Latent Correlations and differences between versions of Conquest

[benjamin.euen]

Hi,

I am using Conquest for multidimensional item response models up to 10 dimensions (using achievement variables) with a conditional latent regression model (using background variables).

Could you provide me some reference papers with a description of the mathematical details used in ConQuest to estimate the correlations between latent variables (manual p.6 and p.110) to be estimated a) in the first run to estimate item parameters and b) in the second run to estimate plausible values using conditional latent regression model?

Furthermore I figured out that analysis with the same data and parameters turned to different results depending on the version of ACER ConQuest. Comparing results calculated with ACER ConQuest 3.0.1 with an older version the outputs show slight differences in the final deviance, the item difficulties and the latent correlations as well as bigger changes in the MNSQ’s. Was there an adaption of the technical details?

Thanks!
Benjamin

[Eveline Gebhardt]

Hi Benjamin

Have a look at these

Multilevel Item Response Models: An Approach to Errors in Variables Regression
Journal of Educational and Behavioral Statistics March 20, 1997 22: 47-76.

Raymond  j Adams Mark Wilson Wen-chung Wang (1997). The Multidimensional Random Coefficients Multinomial Logit Model. Applied Psychological MeasurementMarch, 21 no. 1 1-23.

Eveline

[hartmann]

Benjamin,

as far as I know, all ConQuest versions will provide identical output if you specify certain parameters in the input syntax (parameter change, no. of iterations, no. of nodes, node range …). It seems to me that the standard values for some of these have changed with newer program versions. (Please correct me if I'm wrong here.)

Cheers,
Stefan