KCI independence test for mixed data

[samvanni]

We want to run FCI with the KCI independence test and have a couple of binary variables (vegetarianism, gender) and a couple of scaled discreet variables (like smoking, where 0 is not smoking, 1 smoking rarely, 4 smoking frequently) and a couple of normal numerical variables.
We found several old discussions on mixed data, the general advice was to check out Tetrad and that causal-learn does not really support mixed data yet. Have there been any updates on that? What would you suggest now to use on the data we described?

[kunwuz]

Thanks for your great question. In principle, nonparametric tests such as KCI and its faster versions (FastKCI and RCIT) can deal with mixed data, since they put no constraints on the distribution. However, I'm not sure whether that is optimal in practice. I would still suggest trying both kernel-based tests in causal-learn and conditional/degenerate Gaussian scores in Tetrad/Pytetrad, but I do not have a definite answer on what's better.