The Nadaraya Watson regression curve estimator is given as a ratio m = f/g. Under very mild assumptions (in particular not including any continuity of the regression function or design density), the uniform asymptotic deviation of the expectation Em from the ratio Er/Eg (a quantity usually appearing in inspections of the asymptotic properties of m) is investigated. Our approach covers several types of data-dependent bandwidths and remains valid for classes of regression functions.