`fill` with `value=f::Function` isn't comfortable with entirely `missing` data

[nickrobinson251]

Here are some common functions we may want to use with fill but which (arguably) have awkward behaviour if all data is missing -- stemming from fill using drop and subsequently ending up calling the function on a empty array:

julia> using Impute, Statistics

julia> x = [missing 2.0; missing 4.0]
2×2 Array{Union{Missing, Float64},2}:
 missing  2.0
 missing  4.0

julia> Impute.fill(x, value=mean, dims=1)
2×2 Array{Union{Missing, Float64},2}:
 NaN  2.0
 NaN  4.0

julia> x = [missing 2; missing 4];

julia> Impute.fill(x, value=mean, dims=1)
ERROR: InexactError: Int64(NaN)

julia> x = [missing 2.0; missing 4.0];

julia> Impute.fill(x, value=median, dims=1)
ERROR: ArgumentError: median of an empty array is undefined, Union{Missing, Float64}[]

julia> x = [missing 2.0; missing 4.0];

julia> Impute.fill(x, value=middle, dims=1)
ERROR: ArgumentError: collection must be non-empty

(Impute v0.3, tested on Julia v1.2)

Apologies for not having a solution here... but some question we might want to think about

• Should these all be errors?
• If so, should they error in a "nicer" way?
• If not, should they simply fail to impute, i.e. leave the missings, a la locf if the first element is missing?

[rofinn]

I'd be okay with retaining the missing in this case. One thing I've been considering is that if we split out an Iterators interface then we'd likely want to use an OnlineStat instead. Unfortunately, this would just use that default value (e.g., value(Mean()) -> 0.0) which also might not be what we want? I guess in that case we could check nobs(value) before returning.

[rofinn]

Alright, #69 fixed this by having impute(data::AbstractArray{Missing}, imp) just return in the input data cause most imputation methods are going to need at least some non-missing values to work properly.