I am using Logistic Regression with count representation. So, for any feature-tuple, I have few 0's(negative class) and few 1's. I duplicate each row, one for target 0 and other for target 1, And I use, number of 0's as sample weights for the row with target 0 and number of 1's as sample weights for the row with target 1.

feature1--feature2--#0s--#1s x1--x2--20--10

gets pre-processed as

feature1--feature2--target--sample_weights x1--x2--0--20 x1--x2--1--10 I am using sandwich estimator for estimating covariance matrix of the coefficients as described in https://web.stanford.edu/class/stats200/Lecture26.pdf

sigma=inverse(X.T * W_hat * X ) * (X. T * W_tilda * X ) * inverse(X.T * W_hat * X )

Where, X, an n by k matrix, contains all observations in row-manner (so, i_th observation (x_i1,x_i2,...,x_ik) constitute the i_th row)

W_hat, an n by n matrix, is diag((y_1_hat*(1-y_1_hat), (y_2_hat*(1-y_2_hat)........(y_n_hat*(1-y_n_hat))

W_tilda, an n by n matrix, diag((y_1-y_1_hat)^2, (y_2-y_2_hat)^2,.......,(y_n-y_n_hat)^2)

How, can I estimate covariance matrix for my model which contains same feature tuple for both targets along with sample weights?

I found one article, which suggests multiplying with each row of X and W_hat and W_tilda with sqrt(sample_weights). But, I can't have duplicate rows in my X matrix(otherwise , it would be singular). https://www.parisschoolofeconomics.eu/docs/dupraz-yannick/using-weights-in-stata(1).pdf

Anyways, I am not sure if multiplying with sqrt(sample_weights) is the correct way of handling this situation.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.