We may have data from different data sources. Some samples can get data from every possible data source. But others can only obtain information from one source. Each source may contain hundreds of variables.
The following figure shows an example. There are 10,000 rows in the data. Part 1, 5000 rows, V1-V100 are all missing. Part 2 data has both V1-V100 & V101-V200.
I'm considering model stacking. First, I can use part1B+part2B to train a model. This model could use more data points and should produce estimates with smaller variance (since a larger sample is used here). The second model can use part2A+part2B data. 2nd model contains more variables, and hopefully we can get a model with 'smaller' bias.
How should I split data into training, validation and testing in this case?