with rxlev access min = -100dBm, you still have a poor QoS ?
I would check the path balance of this cell, perform drive-tests, check interference, and check the QoS per TRX. Overall : try to be certain that the hardware is good and the frequencies are clear.
If everything is fine, I would check power control parameters, possibly inhibit DL and UL power control for few hours and check the imapact. I would also implement synth. frequency hopping on a TRX, and force traffic to go on this SFH TRX (to benefit from frequency diversity).
Then I would continue reducing the RXlev access min until I see an improvement. Compare those results with your drive-tests. Where is, geographically, the problem ? Can’t you add an aditionnal site to cater for this area ?
For information, what are the values for:
the Call Drop Rate,
the TCH assignment failure rate, the path balance,
the average timing advance,
the average Rx Lev downlink,
and the traffic in erlang ?