I am using Mplus for an EFA on 68 items. The higher the number of factors, the better the fit gets and, particularly, the Chi-square difference is always significant for each addition of one factor I tried (until now 12).
Parallel analysis run in spss suggested 6 factors and MAP-analysis suggested 8 factors.
So my question is: when should one stop adding factors in Mplus? Which are the best indices? I fear that if I only base my argumentation on the interpretability of the factor solution this may sound to a reviewer like chosing the result that fulfil my expections.
I'm new to Mplus and have got very much the same problem. My questionnaire contains 53 Items, n is 282. Parallel analysis and MAP-Test in SPSS indicate a 5-factor-model but the fit in Mplus for this model is bad.
Consequently I tried more factors. Though the fit gets better the more factors I add, the additional factors contain only 1-2 items or consist of double-loading-items only (though rotation is geomin).
In my case, is it useful to use RMSEA, CFI, TLI and SRMR to determine the number of factors or should I rely on MAP-test and parallel analysis?
I don't see you mention the substantive theory behind the questionnaire which should guide your item creation and the expected number of factors. Applying factor analysis to a set of items without that background is not likely to be successful - not all data can be well fitted by a factor model. Perhaps you are in the early stages of a questionnaire development in which case you should consider deleting and adding items to measure the hypothesized factors better. I think it is always useful to consider fit statistics like chi-square and RMSEA in addition to more descriptively-oriented checks like parallel analysis.
You may also ask this general (not Mplus-specific) modeling question on SEMNET.
in your Topic 1 video of short courses you recommend to determine the number of factors in an EFA by picking the solution with lowest amount of factors which has an unsignificant chi-square and other acceptable fit-measures. Do you know any references which i could citate to justify this procedure?
I don't know that this needs a reference given that it is like "a first principle" of modeling - finding the best fitting but still parsimonious model (so in line with the BIC criterion). I can't off-hand point to an EFA book on ML - I think Lawley-Maxwell's classic factor analysis text would cover it but also other modern factor analysis books.