Why don't certain sensor arrangements work with my Kalman filter simulations?

I'm simulating a quarter car model suspension system. The model details can be found at http://nopr.niscair.res.in/bitstream/123456789/7546/1/IJEMS%2013(3)%20173-179.pdf. I'm simply adding a Kalman filter to estimate the states as I'm also including measurement noise and the road profile as a the process disturbance.

The problem is that with the sensor arrangement that I believe is realistic, the state estimates are way off. The arrangement is (the 6 state variables in order: passenger displacement, its velocity, sprung mass displacement, its velocity, unsprung mass displacement and its velocity) the measurement matrix

C=[1 0 -1 0 0 0; 0 0 1 0 0 -1]

which is simply the relative distances between the passenger and the body of the vehicle, and the tire and the body of the vehicle. Here's a list of some good and bad measurement matrices:

C=[1 0 1 0 0 0; 0 0 1 0 0 1] OK

C=[-1 0 -1 0 0 0; 0 0 -1 0 0 -1] OK

C=[1 0 -1 0 0 0; 0 0 1 0 0 1] OK

C=[-1 0 1 0 0 0; 0 0 1 0 0 -1] NOT OK

C=[1 0 -1 0 0 0; 0 0 -1 0 0 1] NOT OK

... and so on...

So, to conclude, I don't understand why some sensor arrangements work and some don't, all of them make an observable system but only some of them produce accurate state estimates. I might use another arrangement, why not, if it's a plausible one such as the 'realistic' one that I had introduced. However, for example perfectly functional and nicely illustrative measurement matrices such as

C=[1 0 0 0 0 0; 0 0 1 0 0 0; 0 0 0 0 1 0]

are not realistic in my opinion.

Here are the Simulink file and the script to initiate: https://dl.dropbox.com/u/51236239/MATLAB/Suspmodel1FAILSENSOR.mdl https://dl.dropbox.com/u/51236239/MATLAB/SuspscriptFAILSENSOR.m