Hello, I would like to know if it is possible to use items and parcels as indicators in lca at the same time. I have six items which are loading on three factors (each factor has two indicators). For the third factor, the loading of the second indicator are low in comparism to the other factor loadings (.646). For the LCA I used first all items as indicators, but as the loading of the one item are low (.646), this item did not differentiate very well (e.g. in all 4 classes, for this item, persons had probabilities of more or less 55% to be in category 2). When I use for factor 3 one item and one parcel instead of the two items, the 4-class solution is much better differentiated. 1) Is it ok to use an item in combination with a parcel as indicators for one factor and only items for the other factors, or does this bias the results of the class-solution?
2) Another question is, if there is any literature about the optimal size of the classes. I have one class which includes about 7.9% of the sample and I am wondering if this is too small.
1) you don't want to use both an item and a parcel containing the item because then you have a direct correlation between the two variables which is not modeled in the LCA.
2) The minimum percentage of the classes is related to the sample size - a large sample allows small percentages. There is no fixed rule and only a Monte Carlo study could tell what could be well recovered.
Thank you very much for the prompt answer. I think i was unclear, sorry. My parcel does not contain the same item as I use already in the analysis - it contains two other items, which are indicators for the same factor...(factor is measured by: item1; parcel=item2 and item3) so under that circumstances, may I use the item and the parcel as indicators? Thank you!!
I have selected the LCA model (4 classes) with the best fit, and have a class that is 4% of the sample (n=555). This model is substantively interpretable and actually quite illuminating.
However, a reviewer has said the norm is that a class should be more than 5%. I have never heard this minimum class size rule and cannot find it in articles by you or in Collins and Lanza’s book.
In the exchange above, you say that a large sample allows small percentages, and that there is no fixed rule regarding class size. Could you please suggest a source (beyond this discussion board) that I can cite to the reviewer?
I don't think a fixed rule like that makes much sense. It is true that few people in a class may not provide reliable estimates of class-specific parameters, and 4% of 555 isn't a large number of people, but if it is a clearly interpretable class I think it would be a pity to not consider it - and in a larger sample down the road see if it re-appears.
I don't know about writings on this - I think it is hard to say anything general about it.
Could you provide any references for why it would still be helpful to interpret classes that have smaller sample sizes? (e.g., n = 18-20) It would be very useful for manuscript submission as Tatiana brought up.
I would respond like I did on 10/9/14. I'm afraid I don't know of a reference that would say the same thing. You may want to ask on SEMNET or email LCA authors of papers that we have posted on our website under Papers, LCA.