Guanlan Xu

Guanlan Xu

Associate Research Scientist at Pearson

Talks and presentations

Integrating Bifactor Models into G-Theory Frameworks

April 23, 2022

Paper, National Council on Measurement in Education 2022, SAN DIEGO, CA

We integrated bifactor models into multi-occasion G-theory SEM frameworks to allow for expanded score consistency indices; separation of systematic variance into general and group factor effects and measurement error into transient, specific-factor, and random-response effects; and generalization of results to broader domains from which items and occasions were sampled.

Accuracy of Absolute Error Estimates within a G-theory SEM Framework

April 09, 2022

Paper, National Council on Measurement in Education 2022, SAN DIEGO, CA

We evaluated the accuracy of Jorgensen (2021)’s method for estimating variance components for absolute error using structural equation models. Results from numerous self-concept and personality measures revealed that Jorgensen’s procedure yielded variance components, generalizability coefficients, and dependability coefficients equivalent to those produced by variance-component programs within R, SAS, and SPSS.

Matching G-Theory Analyses to the Numbers of Response Options Available

August 10, 2019

Presentation, American Psychological Association 2019, Chicago

We used scores from a multi-dimensional self-concept measure to compare split-half versus item level scores in task-by-occasion, G-theory analyses that took multiple sources of measurement error into account. Split analyses were performed on raw-score metrics, and item analyses on both raw-score and continuous-latent-response-variable (CLRV) metrics. Reliability coefficients for raw-split and CLRV-item scores always exceeded those for raw-item scores, but differences diminished as number of scale points increased. Reliability coefficients for raw-split dichotomous scores were lower than those for CLRV dichotomous scores but comparable or higher as number of scale points increased. Overall, these results highlight advantages of using splits in G-theory analyses in practical settings in which raw scores are needed for decision making.

Using the Bifactor Model to Quantify Effects of Response Styles

August 10, 2019

Paper, American Psychological Association 2019, Chicago

The bifactor model was used to separate effects of target constructs of interest and response styles on results obtained from the 100-item version of the International Personality Item Pool Big Five Model Inventory (IPIP-BFM-100; Goldberg, 1999). Results showed that responses were affected by both the target constructs and approaches to answering negatively and positive phrased items. Overall, response styles accounted for 7% to 21% of the variance in subscale scores. Effects of response styles varied across subscales, with relative effects of positive and negatively keyed items largely dependent on the proportion of such items included in a given subscale. Overall, these results highlight that the bifactor model can provide additional insights into the nature of item scores and aid in the construction of measures to reduce and better balance item wording effects.

Single-Occasion and Longitudinal CFA Analysis of Responses to the Mini-IPIP-BFM

August 10, 2019

Poster, American Psychological Association 2019, Chicago

We evaluated psychometric evidence for two versions of the 20-item Mini-International Personality Item Pool Big-Five Model Inventory (mini-IPIP; Donnellan, Oswald, Baird, & Lucas, 2006) in relation to the 50-item version (IPIP-50) from which they were created (Goldberg, 1992). Results showed that the mini-IPIP provided good psychometric properties in relation to the original 50-item version. Psychometric properties of the mini-IPIP were improved further by replacing one item in the mini-IPIP Intellect/Imagination scale.

Using Data Augmentation to Improve IRT-3PL Calibrations with Small Samples

August 08, 2019

Poster, American Psychological Association 2019, Chicago

We investigated the effectiveness of Duplicate, Erase, and Replace Augmentation procedures (DupER; Foley, 2010) in calibrating the 3PL model using small datasets in an operational setting. DupER procedures were most effective with medium-sized samples (n = 600) in which imputed datasets adequately reflected the score distribution within the target population.