OK - let me try. Rates don't have returns, but the underlying bonds do. In my current world, we actually calculate tracking errors (as close to VaR as I would care to get) on bond portfolios based on the log returns of the underlying bonds. But one reasonable way to approximate that is to add up the product of the basis point change in a set of key rates and the corresponding key rate durations. So this may be the reason why you have come across the basis point change approach to interest rate risk. More generally, if you do your VaR calculations based on a full fledged interest rate model (as opposed to based on key rate durations and some sort of principal components), you should let the model dynamics guide your interest rate shocks. If the model is lognormal (which I would not necessarily recommend), then you should focus on the relative interest rate moves, although I still wouldn't refer to them as "returns".