Distinguishing “missing at random” and “missing completely at random”

Daniel F. Heitjan, Srabashi Basu

Research output: Contribution to journalArticle

129 Citations (Scopus)

Abstract

Missing at random (MAR) and missing completely at random (MCAR) are ignorability conditions—when they hold, they guarantee that certain kinds of inferences may be made without recourse to complicated missing-data modeling. In this article we review the definitions of MAR, MCAR, and their recent generalizations. We apply the definitions in three common incomplete-data examples, demonstrating by simulation the consequences of departures from ignorability. We argue that practitioners who face potentially non-ignorable incomplete data must consider both the mode of inference and the nature of the conditioning when deciding which ignorability condition to invoke.

Original languageEnglish (US)
Pages (from-to)207-213
Number of pages7
JournalAmerican Statistician
Volume50
Issue number3
DOIs
StatePublished - Jan 1 1996

Fingerprint

Ignorability
Missing Completely at Random
Missing at Random
Incomplete Data
Data Modeling
Missing Data
Conditioning
Simulation

Keywords

  • Bayesian inference
  • Coarse data
  • Frequentist inference
  • Ignorability
  • Incomplete data
  • Likelihood inference
  • Missing data

ASJC Scopus subject areas

  • Statistics and Probability
  • Mathematics(all)
  • Statistics, Probability and Uncertainty

Cite this

Distinguishing “missing at random” and “missing completely at random”. / Heitjan, Daniel F.; Basu, Srabashi.

In: American Statistician, Vol. 50, No. 3, 01.01.1996, p. 207-213.

Research output: Contribution to journalArticle

@article{9d5332424ecb4f219341eb553464d792,
title = "Distinguishing “missing at random” and “missing completely at random”",
abstract = "Missing at random (MAR) and missing completely at random (MCAR) are ignorability conditions—when they hold, they guarantee that certain kinds of inferences may be made without recourse to complicated missing-data modeling. In this article we review the definitions of MAR, MCAR, and their recent generalizations. We apply the definitions in three common incomplete-data examples, demonstrating by simulation the consequences of departures from ignorability. We argue that practitioners who face potentially non-ignorable incomplete data must consider both the mode of inference and the nature of the conditioning when deciding which ignorability condition to invoke.",
keywords = "Bayesian inference, Coarse data, Frequentist inference, Ignorability, Incomplete data, Likelihood inference, Missing data",
author = "Heitjan, {Daniel F.} and Srabashi Basu",
year = "1996",
month = "1",
day = "1",
doi = "10.1080/00031305.1996.10474381",
language = "English (US)",
volume = "50",
pages = "207--213",
journal = "American Statistician",
issn = "0003-1305",
publisher = "American Statistical Association",
number = "3",

}

TY - JOUR

T1 - Distinguishing “missing at random” and “missing completely at random”

AU - Heitjan, Daniel F.

AU - Basu, Srabashi

PY - 1996/1/1

Y1 - 1996/1/1

N2 - Missing at random (MAR) and missing completely at random (MCAR) are ignorability conditions—when they hold, they guarantee that certain kinds of inferences may be made without recourse to complicated missing-data modeling. In this article we review the definitions of MAR, MCAR, and their recent generalizations. We apply the definitions in three common incomplete-data examples, demonstrating by simulation the consequences of departures from ignorability. We argue that practitioners who face potentially non-ignorable incomplete data must consider both the mode of inference and the nature of the conditioning when deciding which ignorability condition to invoke.

AB - Missing at random (MAR) and missing completely at random (MCAR) are ignorability conditions—when they hold, they guarantee that certain kinds of inferences may be made without recourse to complicated missing-data modeling. In this article we review the definitions of MAR, MCAR, and their recent generalizations. We apply the definitions in three common incomplete-data examples, demonstrating by simulation the consequences of departures from ignorability. We argue that practitioners who face potentially non-ignorable incomplete data must consider both the mode of inference and the nature of the conditioning when deciding which ignorability condition to invoke.

KW - Bayesian inference

KW - Coarse data

KW - Frequentist inference

KW - Ignorability

KW - Incomplete data

KW - Likelihood inference

KW - Missing data

UR - http://www.scopus.com/inward/record.url?scp=0030343462&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=0030343462&partnerID=8YFLogxK

U2 - 10.1080/00031305.1996.10474381

DO - 10.1080/00031305.1996.10474381

M3 - Article

AN - SCOPUS:0030343462

VL - 50

SP - 207

EP - 213

JO - American Statistician

JF - American Statistician

SN - 0003-1305

IS - 3

ER -