K fold Cross Validation Accuracy (mnrfit)

Hello,

I am trying to calculate cross validation accuracy of an ordinal model (using mnrfit and mnrval) as follows:

K=10;
indices=crossvalind('Kfold',DS,K);

for i_fold=1:K

      test = (indices == i_fold);
      train = ~test;

train_x=variable(train,:);
train_y=DS(train,:);
test_x=variable(test,:);
test_y=DS(test,:);

[b,dev,stats] = mnrfit(train_x,train_y,'interactions','off','model','ordinal','link','logit');
y_hat = mnrval(b,test_x,stats,'interactions','off','type','cumulative','model','ordinal','link','logit');
ccdf_yfit=1-y_hat;
class1 = ccdf_yfit > 0.5;

for i=1:length(test_x)
index_pred(i)=max(find(class1(i,:)))+1;
index_pred(i)=index_pred(i)';
Diff=abs(test_y-index_pred);

 Pen_Acc=zeros(K,1);

   if Diff == 0
       Pen_Acc(i_fold)=1;
   elseif Diff == 1
       Pen_Acc(i_fold) = 1-0.2;
   elseif Diff == 2
       Pen_Acc(i_fold) = 1-0.4;
   elseif Diff == 3
       Pen_Acc(i_fold) = 1-0.6;
   elseif Diff == 4
       Pen_Acc(i_fold) = 1-0.8;
   else
       Pen_Acc(i_fold) = 0
  end

end
end

DS are my response labels, and variable is my predictor.

My problem is, as I try to calculate 'Diff', I get an error telling me that test_y and index_pred are not the same size. For each value of i_fold, I want a vector of predicted classes index_pred that I can compare to the real labels in the 'test' subset. Any idea on what I am doing wrong / how I can achieve this?

Many thanks, IC

[Sedo]

Did you ever find an aswer?