SMABS 2004 Jena University
SMABS 2004 Home Organization About Jena Sponsors Links Imprint / Contact SMABS Home

European Association of Methodology

Department of methodology and evaluation research

Jena University

Contributions: Abstract

How does censoring of continuous parallel measurements bias validity of their common variable and scale reliability? A simulation study

Stephane VautierDelphine MoncanyHenri Sztulman
Université de Toulouse Le Mirail

Empirical studies often face data with strong floor or ceiling effects. When data are collected with analogue rating scales, the concept of censored variables seems useful to model these distribution phenomena, invoking thresholds. We assessed to what extent the censoring process may bias (a) the validity of the latent variable built using linear regression of the censored variables, and (b) the scale reliability. The degree of censoring was the independent variable. Moreover, PRELIS 2 allows one to estimate covariance matrices corrected for censoring. We also evaluated to what extent the use of corrected input data could reduce the censoring effect.

Multinormal zero-mean data were generated on 5000 cases for 5 parallel variables of unit variance, where the error variance was set at 1/3. Then, 5 below-censoring rules where applied to the data, using threshold values of -1, -0.5, 0, 0.5, and 1. Each of the 5 resulting data sets provided two covariance matrices (raw, corrected), further used as input data for ML estimation. Validity of the modeled true-score variable (h1) was defined as its correlation with the original common variable (Y6) used to generate the data. Scale reliability was estimated according to Raykov's (2001) use of dummy h variables in LISREL. Raykov's program was adapted to a model for parallel measurements. The correlation between h1 and Y6 was estimated based on its regression on h1, through the use of two other dummy variables.

The magnitude of the censoring effect was positively associated with the degree of censoring. Biases became marked before 0, i.e., the theoretical mean of the measurement variables, ranging from .73 to 1.00 for validity. Reliability measures ranged from .79 to .90 (theoretical reliability = .91). With corrected input data, censoring effects were clearly attenuated, but not neutralized.