Latent Class Analysis with lclogit: extreme dataset sensitivity regarding class outcomes

[Sebastian Heinen]

Hi,

I have a discrete choice experiment dataset with 917 respondents. When using lclogit to conduct a latent class analysis (LCA), the results differ widely if I exclude just a few respondents. To double-check, I simply deleted one single observation (the first one). And with n = 916, I get completely different class sizes and coefficients.

When I run (conditional) logit regressions, I receive almost identical results when I exlcude one or a couple of respondents.

What can I do? Obviously, LCA can't be that sensitive to excluding individual respondents, right?

Here are my data structure and commands:

id	choice	cc	cc2	options	Ch 1	...
1	1	1	1	1	1
1	0	1	1	2	3
1	0	2	2	1	2
1	1	2	2	2	3
2	1	1	3	1	1
2	0	1	3	2	2
2	1	2	4	1	2

lclogit choice ch1_low ch1_high ch2_low ch2_high ch3_low ch3_high ch4_none ch4_alot ch5_none ch5_alot ch6_none ch6_pub ch6_priv, id(id) group(cc2) nclasses(4)

To explain: ch1_low to ch6_priv are the effects-coded dependent variables. cc are the choice cards and cc2 are the accumulating choice cards.

I already tried to increase the iterations, this didn't lead to more similar results.

The results for n=917 and n=916

Best,
Sebastian

[Sebastian Heinen]

It appears that the problem is that the Stata algorithm identifies various local maxima: https://www.john-uebersax.com/stat/local.htm#fcall6