Curva ROC en modelos de respuesta cualitativa (Logit, Probit)

[atorre]

Hola MMS,
¿se podría incorporar a los resultados de las estimaciones de modelos Logit y Probit las curvas ROC?

Sigue a continuación una NameBlock con la implementación del cálculo:

NameBlock ROC =
[[

Set _.interp.Set = SetOfText("linear", "akima", "cspline");

Set _Interp
(

Set xSet, Set xSet with we want to know y for all funSet
Set setFunSet, Set of tables with same structure SetOfReal(x, y)
Set interpSet Set with gsl valid interpolate method (see gsl_interp)

)
{

Set setxyFunSet = EvalSet(setFunSet, Set(Set funSet)
{

SetOfMatrix(SetCol(Traspose(funSet)[1]), SetCol(Traspose(funSet)[2]))

});

EvalSet(xSet, Set(Real x)
{

Set ySet = EvalSet(setxyFunSet, Real(Set xyFunSet)
{

Matrix xFunSet = xyFunSet[1];
Matrix yFunSet = xyFunSet[2];

Set ySet = EvalSet(interpSet, Real(Text code)
{

Code gslInterp = gsl_interp(code, xFunSet, yFunSet);
gslInterp(0, x)

});
SetAvr(ySet)

});
SetOfReal(x)<<ySet

})

};

Set Curve(Matrix y, Matrix py, Set sliceSet)
{

Set graf = EvalSet(sliceSet, Set(Real slice)
{

Matrix yEst = GE(py, Rand(Rows(py), 1 , slice, slice));
Real VP = MatSum(And(yEst, y));
Real FN = MatSum(And(Not(yEst), y));
Real VN = MatSum(And(Not(yEst), Not(y)));
Real FP = MatSum(And(yEst, Not(y)));

Real TVP = VP/(VP+FN);
Real TFP = 1-VN/(VN+FP);=FP/(VN+FP)
SetOfReal(TFP, TVP, slice)

});
Set claTFP = Classify(graf, Real(Set reg1, Set reg2)
{

Real tfp1 = reg1[1];
Real tfp2 = reg2[1];
Compare(tfp1, tfp2)

});
EvalSet(claTFP, Set(Set class)
{

Real maxTvp = SetMax(Traspose(class)[2]);
Select(class, Real(Set reg){EQ(reg[2], maxTvp)})[1]

})

};

Set Curve.Abs
(

Set ROCSet, Table with reg = SetOfReal(TFP, TVP, slice, ...)
Real M, Total
Real m Infected

)
{

Real p = m/(M-m);
EvalSet(ROCSet, Set(Set reg)
{

Real TFP = reg[1];
Real TVP = reg[2];
Real slice = reg[3];
Real infected = TVP*m;
Real population = (p*TVP+TFP)*M/(1+p);
SetOfReal(TFP, TVP, population, infected, slice)

})

};

Set Eval.TFPSet(Set tfpSet, Set ROCSet)
{

Set tROCSet = Traspose(ROCSet);
Set txFP = tROCSet[1];
Set txVP = tROCSet[2];
Set txSL = tROCSet[3];

Set setFunSet = SetOfSet
(

Traspose(SetOfSet(txFP, txVP)),
Traspose(SetOfSet(txFP, txSL))

);
_Interp(tfpSet, setFunSet, _.interp.Set)

};

Set EvalAbs.PopSet(Set popSet, Set ROCAbsSet)
{

Set tROCSet = Traspose(ROCAbsSet);
Set txPop = tROCSet[3];
Set txInf = tROCSet[4];
Set txSL = tROCSet[5];

Set setFunSet = SetOfSet
(

Traspose(SetOfSet(txPop, txInf)),
Traspose(SetOfSet(txPop, txSL))

);
_Interp(popSet, setFunSet, _.interp.Set)

};

Real Get.Area(Set ROCSet)
{

Set txFP = Traspose(ROCSet)[1];
Set txVP = Traspose(ROCSet)[2];
Set setFunSet = SetOfSet( Traspose(SetOfSet(txFP, txVP)) );
Real get.tvp(Real tfp)
{

Set tfpSet = SetOfReal(tfp);
Real tvp = _Interp(tfpSet, setFunSet, _.interp.Set)[1][2];
tvp

};
IntegrateQAG(get.tvp, 0, 1)

}

]];

Sería ideal poder acceder a estos resultados con una opción de botón derecho desde la estimación.

Siguen un par de links en los que se describe el concepto de las curvas ROC:

​http://es.wikipedia.org/wiki/Curva_ROC
​http://www.fisterra.com/mbe/investiga/curvas_roc/curvas_roc.htm