Adding Weights

[ayushpatnaikgit]

How can we incorporate weights in the models? I want to do bootstrap to calculate the standard errors. The weights will be drawn from the multinomial distribution.

[sourish-cmi]

Can you please provide more detail about the models you are talking about? Currently, your question bit vague.

[ayushpatnaikgit]

For any model.
Suppose I have a data frame with 3 columns, X, Y and weights.

df = DataFrame(X = 1:10, Y = 20:29, weights = 1:10)

In GLM.jl, I can do:

glm(@formula(Y ~ X), df, Normal(), IdentityLink(), wts=df.weights)

However, in CRRao, there is no option for wts. We'll need to add keyword arguments in CRRao that will be passed to GLM. Not sure how weights are considered for bayesian regression.

With bootstrap, what I want to do is:

using Distributions
replicates = 1000
dist_coef = [coef(glm(@formula(Y ~ X), df, Normal(), IdentityLink(), wts=df.weights .* rand(Multinomial(10, 10))))[1] for i in 1:replicates]
SE = std(dist_coef)

[sourish-cmi]

Yes, this feature can be added. GLM can be added quickly. Not sure how to handle the Bayesian model at this stage.

[ayushpatnaikgit]

@ShouvikGhosh2048 can you allow fit to accept keyword arguments and then pass them to glm?

[ShouvikGhosh2048]

@ayushpatnaikgit
Should we allow arbitrary GLM keyword arguments?

Like this:

function fit(formula::FormulaTerm, data::DataFrame, modelClass::LinearRegression; kwargs...)
    model = lm(formula, data; kwargs...)
    return FrequentistRegression(:LinearRegression, model, formula)
end

Or just weights?

[ayushpatnaikgit]

I think all kwargs like you have shown.