juliapsychometrics/raschmodels.jl

Rasch modeling with all the bells and whistles. Implementations for Rasch model, partial credit model, rating scale model, and its linear extensions (upcoming). Classical and Bayesian estimation.

raschmodels.jl's Issues

Implement conditional maximum likelihood estimation for Polytomous Rasch Models

Add support for CML estimation for the following models

• Partial Credit Model
• Rating Scale Model

Implement Warm's weighted likelihood estimation for person parameters

The conditional maximum likelihood estimation only allows to estimate item parameters. Warm's weighted likelihood estimation (Warm, 1989) should be implemented to also allow for person parameter estimation.

Improve TTFX

Since we do not need to support julia LTS we can bump the required julia version to 1.8 and use SnoopPrecompile to precompile model calculations.

This should make the first run for users noticably faster.

Code formatting

We should probably define some sensible defaults for JuliaFormatter.
For now I just used the default settings in VSCode.
Realistically we should use the same style for every package in JuliaPsychometrics.

Do you have any preferences @t-alfers?

Remove duplicated code

Duplicated code should be simplified. This is relevant for:

• AbstractItemResponseModels.iif

RaschModels.jl/src/iif.jl

Lines 1 to 12 in 9c094e4

	function AbstractItemResponseModels.iif(model::RaschModel{PointEstimate}, theta::Real, i, y::Real)
	checkresponsetype(response_type(model), y)
	parname = Symbol("beta[$i]")
	beta = model.pars.values[parname]
	return _iif(theta, beta, y)
	end
	function AbstractItemResponseModels.iif(model::RaschModel{SamplingEstimate}, theta::Real, i, y::Real)
	checkresponsetype(response_type(model), y)
	beta = vec(model.pars["beta[$i]"])
	return _iif(theta, beta, y)
	end

• AbstractItemResponseModels.irf

RaschModels.jl/src/irf.jl

Lines 1 to 12 in 9c094e4

	function AbstractItemResponseModels.irf(model::RaschModel{PointEstimate}, theta::Real, i, y::Real)
	checkresponsetype(response_type(model), y)
	parname = Symbol("beta[$i]")
	beta = model.pars.values[parname]
	return _irf(theta, beta, y)
	end
	function AbstractItemResponseModels.irf(model::RaschModel{SamplingEstimate}, theta::Real, i, y::Real)
	checkresponsetype(response_type(model), y)
	beta = vec(model.pars["beta[$i]"])
	return _irf(theta, beta, y)
	end

These two instances can be fixed by implementing a new function getitempars for fetching item parameters for various data structures. Eventually getitempars could be part of the JuliaPsychometrics/AbstractItemResponseModels.jl interface.

Fetching item parameters for CML fits

getitempars fails for models fit by CML().

I think this is due to the new structure of the parameter results implemented in #37.

Add bayesian model fitting via variational inference

Variational inference is supported by Turing.vi() and can be easily implemented using a custom _fit function.

See https://turing.ml/dev/tutorials/09-variational-inference/ for details about VI in Turing.jl

This might require an additional EstimationType in JuliaPsychometrics/AbstractItemResponseModels.jl since VI does return a T <: Distribution not posterior samples or a single point estimate.

Duplicate _fit_by_alg

Why are both of these methods required? If we leave out the second function definition, then _fit_by_alg will dispatch on the correct method of _fit_by_cml anyways.

RaschModels.jl/src/fit.jl

Lines 161 to 176 in ccca8a0

	function _fit_by_alg(modeltype, data, alg::CML, args...; kwargs...)
	estimate = _fit_by_cml(modeltype, data, alg, args...; kwargs...)
	return estimate, PointEstimate
	end
	# for CML estimation a separate logic is necessary to follow if missing values in dataset
	function _fit_by_alg(
	modeltype,
	data::MatrixWithMissings{T},
	alg::CML,
	args...;
	kwargs...,
	) where {T}
	estimate = _fit_by_cml(modeltype, data, alg, args...; kwargs...)
	return estimate, PointEstimate
	end

Add documentation

We should at minimum add documentation for all user facing functions in the package.

TODO

• irf
• iif
• expected_score
• information
• fit

Implement linear model variants

This issue splits #5.

Separately implement linear variants of simple Rasch models:

TODO

• Linear Logistic Test Model
• Linear Rating Scale Model
• Linear Partial Credit Model

Item response function always returns 0

For RaschModel{SamplingEstimate} the irf function always returns 0.
Reason being that scoring_function defaults to identity here:

RaschModels.jl/src/models/RaschModel.jl

Lines 54 to 59 in 245cd73

	function irf(model::RaschModel{SamplingEstimate}, theta, i, y = 1)
	n_iter = length(getitempars(model, i))
	probs = zeros(Float64, n_iter)
	add_irf!(model, probs, theta, i, y)
	return probs
	end

scoring_function is needed in add_irf! for expected_score, but here we can just make it default to x -> one(x) to fix the issue

Improve performance of bayesian model estimation

Before the first release we should try to make models as performant as possible.

Update to AbstractItemResponseModels 0.2.0

AbstractItemResponseModels.jl 0.2.0 implements two new interface functions getitemlocations and getpersonlocations that should be implemented once JuliaPsychometrics/AbstractItemResponseModels.jl#17 is merged.

Implement conditional maximum likelihood estimation for Rasch Model

Add an additional estimation method for classical model estimation.

Add simple Bayesian Rasch Models

We still need to implement some additional models:

TODO

• Rating Scale Model
• Partial Credit Model

RaschModel struct should not be mutable

It is not required that RaschModel is a mutable struct.

RaschModels.jl/src/models/RaschModel.jl

Lines 1 to 5 in 6606a56

	mutable struct RaschModel{ET<:EstimationType,DT<:AbstractMatrix,PT} <: AbstractRaschModel
	data::DT
	pars::PT
	parnames_beta::Vector{Symbol}
	end

Scoring function for RaschModel

scoring_function is ignored for y = 0 in expected_score.

Implement scoring function

Implement an additional argument for expected_score and information for scoring functions