Hi,

A browse-through of the formula for CFI may reveal some clues.

CFI is a goodness of fit index which take the ratio of current fit to null model fit and subtracts it from one:

CFI = 1 - ((fit-hypothesized)/(fit-null))

so if you are getting CFI = 0, then a little algebra shows that for your hypothesized model ((fit-hypothesized)/(fit-null)) must be 1.

In the CFI:

fit-hypothesized is (chi-square - df) for the hypothesized model

and fit-null is (chi-square - df) for the null model.

df = degrees of freedom for the chi-square

There are a couple of other tweaks to the formula to look for (see linked paper below). The value in the numerator of this quantity

((fit-hypothesized)/(fit-null))

is really the maximum of fit-hypothesized or 0. So if your fit-hypothesized is negative (chi-square - df), CFI uses 0 instead of a negative value.

A similar rule applies to the denominator such that it takes the maximum of fit-null, fit-hypothesized, or 0. This means that if fit-null < 0 and fit-hypothesized < 0, then CFI for you will be undefined, and may be set to 0 by the software you are using:

CFI = 1 - (0/0) = undefined

Hope this helps,

Scot

https://digitalcommons.wayne.edu/cgi/viewcontent.cgi?referer=https://www.google.com/&httpsredir=1&article=1810&context=jmasm

Very helpful! Thank you!

Hi Stephanie,

Your experience sounds frustrating! I am sorry you have encountered this issue. Sometimes it can be tricky to get LGCs to run correctly in AMOS. A few thoughts:

The null model in AMOS is an independence model and one where all correlations between variables are 0. Hence, the model would have to be very poor to produce a CFI of 0.000, which is unlikely.

I would first check that your model has been specified correctly (e.g. you have the proper loadings on the intercept, slope, and disturbances). Also verify you are not attempting to estimate too many parameters using the formula (v(v+1)/2) - p where v is the number of observed variables in your model and p is the number of unique parameters you are attempting to estimate. If you need a reference I would encourage you to check out "Latent curve models: A structural equation perspective" by Bollen and Curran. Its a great resource.

AMOS will often attempt to estimate a model, even if it is incorrectly specified, and the resulting output will not make sense. Hence, whether or not the model ran does not mean the results can be trusted. If you post your path diagram or email it to me at [email protected] I am happy to help double-check it with you.

If model specification does not seem to be the problem, check the variance-covariance matrix to ensure there are no negative values which would indicate an impermissible solution. Sometimes this can occur due to having a small N (