Yeah, compute the means first and then do the comparison over the valid regions. You can augment sizes later however you so choose...
>> wt=[0.5 0.5]; % filter weights for window 2 means
>> mn=filter(wt,1,p,nan) % compute means w/ valid ranges only
mn =
NaN NaN NaN
NaN NaN 3.5000
NaN 1.5000 4.5000
1.5000 1.0000 3.0000
1.5000 2.0000 1.5000
2.5000 1.5000 NaN
2.0000 1.0000 NaN
2.5000 NaN NaN
>> w=double(p(3:end,:)>mn(2:end-1,:)); % weighting comparison over valid range
>> w(isnan(p(3:end,:))|isnan(mn(2:end-1,:)))=nan % identify missing value locations
w =
NaN NaN 1
NaN 0 0
1 1 0
1 0 NaN
0 1 NaN
1 NaN NaN
>>
NB: There are only N-2 valid locations; the available data from the third to the last is compared to the first valid mean (2nd in the array) to the next-to-last since there's no observation against which to compare the last valid computed mean. This leaves a computed array of six observations from your initial eight points; as noted you can adjust this however needed to fit the remainder of the application by augmenting whichever end needs it for desired alignment.
