A method of measuring overlay uses a plurality of asymmetry measurements from locations (LOI) on a pair of sub-targets (1032, 1034) formed on a substrate (W). For each sub-target, the plurality of asymmetry measurements are fitted to at least one expected relationship (1502, 1504) between asymmetry and overlay, based on a known bias variation deigned into the sub-targets. Continuous bias variation in one example is provided by varying the pitch of top and bottom gratings (P1/P2). Bias variations between the sub-targets of the pair are equal and opposite (P2/P1). Overlay (OV) is calculated based on a relative shift (xs) between the fitted relationships for the two sub-targets. The step of fitting asymmetry measurements to at least one expected relationship includes wholly or partially discounting measurements (1506, 1508, 1510) that deviate from the expected relationship and/or fall outside a particular segment of the fitted relationship.