I'm getting some strange results from a closest facility problem (~40 facilities, ~1100 incidents) in Network analyst. In addition to a Total_Length, which is my impedance measure, I am calculating lengths for each route within several mutually exclusive and exhaustive categories. In other words, when summed for any edge in my network dataset, they add to Total_Length for that edge. Let's call them LengthA, LengthB, LengthC etc. I do this by setting them as Accumulation Attributes.
The resulting Routes look fine, except that the Total_Length of each is not the same as the sum of the lengths by category. The difference is sometimes too large to be because e.g. I am assigning locations to the nearest network location within 5000 meters.
Here's the part that makes me think someone out there will know the answer:
All datasets I am using are in the same feature dataset in a sinusoidal projection. The percent difference between Total_Length and Sum(LengthA, LengthB, LengthC...) is highly correlated (0.85) with the percent difference between the analogous straight line distances calculated using latitude/longitude in the Haversine formula and using projected (sinusoidal) coordinates.
I can't think of any reason why the Plate Carree ("Geographic") or any projection other than the original sinusoidal would be used for one set of calculations but not the other.
I'm running 9.3 SP1 with an ArcInfo license.
Any ideas would be greatly appreciated.