Dealing With Missing Data

[alphamaximus]

I’m working on a dataset that has a lot of missing data. Rather than dropping those rows with missing data, I’d like to estimate the missing data using a Bayesian approach.

I found two examples on how to do this:

1. @fonnesbeck’s Intro Stat Modeling 2017 - Dealing with Missing Data
2. Ruslan Salakhutdinov and Andriy Mnih’s Probabilistic Matrix Factorization for Making Personalized Recommendations

In the first approach, a predictive regression model is used for each column. In the second approach, a single factor model is used for everything.

It seems that the second approach is more efficient and straightforward, but perhaps the second approach is also more computationally intensive.

I was wondering what other people thought about the pros and cons of each approach and when would one would use one vs the other.

Thanks!

[junpenglao]

related discussion:

Large Scale Factor Analysis with minibatch ADVI

Hi, I’m new to PyMC3 and am using it in my master dissertation. I have a dataset of roughly 1.3 million datapoints, each of 20 dimensions, and 40% of the data matrix are missing. My goal is to reduce its dimensionality for further analysis. The way I deal with missing values is simple - assign a 20 dimensional Bernoulli mask, that either returns the “true” value (when entry is observed) or a placeholder (when the entry is missing). To make it run on ADVI, I have integrated away the Bernoulli la…