^{\prime }\varvec{R}_{22}\varvec{a}_d+\varvec{a}_d^{\prime }\varvec{R}_{23}\varvec{a}_d +\varvec{a}_d^{\prime }\varvec{R}_{31}\varvec{X}^{\prime }\varvec{a}_d+\varvec{a}_d^{\prime }\varvec{R}_{32}\varvec{a}_d +\varvec{a}_d^{\prime }\varvec{R}_{33}\varvec{a}_d \\= & {} \varvec{x}_d\varvec{Q}\varvec{x}_d^{\prime }-\varvec{x}_d\varvec{Q}\varvec{x}_d^{\prime }\frac{\sigma _u^2}{b_d} -\varvec{x}_d\varvec{Q}\varvec{x}_d^{\prime }\frac{a_d}{b_d} -\varvec{x}_d\varvec{Q}\varvec{x}_d^{\prime } \frac{\sigma _u^2}{b_d}\\&+\Big (\sigma _u^2-\frac{\sigma _u^4}{b_d} +\frac{\sigma _u^4}{b_d^2}\varvec{x}_d\varvec{Q}\varvec{x}_d^{\prime }\Big ) \\&+ \Big (-\frac{\sigma _u^2a_d}{b_d}+\frac{\sigma _u^2a_d}{b_d^2} \varvec{x}_d\varvec{Q}\varvec{x}_d^{\prime }\Big ) -\varvec{x}_d\varvec{Q}\varvec{x}_d^{\prime }\frac{a_d}{b_d}\\&+\Big (-\frac{\sigma _u^2a_d}{b_d}+\frac{\sigma _u^2a_d}{b_d^2} \varvec{x}_d\varvec{Q}\varvec{x}_d^{\prime }\Big )\\+ & {} \Big (a_d-\frac{a_d^2}{b_d}+\frac{a_d^2}{b_d^2}\varvec{x}_d\varvec{Q}\varvec{x}_d^{\prime }\Big ). \end{aligned}$$

Further simplification yields:

$$\begin{aligned} \text {MSE}(\hat{\mu }_d^{bp}|\varvec{X})= & {} g_{1d}(\varvec{\beta },\sigma _u^2)+g_{2d}(\varvec{\beta },\sigma _u^2),\\ g_{1d}(\varvec{\beta },\sigma _u^2)= & {} \sigma _u^2-\frac{\sigma _u^4}{b_d}-2 \frac{\sigma _u^2a_d}{b_d}+a_d-\frac{a_d^2}{b_d} = \sigma _u^2+a_d-\frac{\sigma _u^4+a_d^2+2\sigma _u^2a_d}{b_d}\\= & {} (\sigma _u^2+a_d)-\frac{(\sigma _u^2+a_d)^2}{(\sigma _u^2+a_d)+\sigma _d^2} =\frac{\sigma _d^2(\sigma _u^2+a_d)}{\sigma _u^2+a_d+\sigma _d^2}=\gamma _d\sigma _d^2,\\ g_{2d}(\varvec{\beta },\sigma _u^2)= & {} \varvec{x}_d\varvec{Q}\varvec{x}_d^{\prime }\Big (1-2\frac{\sigma _u^2}{b_d} -2\frac{a_d}{b_d}+\frac{\sigma _u^4}{b_d^2}+\frac{a_d^2}{b_d^2} +2\frac{\sigma _u^2a_d}{b_d^2}\Big ) \\= & {} \varvec{x}_d\varvec{Q}\varvec{x}_d^{\prime }\frac{1}{b_d^2}\big [b_d^2-2\sigma _u^2b_d -2a_db_d+\sigma _u^4+a_d^2+2\sigma _u^2a_d\big ] \\= & {} \varvec{x}_d\varvec{Q}\varvec{x}_d^{\prime }\frac{1}{b_d^2}\big [(\sigma _u^4+a_d^2 +\sigma _d^4+2\sigma _u^2\sigma _d^2+2\sigma _u^2a_d+2a_d\sigma _d^2)\\&-2(\sigma _u^4+\sigma _u^2a_d+\sigma _u^2\sigma _d)\\&-2(a_d\sigma _u^2-a_d^2-a_d\sigma _d^2)+\sigma _u^4+a_d^2+2\sigma _u^2a_d\big ]\\= & {} \frac{\sigma _d^4}{(a_d+\sigma _u^2+\sigma _d^2)^2}\,\varvec{x}_d\varvec{Q}\varvec{x}_d^{\prime }. \end{aligned}$$

Appendix C

See Tables 4 and 5.

Table 4 Province poverty proportions for men (top) and women (bottom) based on the SLCS of 2008
Full size table
Table 5 Province means (left) and std.errors (right) of covariates for men (top) and women (bottom) based on the SLFS of 2008
Full size table

Rights and permissions

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Burgard, J.P., Esteban, M.D., Morales, D. et al. A Fay–Herriot model when auxiliary variables are measured with error. TEST 29, 166–195 (2020). https://doi.org/10.1007/s11749-019-00649-3

Download citation

  • Received:

  • Accepted:

  • Published:

  • Issue Date:

  • DOI: https://doi.org/10.1007/s11749-019-00649-3

Keywords

Mathematics Subject Classification

Search

Navigation