Factor replication as a criterion for number of factors in exploratory factor analysis of items.

摘要

The number of factors to extract has been an issue in exploratory factor analysis of items. Several methods have been used regularly, but none has proven to work well in a wide variety of conditions in item analysis. A variation of a cross-validation technique, called factor replication, is described. It was applied to simulated item data and compared to parallel analysis, Wrigley's criterion, MAP, Bartlett's test of significance, the screen test, and the roots >1 criterion. The overall performance of factor replication was generally accurate (74%), as was parallel analysis (75%), although both methods had unique strengths and weaknesses. The strength of factor replication appears to be that it performs well across various research conditions with no glaring weaknesses. Some caution regarding factor replication likely needs to be used, however, with weak loadings, when N is small, and number of factors is large. Parallel analysis has some clear advantages over the other methods in several areas. It is the strongest method when factors have weak loadings, and many items. It is probably also the strongest method with low N. It should not be used when there is less than approximately 5 items per factor. The performance and accuracy (66%) of Wrigley's criterion was roughly equivalent to factor replication. The other methods performed inadequately.