fof-visually-guided/elutils/+stats/bootsigmoid_val.m

function [p,D]=bootsigmoid(A,B,varargin)
% [p,D]=bootsigmoid(A,B,varargin)
% Takes two sets of N x 2 binomial data A and B, and fits sigmoids to them
% both.  It then uses the mean and covariance of the fits to calculate the
% distance (using the projection of the fits onto fisher's linear
% discriminant) of the fits.  
% Then we permute the rows of A and B to generate permuted data sets and
% perform the same fits and estimates of "distance". Finally, the distance
% between A and B fits are compared to the distribution of distances
% generated by permuting A and B.


BOOTS=100000;

utils.overridedefaults(who,varargin);
import stats.*

[dAB, D]=get_dist(A,B);
n_A=size(A,1);
n_B=size(B,1);

permAB=nan(BOOTS,1);

M=[A;B];

parfor bx=1:BOOTS
    
    rperm=randperm(n_A+n_B);
    rA=M(rperm(1:n_A),:);
    rB=M(rperm((n_A+1):end),:);
    
    
    permAB(bx)=get_dist(rA,rB);
    
end

p= stats.get_p(dAB,permAB);

D.permAB=permAB;
D.p=p;

end %sfunction

function [dAB,D]=get_dist(A,B)

n_A=size(A,1);
n_B=size(B,1);


aidx=randperm(n_A);
bidx=randperm(n_B);

Alim=floor(n_A/2);
Blim=floor(n_B/2);

trainAdx=aidx(1:Alim);
trainBdx=bidx(1:Blim);

testAdx=aidx(Alim+1:end);
testBdx=bidx(Blim+1:end);

[nbetaA,~,~,covA,~]=nlinfit(A(trainAdx,1),A(trainAdx,2),@sig4,[0 1 0 10]);
[nbetaB,~,~,covB,~]=nlinfit(B(trainBdx,1),B(trainBdx,2),@sig4,[0 1 0 10]);

vAB=flda(nbetaA,nbetaB,covA,covB,Alim,Blim);


[betaA,~,~,~,~]=nlinfit(A(testAdx,1),A(testAdx,2),@sig4,[0 1 0 10]);
[betaB,~,~,~,~]=nlinfit(B(testBdx,1),B(testBdx,2),@sig4,[0 1 0 10]);

dAB=abs(betaA*vAB-betaB*vAB);

D.trainbetaA=nbetaA;
D.trainbetaB=nbetaB;
D.testbetaA=betaA;
D.testbetaB=betaB;
D.covA=covA;
D.covB=covB;
D.vAB=vAB;
D.dAB=dAB;

end % subfunction