Imputation approaches for records with completely missing dimensions I have two sets of data provided by the government - one spans the years 2016-2020, while the other only covers 2018-2020. Data from each dataset, for each year, are used to predict some outcome in the next year. There are common joining criteria between these two datasets.
I have reason to believe the 2018-2020 dataset should be very useful in my predictive task. However, these data are completely missing (i.e. I cannot find them anywhere on the Internet and it's likely the government never collected this information for 2016 and 2017).
Without simply dropping the years 2016 and 2017 from my analysis, how could I best incorporate these 2018-2020 data? I saw this question but it didn't speak to an approach I thought of, which was to simply include the 2018-2020 as I normally would, and add a "this-type-of-data-was-completely-missing" flag in 2016 and 2017, in order to perhaps capture that any potential variation in the outcome that way. What potential pitfalls could there be with this approach?
 A: That use of a "missing data" flag is not a very good choice.
Multiple imputation, as recommended on the page you link, can seem daunting at first. It's not that difficult to understand (at least in outline) or to implement, however. You do multiple probabilistic imputations of the missing data, and fit your model separately on each imputed data set. You then combine information from those models to take into account both the inherent modeling variability and the variability introduced by the imputation.
Stef van Buuren, a respected expert on imputation, has an online book that starts with an explanation of how the type of coding you propose can lead to problems even in simple models. I suspect that in your situation, where you use prior years' data to predict subsequent years' values in some way, would be even worse.
So use multiple imputation for the 2016 and 2017 data if those data are important for your modeling. Most serious statistical software supports that, either directly or as an add-on. In R, you can use van Buuren's package mice.
