Unfortunately, the MATLAB functions that compute the sample cross-correlation from two provided sequences (such as "xcorr" in the Signal Processing Toolbox) cannot deal with incomplete data.
One possible solution is to find the NaN value's in the signals and to ignore the measurements (at that time value) for both sequences. The following code shows how you can remove those NaN's:
s2NanLoc = find(isnan(s2));
s1NanLoc = find(isnan(s1));
nanLocs = unique([s1NanLoc; s2NanLoc]);
s1(nanLocs) = ;
s2(nanLocs) = ;
When I do this with your dataset I am able to get a "td" of 4.