wbelzak/regdif

regDIF: Regularized Differential Item Functioning

I am currently a Senior Psychometrician at Duolingo.

I have a PhD and MA in Quantitative Psychology from UNC - Chapel Hill (2021, 2019; Advisor: Dan Bauer), an MA in Psychology from William & Mary (2017; Advisor: Todd Thrash), and a BA in Economics from N.C. State (2013).

I am broadly interested in psychometrics and human assessment. While most of my research has focused on measurement bias and personality/motivation, I have more recently been working on test security, test and rater reliability, and individual differences in decision-making.

My CV can be found here; my research papers can be found on Google Scholar; and my professional history can be found on LinkedIn. Feel free to reach out anytime at [email protected].

regdif's Issues

Negative variance estimate?

Hello,

Thank you for providing the package for DIF assessment. I got some questions when using the package and wonder if you could give me some directions:

1. Why the variance estimates of impact can be negative? For example, the var.gender for ida dataset shown at https://cran.r-project.org/web/packages/regDIF/vignettes/regdif_vignette.html.
   I suppose the variance should be constrained positive throughout the penalized EM optimization.

2. Could you give an example of the R code for providing start.values? I tried to given starting values for the var.gender but not worked.

Regards,
CW

Comparison of mirt and regDIF

Hello,

I tried to simulate two-group data without any DIF items (for testing purposes). Male group from N(0,1) and female group from N(1,1). 2PLM was used as the generating model. The true slope parameters are all one and the intercept parameters are all zero. Attached is the simulated data and the R syntax (change test.txt to test.R; text.docx is the data).
test.docx
test.txt

The results show that the mirt can well obtain the estimates, but the regDIF yielded
mean.cov1 = 0.5251, var.cov1 = -0.0602, X1.int. = 0.4376, and other intercepts were around 0.5.

1. It seems the regDIF cannot identify the focal-group mean and item intercepts.
2. What does it mean by var.cov1 = -0.0602?

Regards,
CW