OBJECTIVE: To assess the reproducibility of time and frequency domain variables derived from the signal averaged P wave. DESIGN: Longitudinal within patient study. SETTING: Regional cardiothoracic centre. PATIENTS: 20 patients (10 with documented paroxysmal atrial fibrillation and 10 normal controls) were studied on three occasions to assess the reproducibility of repeated signal averaged P wave recordings. Digital P wave recordings were made on a further 10 patients on a single occasion and the recordings signal averaged twice in order to assess the reproducibility of the averaging system itself in the absence of biological variation. MAIN OUTCOME MEASURES: P wave duration, spatial velocity, and energies contained in frequency bands from 20, 30, and 60-150 Hz of the P wave spectrum were measured after P wave specific signal averaging. Coefficients of reproducibility were calculated for paired signal averaged P waves derived by signal averaging the same digital recordings on two separate occasions, for recordings performed in the same patients immediately after each other ("back to back") and those performed one week apart. RESULTS: System reproducibility when the same digital P wave recordings were signal averaged on two separate occasions was high (< 11% for all variables). For P wave duration the coefficient of reproducibility was 11.4% for back to back recordings and 13.1% for those one week apart. The reproducibility of spatial velocity and P wave energy was low. Variation in P wave morphology was noted when successive P waves from the same subject were examined. If recordings with the same P wave morphology were analysed the reproducibility of spatial velocity and P wave energy improved but remained significantly poorer than that for P wave duration. CONCLUSIONS: P wave duration is reproducible within subjects in the short and medium term. Frequency domain and spatial velocity analysis are poorly reproducible, due more to spontaneous variation in P wave morphology than to instability of the signal averaging process. This may limit the utility of signal averaged P wave variables other than duration for the prediction of atrial arrhythmia.