Artificial Neural Networks for Short-Form Development of Psychometric Tests: A Study on Synthetic Populations Using...

Area of Science:

• Psychometric Research
• Machine Learning
• Artificial Intelligence

Background:

• Traditional short-form development methods in psychometrics rely on exploratory models with assumptions often unmet by psychological data.
• Methodological choices in test shortening are critical for researchers.
• Existing computational approaches may not fully capture data characteristics.

Purpose of the Study:

• To propose a machine learning-based autonomous procedure for short-form development.
• To investigate the item-selection performance of autoencoders for test shortening.
• To compare the proposed method with existing computational approaches.

Main Methods:

• Utilized autoencoders, a type of artificial neural network, for item selection in short-form development.
• Tested the procedure on artificial data simulated from a factor-based population.
• Compared the autoencoder-based method with traditional computational approaches.

Main Results:

• Autoencoders require minimal assumptions about data characteristics.
• The proposed procedure automatically selects items that effectively reconstruct original responses.
• The developed short forms preserve the internal structure of the long-form test.
• Autoencoders provide a method to predict long-form item responses from the short form.

Conclusions:

• Machine learning, specifically autoencoders, offers an effective and integrative approach to autonomous short-form development.
• This method addresses limitations of traditional techniques by making fewer assumptions on data.
• The autoencoder-based procedure aids researchers in creating robust short forms that maintain psychometric integrity.