Myocardial blood flow (rMBF) can be measured using dynamic positron emission tomography (PET) and bolus injection of H2(15)O. Recent studies indicate that large errors in the estimates of flow (f) can be produced by time shifts between the true arterial input function and the measured input function [A(t)]. We have investigated this phenomenon further using A(t) derived from patient data, and using simulated myocardial time activity curves [M(t)]. We found that with judicious choice of scan parameters and region of interest (ROI) placement, these errors can be greatly reduced. In particular, when A(t) is measured from the left ventricular (LV) cavity, the bias in f is negligible over a wide range of circumstances. However, when A(t) is not measured from the LV cavity, the bias in flow can be large for short scans (< 2 min) or low flow values (f < 0.4 ml/g/min). We show that the bias is primarily due to the spill-over term in the model that is most commonly used to compute rMBF and suggest some correction methods. We conclude that it is possible to avoid errors in estimates of flow due to time delay.