Loss of power in logistic, ordinal logistic, and probit regression when an outcome variable is coarsely categorized

Aaron B. Taylor, Stephen West, Leona S. Aiken

Research output: Contribution to journalArticle

38 Citations (Scopus)

Abstract

Variables that have been coarsely categorized into a small number of ordered categories are often modeled as outcome variables in psychological research. The authors employ a Monte Carlo study to investigate the effects of this coarse categorization of dependent variables on power to detect true effects using three classes of regression models: ordinary least squares (OLS) regression, ordinal logistic regression, and ordinal probit regression. Both the loss of power and the increase in required sample size to regain the lost power are estimated. The loss of power and required sample size increase were substantial under conditions in which the coarsely categorized variable is highly skewed, has few categories (e.g., 2, 3), or both. Ordinal logistic and ordinal probit regression protect marginally better against power loss than does OLS regression.

Original languageEnglish (US)
Pages (from-to)228-239
Number of pages12
JournalEducational and Psychological Measurement
Volume66
Issue number2
DOIs
StatePublished - Apr 2006

Fingerprint

Probit Regression
Ordinal Regression
Logistic Regression
Least-Squares Analysis
Sample Size
Logistics
Logistic Models
logistics
regression
Least Squares Regression
Ordinary Least Squares
Regain
Psychology
Ordered Categories
Monte Carlo Study
Categorization
Research
Regression Model
Dependent

Keywords

  • Logistic regression
  • OLS regression
  • Probit regression
  • Statistical power
  • Variable categorization

ASJC Scopus subject areas

  • Mathematics (miscellaneous)
  • Psychology(all)
  • Developmental and Educational Psychology
  • Psychology (miscellaneous)

Cite this

Loss of power in logistic, ordinal logistic, and probit regression when an outcome variable is coarsely categorized. / Taylor, Aaron B.; West, Stephen; Aiken, Leona S.

In: Educational and Psychological Measurement, Vol. 66, No. 2, 04.2006, p. 228-239.

Research output: Contribution to journalArticle

@article{b86ac81496d6406596e2f0135d409737,
title = "Loss of power in logistic, ordinal logistic, and probit regression when an outcome variable is coarsely categorized",
abstract = "Variables that have been coarsely categorized into a small number of ordered categories are often modeled as outcome variables in psychological research. The authors employ a Monte Carlo study to investigate the effects of this coarse categorization of dependent variables on power to detect true effects using three classes of regression models: ordinary least squares (OLS) regression, ordinal logistic regression, and ordinal probit regression. Both the loss of power and the increase in required sample size to regain the lost power are estimated. The loss of power and required sample size increase were substantial under conditions in which the coarsely categorized variable is highly skewed, has few categories (e.g., 2, 3), or both. Ordinal logistic and ordinal probit regression protect marginally better against power loss than does OLS regression.",
keywords = "Logistic regression, OLS regression, Probit regression, Statistical power, Variable categorization",
author = "Taylor, {Aaron B.} and Stephen West and Aiken, {Leona S.}",
year = "2006",
month = "4",
doi = "10.1177/0013164405278580",
language = "English (US)",
volume = "66",
pages = "228--239",
journal = "Educational and Psychological Measurement",
issn = "0013-1644",
publisher = "SAGE Publications Inc.",
number = "2",

}

TY - JOUR

T1 - Loss of power in logistic, ordinal logistic, and probit regression when an outcome variable is coarsely categorized

AU - Taylor, Aaron B.

AU - West, Stephen

AU - Aiken, Leona S.

PY - 2006/4

Y1 - 2006/4

N2 - Variables that have been coarsely categorized into a small number of ordered categories are often modeled as outcome variables in psychological research. The authors employ a Monte Carlo study to investigate the effects of this coarse categorization of dependent variables on power to detect true effects using three classes of regression models: ordinary least squares (OLS) regression, ordinal logistic regression, and ordinal probit regression. Both the loss of power and the increase in required sample size to regain the lost power are estimated. The loss of power and required sample size increase were substantial under conditions in which the coarsely categorized variable is highly skewed, has few categories (e.g., 2, 3), or both. Ordinal logistic and ordinal probit regression protect marginally better against power loss than does OLS regression.

AB - Variables that have been coarsely categorized into a small number of ordered categories are often modeled as outcome variables in psychological research. The authors employ a Monte Carlo study to investigate the effects of this coarse categorization of dependent variables on power to detect true effects using three classes of regression models: ordinary least squares (OLS) regression, ordinal logistic regression, and ordinal probit regression. Both the loss of power and the increase in required sample size to regain the lost power are estimated. The loss of power and required sample size increase were substantial under conditions in which the coarsely categorized variable is highly skewed, has few categories (e.g., 2, 3), or both. Ordinal logistic and ordinal probit regression protect marginally better against power loss than does OLS regression.

KW - Logistic regression

KW - OLS regression

KW - Probit regression

KW - Statistical power

KW - Variable categorization

UR - http://www.scopus.com/inward/record.url?scp=33644787114&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=33644787114&partnerID=8YFLogxK

U2 - 10.1177/0013164405278580

DO - 10.1177/0013164405278580

M3 - Article

AN - SCOPUS:33644787114

VL - 66

SP - 228

EP - 239

JO - Educational and Psychological Measurement

JF - Educational and Psychological Measurement

SN - 0013-1644

IS - 2

ER -