Hi, I’m new to bayesian statistics and pymc. Say I have N variable correlated according to a MVN distribution. I have train dataset that I can use to find the posterior distribution of the MVN. Then I have new data which has missing observation on some subset of the N variables. I want to predict t…

If you set the masked values to nan and pass that as observed, pymc will automatically infer those during the course of sample. You can find them in trace.posterior. For test_data you will need to write the conditional model yourself. This blogpost may give some ideas, although it doesn’t have your…

Thanks. That solved my problem. The way you do it is by manually defining the transformation. It’s great since MVN has a nice closed form transformation for it. Now I’m curious, what if I my incomplete observation are from another distribution parameterized by the MVN? something like this: \text{…

You would have to rewrite the graph downstream of the conditional mvnormal but it should work. Simply gonna have more sources of randomness between the latent mvnormal and the final observations.

[image] messier12: I have train dataset that I can use to find the posterior distribution of the MVN. Then I have new data which has missing observation on some subset of the N variables. I want to predict the values for those missing data. What’s the way to do this in MCMC way? I am going t…

[image] bob-carpenter: just another latent or unknown value that you want to do posterior inference for conditioned on some observed data. That’s what setting the data as nan and calling sample would achieve. But for pipelining predictions of new data, using forward sampling will be much fas…

[image] ricardoV94: I didn’t know about the cut operator. How is that implemented? I don’t know how it’s implemented, but the effect is exactly what you describe. This comes up all the time in pharmacometric/pharmacodynamic (PK/PD) models. There you have a very well-informed PK model that g…

How to find conditional distribution of missing values?

Questions version agnostic

jessegrabowski November 18, 2025, 1:34pm 3

There a discussion about how to do this here. The whole thread is maybe interesting, but I linked the post with the best code.

Topic		Replies	Views
Calculating conditional posterior predictive samples in high-dimensional observation spaces v5 modeling	13	979	July 7, 2023
Handling missing values in predictor when outcome is a Multivariate Normal distribution v5	7	205	October 25, 2024
Sampling plausible posterior x parameter values in linear model for given y version agnostic	6	137	July 30, 2024
Automatic imputation and posterior predictive sampling with new data Questions	5	2103	October 29, 2020
OOS predictions with missing input values Questions	6	4182	November 22, 2018

How to find conditional distribution of missing values?

Related topics