[Ft] give cox baseline

[rimhajal]

it would be nice to have a function that adds the baseline H_0 function. this is a rough draft of the breslow estimator for that end:

function _breslow_est(model::SurvivalModels.CoxModel, t)
    r = 0.0
    exp_pred = exp.(model.des * model.par)
    for ti in model.times
        if ti < t
            num = 0.0
            den = 0.0
            for j in eachindex(model.times)
                num += model.status[j] * (model.times[j]=ti)
                den += exp_pred[j] * (model.times[j]>=ti)
            end
            r += num / den
        end
    end
    return r
end

survival_curve(model::SurvivalModels.CoxModel, t, X) = exp(-_breslow_est(model, t) * exp(X' * model.par))

[lrnv]

Thisone looks like it is doing the same thing but a bit faster:

function _breslow_est(model::SurvivalModels.CoxModel, t)
    r = 0.0
    us = unique(model.times)
    max_iu = searchsortedfirst(us, t)
    contribs = 1 ./ reverse(cumsum(reverse(exp.(model.des * model.par))))
    for i_u in 1:(max_iu-1)
        u = us[i_u]
        iiu = searchsortedfirst(model.times, u)
        count = sum(model.status[1:(iiu-1)])
        r += count * contribs[iiu]
    end
    return r
end

[rimhajal]

It's still the wrong values no?

[lrnv]

It is the same value, I do not know if it is right or wrong for the moment...

[lrnv]

@rimhajal i do not remember if we had a correctly working verison at the end ? if so, can you copy it here ?

[rimhajal]

the results were a lot better (when comparing the Cox survival curve to Kaplan Meier) but they didn't match the values in R i think.

function _breslow_est(model::SurvivalModels.CoxModel)
    t = unique(model.times)
    k = length(t)
    h = zeros(k)
    elp = exp.(model.des * model.par)
    for i in 1:k
        risk = elp .* (model.times .>= t[i])
        h[i] = sum(model.status .* (model.times .== t[i])) / sum(risk)
    end
    return cumsum(h)
end

[lrnv]

Ha now i remember, this version is actually OK, TY (even if a bit slow i think). Do you also have the website link we used for future reference ? I do not remember it..

[rimhajal]

you mean this one https://missingdatasolutions.rbind.io/2022/12/cox-baseline-hazard/ ?