Imputations over prices or other explanatory variables

[fedemolina]

In practice with real data (for example, retail) there are many missing values so, could an imputation model be useful for practitioners? (For univariate time series, imputeTS does an extremely nice job, with multivariate data there are many options). If so there are two different moments when this could be made. 1) Before the data is transformed to have the structure necessary for aids estimation. 2) After the data is transformed.

For example (in retail) we have:

geo: geographic location channel: number of supermarket checkout store: id store sku: id product date: date (day, week, ...)

And we can estimate a model for every geo-channel (or we could incorporate them as explanatory variables).

So in the first situation the imputation would be done over any store-sku-date. In the second case it would be done over the sku-date.

I could add a reproducible example to show the differences between (1) and (2)

[arne-henningsen]

I suggest to use the R package "Amelia". You can run it before and after the transformation and check which of the two approaches gives more reasonable imputations; this may depend on the data set.