Estimates should be made in abstract units (e.g. 'function points'), then converted to man-days or dollars or whatever using a scaling factor specific to the estimator. That way, the actual value of the estimate is moot; it's the consistency of the estimates that counts.