I recently received a comment from a reviewer over the limitation statement we had included in the manuscript “In this cross-sectional study, we could not determine the temporal relationship between low HDL-c and the covariates”, that would be nice for as to conduct deep learning and Mendelian randomization methods to ascertain potential causal effect. I was actually hearing these methods for the first. How can one apply them in a cross-sectional study to determine potential causal effect?
Mendelian randomization is essentially instrumental variable analysis with genes/genetic variants. There are several introductory articles written on the topic, but perhaps this one is a nice starting point:
You use the fact that genetic variants related to your exposure of interest are for the most part randomly distributed, unaffected by issues such as confounding. You then use the variants to estimate a presumed causal effect of your exposure on the outcome. In reality and depending on your study population there are still several fine points to think about, but this is the gist. Perhaps these analyses are possible in your study, but you will probably have to perform completely different analyses than you have done so far.
I’m not familiar with performing deep learning in this context, but essentially it’s just a different approach to statistical modeling. Different modeling doesn’t solve the cross-sectional nature of the measurements in your study, so the best that come out of this is just another estimate of your exposure-outcome relation obtained through some slightly different model but all the caveats you mention still stand.
1 Like
Very much agree. Deep Learning has no special property that somehow enables causal conclusions. It’s a very interesting/successful method for e.g. prediction/classification (and other correlational tasks) when your input data is images, audio, text, time series and also for tabular data. Of course, people are writing plenty of papers that are trying to add causal inference on top of it, but it’s completely unclear how that would help with unclarity on temporal ordering of things. I don’t want to be too negative of a reviewer I don’t know, but if that’s the whole comment, then it probably just means “Have you considered a more innovative approach? I don’t really know much about this type of problem, but here’s some buzzwords I have heard recently that may or may not be relevant: …”
3 Likes
Thank you Scboone and Bjoern, your contribution helped me address the comment from the reviewer and the manuscript has been formally accepted for publication.