haroine / icarus Goto Github PK

The idea is to change the recommended margin input method to something like:

marginList <- list(
"categ" = c(0.35, 0.4, 0.25),
"sexe" = c(0.60, 0.40),
"service" = c(0.45, 0.55),
"salaire" = 470000
)

• Add tests with the list method, with categorical and continuous variables

• Modify addMargin

• marginMatrix becomes an ambiguous parameter name: consider changing it (but ensure retrocompatibility!)

• Change recommended way of integrating margins in the docs

Thanks to Toby Hocking for the suggestion!

Following the linearization formulas of Dell et. al: http://www.crest.fr/ckfinder/userfiles/files/Pageperso/xdhault/doc_JMS.pdf, Icarus could support calibration on non-linear totals such as quantiles, Gini indices, etc.

Due to the approximations, we would have to come up with an iterative procedure that would check that the final estimates correspond to the true value.

Some functions of the gustave package https://github.com/martinchevalier/gustave could be reused

Create a function so that creating a margin table would not require the actual creation of a matrix to the right format.

Bonjour,
Une question est il possible dans Icarus de créer de nouvelles valeurs de pondération à partir de zéro ?
Je veux dire par là ne pas avoir de pondération a priori ?

English version below :
Hi,
Just a litle ask, : Is it possible to create ponderation weight from scratch in icarus ?

Change all French variable names in examples and datasets to English

Hi, I need to perform a calibration using margins on a dataset of observations.
Now this operation is performed using the macro CALMAR2 in SAS, I tried using the Icarus package but the algorithm doesn't converge.
I checked all the input parameters and they are the same that I use in SAS, with the 'linear' method the algorithm converges but with the raking no.

q_k is a already a parameter of the (private) calibAlgorithm function. It should be added in calibration and properly tested so it can be used in production.

Return a list instead and suggest in the docs to use operator %<-% to facilitate multiple assignment of returned calibrationMatrix and marginMatrix.

Be careful to not deprecate the side effects functionality too fast.

As of version 0.2.0, calibration only returns a column of weights, so colCalibratedWeights should not be used as a parameter anymore.

When I run the following, I get minimum and maximum weight values that are outside the bounds. Sorry if I've misunderstood the documentation...

library(icarus)
mar1 <- c("categ",3,80,90,60)
mar2 <- c("sexe",2,140,90,0)
mar3 <- c("service",2,100,130,0)
mar4 <- c("salaire", 0, 470000,0,0)
margins <- rbind(mar1, mar2, mar3, mar4)
wCal <- calibration(data = data_ex2,
marginMatrix = margins,
colWeights="poids",
method = "logit",
bounds = c(3, 38),
description=FALSE)
range(wCal)

Bonjour,

Nous avons un problème de convergence avec un nombre assez important de contraintes (environ 800-900: 238*17).
Les résultats de la fonction calibration ont l'air assez cohérent si ce n'est pour une des strates (celle avec 17 options) où l'on obtient:

• Before Calibration : que des valeurs nulles sauf une à 1,08
• After Calibration: que des valeurs nulles sauf la même à 1,07
• Margin: Correct correspondants aux poids respectifs des différentes modalités de la Strate dans la marginMatrix

Lorsque nous utiliser la méthode avec pénalité en relâchant les contraintes sur cette strate étrange, nous obtenons rigoureusement le même résultat.

Nous ne parvenons pas à en déterminer la cause.

Pourtant la convergence a l'air bonne à plusieurs égards. Centrée autour de 1 avec des margins matrix avant-après très similaires.

Auriez-vous un avis sur notre problème?
Merci d'avance.

Currently, implementation of min bounds calibration is purely R and (relying on slam's sparse matrices and Rglpk).

Implementing the algorithm in C++ could:

• decrease the execution time
• decrease memory usage
• decrease the number of dependencies required to make minimum calibration work

Percentages in Icarus have now to be entered with a sum equal to 1 (e.g. "0.4 - 0.2 - 0.4"), whereas in Calmar and Calmar 2, percentages sum have to be equal to 100 (e.g. "40 - 20 - 40").

Icarus should support both input types as long as parameter pct is set to TRUE.