Review Article
Volume 5 Issue 3 - February 2018
DOI: 10.19080/BBOAJ.2018.04.555662
Biostat Biometrics Open Acc J
Copyright © All rights are reserved by Yahia S El-Horbaty
Some Estimation Methods and Their Assessment
in Multilevel Models: A Review
Yahia S El-Horbaty* and Eman M Hanafy
Department of Mathematics, Insurance and Applied Statistics, Helwan University, Egypt
Submission: January 27, 2018; Published: February 27, 2018
*Corresponding author: Yahia S El-Horbaty, Department of Mathematics, Insurance, and Applied Statistics, Helwan University, Egypt;
Email:
Biostat Biometrics Open Acc J 5(3): BBOAJ.MS.ID.555662 (2018) 0069
Introduction
Multilevel modeling is an approach that can be used to
handle clustered or grouped data. Social sciences often involve
problems that investigate the relationship between individual
and society. The general concept is that individuals and the
social groups are conceptualized as a hierarchical system, where
the individuals and groups are defined at separate levels of this
hierarchical system. There are many types of multilevel models,
which differ in terms of the number of levels, type of design
(random intercept, random slopes and random coefficients
regression model), scale of the outcome variable (continuous,
categorical), and number of outcomes (univariate, multivariate).
In this article, the two-level random coefficients regression
model is represented, letting the univariate outcome variable to
be continuous.
The parameters to be estimated in multilevel regression
models are the fixed coefficients that represent the fixed
part of the model. The parameters also include the variance
covariance matrix of the random coefficients and the variances
of the residual errors that represent the random part of the
model. If the components of the random part in the model were
known, the unknown fixed coefficients could be estimated
using generalized least squares (GLS) estimation method [1].
Similarly, if the fixed coefficients were known, then the unknown
variance components of the model could also be estimated using
GLS estimation method. If both are unknown, hence we have to
follow different approaches.
In order to estimate the unknown parameters several
procedures are discussed in Searle et al. [2]. These methods
include the ANOVA method for balanced data which uses the
expected mean squares approach. However, this method is
difficult to deal with under unbalanced data situations. Under
the latter case, Rao [3] proposed the minimum norm quadratic
estimation (MINQUE) method for estimating the variance
parameters that produces quadratic unbiased estimators with
minimum norm (MINQUE). Maximum likelihood method (full
maximum likelihood (ML) and restricted maximum likelihood
(REML)), iterative generalized least squares (IGLS) and the
expectation maximization (EM) estimation method represent
standard methods that are used under both balanced and
unbalanced data. These methods involve iterative procedures
and thus the resulting estimators are not necessarily expressed
in a closed form.
The commonly used methods to estimate the multilevel
model are ML and REML [4]. The ML estimators of the variance
components do not correct for the degrees of freedom lost due
to the estimation of the fixed effects. As a result, the estimates
Biostatistics and Biometrics
Open Access Journal
ISSN: 2573-2633
Abstract
Multilevel linear regression models represent a generalization of linear models in which the regression coefficients are themselves given a
model whose parameters are also estimated from the data. This paper reviews multilevel random coefficients regression models with a focus on
the estimation problem and its assessment. Parameter estimation for the fixed effects and the variance components are highlighted. In addition,
comparisons that are made in the literature to choose among the competing methods are highlighted. This is particularly emphasized when some
of the assumptions underlying the estimation methods are violated.
Keywords: Multilevel models; Parametric estimation; Robust standard errors
Abbreviations: GLS: Generalized Least Squares; MINQUE: Minimum Norm Quadratic Estimation; ML: Maximum Likelihood; REML: Restricted
Maximum Likelihood; IGLS: Iterative Generalized Least Squares; EM: Expectation Maximization; OLS: Ordinary Least Squares; ARE: Asymptotic
Relative Efficiency