Effect of Violation of the Normal Assumption on MI and ML Estimators in the Analysis of Incomplete Data

Shintaro Hojo, Michio Yamamoto, Yutaka Kano

Research output: Contribution to journalArticle

Abstract

Asymptotic distributions of normal-theory-based ML/MI estimators are studied in a simple regression model under general distributions with MAR missing data. The asymptotic variance of the ML/MI estimator of residuals variance is explicitly derived, from which it follows that the kurtosis of the error distribution primarily affects the asymptotic variance. Results of numerical simulations conducted to study finite sample properties of the estimators, conformed largely to the asymptotic results, and they also indicated interesting findings particularly for small samples, which do not follow from the asymptotic property. It is concluded that the ML estimators perform best in the situation studied here.

Original languageEnglish
Pages (from-to)3234-3250
Number of pages17
JournalCommunications in Statistics - Theory and Methods
Volume44
Issue number15
DOIs
Publication statusPublished - Aug 3 2015
Externally publishedYes

Fingerprint

Incomplete Data
Estimator
Asymptotic Variance
Mars
Kurtosis
Missing Data
Small Sample
Asymptotic distribution
Asymptotic Properties
Regression Model
Numerical Simulation

Keywords

  • Distribution misspecification
  • Incomplete data
  • Maximum likelihood
  • Multiple imputation

ASJC Scopus subject areas

  • Statistics and Probability

Cite this

Effect of Violation of the Normal Assumption on MI and ML Estimators in the Analysis of Incomplete Data. / Hojo, Shintaro; Yamamoto, Michio; Kano, Yutaka.

In: Communications in Statistics - Theory and Methods, Vol. 44, No. 15, 03.08.2015, p. 3234-3250.

Research output: Contribution to journalArticle

@article{a896750add4847e1ba912b756363c05a,
title = "Effect of Violation of the Normal Assumption on MI and ML Estimators in the Analysis of Incomplete Data",
abstract = "Asymptotic distributions of normal-theory-based ML/MI estimators are studied in a simple regression model under general distributions with MAR missing data. The asymptotic variance of the ML/MI estimator of residuals variance is explicitly derived, from which it follows that the kurtosis of the error distribution primarily affects the asymptotic variance. Results of numerical simulations conducted to study finite sample properties of the estimators, conformed largely to the asymptotic results, and they also indicated interesting findings particularly for small samples, which do not follow from the asymptotic property. It is concluded that the ML estimators perform best in the situation studied here.",
keywords = "Distribution misspecification, Incomplete data, Maximum likelihood, Multiple imputation",
author = "Shintaro Hojo and Michio Yamamoto and Yutaka Kano",
year = "2015",
month = "8",
day = "3",
doi = "10.1080/03610926.2013.819920",
language = "English",
volume = "44",
pages = "3234--3250",
journal = "Communications in Statistics - Theory and Methods",
issn = "0361-0926",
publisher = "Taylor and Francis Ltd.",
number = "15",

}

TY - JOUR

T1 - Effect of Violation of the Normal Assumption on MI and ML Estimators in the Analysis of Incomplete Data

AU - Hojo, Shintaro

AU - Yamamoto, Michio

AU - Kano, Yutaka

PY - 2015/8/3

Y1 - 2015/8/3

N2 - Asymptotic distributions of normal-theory-based ML/MI estimators are studied in a simple regression model under general distributions with MAR missing data. The asymptotic variance of the ML/MI estimator of residuals variance is explicitly derived, from which it follows that the kurtosis of the error distribution primarily affects the asymptotic variance. Results of numerical simulations conducted to study finite sample properties of the estimators, conformed largely to the asymptotic results, and they also indicated interesting findings particularly for small samples, which do not follow from the asymptotic property. It is concluded that the ML estimators perform best in the situation studied here.

AB - Asymptotic distributions of normal-theory-based ML/MI estimators are studied in a simple regression model under general distributions with MAR missing data. The asymptotic variance of the ML/MI estimator of residuals variance is explicitly derived, from which it follows that the kurtosis of the error distribution primarily affects the asymptotic variance. Results of numerical simulations conducted to study finite sample properties of the estimators, conformed largely to the asymptotic results, and they also indicated interesting findings particularly for small samples, which do not follow from the asymptotic property. It is concluded that the ML estimators perform best in the situation studied here.

KW - Distribution misspecification

KW - Incomplete data

KW - Maximum likelihood

KW - Multiple imputation

UR - http://www.scopus.com/inward/record.url?scp=84939447954&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=84939447954&partnerID=8YFLogxK

U2 - 10.1080/03610926.2013.819920

DO - 10.1080/03610926.2013.819920

M3 - Article

VL - 44

SP - 3234

EP - 3250

JO - Communications in Statistics - Theory and Methods

JF - Communications in Statistics - Theory and Methods

SN - 0361-0926

IS - 15

ER -