Comparison of weighted least squares and robust estimation in structural equation modeling of ordinal categorical data with larger sample sizes

Abstract

The effect of different sample sizes on estimation methods such as weighted least squares and robust weighted least squares that are used in structural equation modeling was studied and compared using information criteria such as Akaike Information Criteria in this study. The simulations were repeated 1000 times with two estimation methods and the average values of criteria were calculated with different sample sizes. The study includes a construct of four factors, with four questions of each that are measured on a five-point Likert scale. Different sample sizes, ranging from 300 to 5000 were selected. According to the simulations results, it is concluded that the robust estimation method provides more effective results at lower sample size. In addition, it was found that as the sample size increases, the efficiency difference between two methods gradually decreases. Moreover, it was detected that there is almost no difference between the two methods for sample sizes over 3000.

Is the sexual sensation seeking scale a reliable instrument? addressing multiple factor structures in a Colombian sample

“…On the first subsample, we computed an exploratory factor analysis (EFA) with Weighted Least Square (WLS) and Promax rotation. We chose WLS due to the limitations of estimating ordinal variables from estimators developed for quantitative data such as the Maximum Likelihood (ML) method (Gazeloglu & Greenacre, 2020). For the same reason, we relied on the polychoric correlation matrix.…”

Health literacy, worry about unmet needs for medical care, and psychological well‐being among older Chinese adults

“…The weighted least squares (WLS) estimator was adopted in the SEM because the key variable, worry about future unmet needs for medical care, is an ordinal variable with only four levels. 27 The mediatory effects of worry about future medical care were tested with the bootstrapping approach with 5000 replications, and 95% biascorrected confidence intervals were calculated. We adopted several model fit indices, including the comparative fit index (CFI) (>0.95), Tucker-Lewis index (TLI) (>0.95) the goodness-of-fit index (GFI) (>0.95), the root mean square error of approximation (RMSEA) (<0.06), and the standardized root mean square residual (SRMR) (<0.08) as indicators for satisfactory model fits.…”

“…All items for the key variables, including health literacy, psychological well‐being, and worry about unmet needs for medical care, have values of skewness and kurtosis smaller than 1. The weighted least squares (WLS) estimator was adopted in the SEM because the key variable, worry about future unmet needs for medical care, is an ordinal variable with only four levels 27 . The mediatory effects of worry about future medical care were tested with the bootstrapping approach with 5000 replications, and 95% bias‐corrected confidence intervals were calculated.…”

Effects of Category Labeling With Emojis on Likert-Type Scales on the Psychometric Properties of Measurements

“…WLS requires large samples to make accurate predictions (Flora & Curran, 2004). By contrast, DWLS and ULS are estimation methods that can be used in relatively small samples (Gazeloglu & As ¸an Greenacre, 2020;Li, 2016;Maydeu-Olivares, 2001). Forero et al (2009) and Kogar and Yılmaz Kogar (2015) determined that ULS gives slightly more precise and accurate estimates than DWLS.…”