Leif Ekblad posted on Tuesday, April 07, 2015 - 3:02 pm
I'm using CFA with 121 variables (discrete values 1-3) and have a 12-factor model. I'm using the WLSMV estimator. My problem is that CFI and TLI decreases with sample size. I extracted different samples sizes from the same dataset (600, 1500, 3000, 6000 and 18000), and then I used DataFit to find the best equation that could estimate CFI and TLI.
This resulted in the following equations:
CFI = 0.981+3.763E-06*N-1.0264E-03 *sqrt(N)
TLI = 0.982+3.982E-06*N-1.0714E-03 *sqrt(N)
From previous research, I know the data is not normally distributed, but rather best fits with two discrete overlapping normal distributions. Could this result in CFI and TLI depending on sample size?
Another strange thing is that the above equations seems to work for many different factor models using different datasets.
Is there a way to eliminate this dependence on the sample size, or might it be acceptable to present these equations in a paper and then to normalize the CFI/TLI values so they can be compared regardless of sample size?
Leif Ekblad posted on Wednesday, April 08, 2015 - 12:34 am
DataFit is a software package that attempts to fit a set of values to a mathematical function.
The main problem I have is that if I use a small sample (like 600), then both CFI and TLI are excellent (0.96), while if I use a large sample (like 18000), then CFI is 0.91, and the match appears to be considerably worse. Before I discovered this the results appeared random as I used different sample sizes to try to find the best solution.