You can use the function isnan to find the appropriate values to replace as in the following code:
p = [
NaN 2 NaN;
2 4 NaN;
2 6 8;
NaN 8 2;
NaN NaN 1;
NaN NaN 1
];
p1 = p(1:end-1,:); % NaN replacement sources
p2 = p(2:end,:); % NaN replacement candidates
replace = isnan(p2) & ~isnan(p1);
p2(replace) = p1(replace)/2; % replace the NaN's that met your criteria
p2 = [p(1,:);p2]; % put back the first row, which never changes
disp(p2)
This code yields your expected result for your example. I am not completely certain what you expect to happen for NaN's with preceding values that aren't the last in the column. If you want to leave those as NaN and not replace them, you will need an additional step. After creating the logical indices variable, replace, use the following code to keep only the last true value in each column.
replace = replace & flipud(cumsum(flipud(replace)))==1;
You can use this test value to see the difference in the two interpretations.
p = [
NaN 2 NaN;
2 4 NaN;
2 6 8;
NaN 8 NaN;
NaN NaN 2;
NaN NaN NaN
];
