The PARFOR loop cannot run due to the way the variable is used

I do not believe there is a way to further vectorize this code. Please let me know if that is untrue. I would like to parallelize the below code. The goal of the routine is to take the structure array "data", extract values from the fields "goPower" and "ngPower" for each element in the struct, perform some operation, and store the result in a 2-D matrix of size 1501 x number_of_elements_in_struct. (1501x7) for example. I expect MatLab to distribute iterations of the PARFOR loop (corresponding to permi looping index) to workers and have each worker perform the inside loop separately. Since each "subject" is different and each iteration of permi does something different, I don't think I have any dependencies. MatLab still says "The PARFOR loop cannot run due to the way the variable "diff" is used. What am I doing wrong?

parfor permi = 1:npermutations
    
    for subject = 1:numel(data)

        % Get indices for subsampled set of Go trials = to # of No-Go trials
        idxs = randsample(size(data(subject).goPower,2),size(data(subject).ngPower,2),true);
        
        % Combine trials
        allTrials = cat(3,data(subject).goPower(:,idxs,:),data(subject).ngPower);
    
        % Making mapping vector
        mapping = [ones(size(allTrials,3)/2,1); ones(size(allTrials,3)/2,1)+1];
        
        % Shuffle the mapping vector
        new_map = mapping(randperm(numel(mapping)));
    
        % Calculate mean difference
        diff(:,subject) = mean( mean( allTrials(:,new_map==2,:) ,2) - mean( allTrials(:,new_map==1,:) ,2) ,3);
    end
    
    % Store values for this iteration
    perm_diffs(permi,:) = mean(diff,2);
end

Answer 1

[Matt J]

Pre-allocate temporary variables like diff within the parfor loop.

parfor permi = 1:npermutations

    diff=nan(M,N); %pre-allocate diff  
    
    for subject = 1:numel(data)