We report a palaeomagnetic study of Icelandic lavas of late Miocene to late Pliocene age to test the geocentric axial dipole hypothesis at high northern latitudes. Cores were sampled from 125 sites in the Fljótsdalur valley in eastern Iceland, and hand samples were taken for 17 new incremental heating 40Ar/39Ar age determinations. 96 per cent of the cores were oriented using both a Brunton compass and a sun compass. Comparison of the magnetic and sun azimuths reveals deviations of ±5°, ±10° and ±20°, respectively, for 42, 16 and 3 per cent of the data points, indicating that core sampling intended for palaeosecular variation (PSV) studies at high northern latitudes should be oriented by sun. A total of 1279 independent specimens were subjected to AF- and thermal-demagnetization for palaeodirectional analysis, and well-grouped site mean directions were obtained for 123 sites of which 113 were found to be independent sites. Applying a selection criteria of k > 50 and N ≥ 5 (Nmean = 9.5), we obtain a combined grand mean direction for 46 normal and 53 reverse (for VGPlat > ±45°) polarity sites of declination = 5.6° and inclination = 77.5° that is not significantly different from that expected from a GAD field. The corresponding palaeomagnetic pole position (VGPlat = 86.3°N, VGPlon = 21.2°E, dp/dm = 4.0°/4.3°) is coincident with the North Pole within the 95 per cent confidence limits. An updated age model is constructed based on the 40Ar/39Ar ages, showing that the majority of the Fljótsdalur lavas fall within 2–7 Ma. We combine the Fljótsdalur data with existing data from the nearby Jökuldalur valley. The 154 palaeodirections are well-dispersed between 1 and 7 Ma and constitute a high-quality data set for PSV analysis. Our results partly support previous conclusions of a generally higher dispersion during reverse polarity intervals. However, when comparing our Matutayma data with Brunhes age data from Jan Mayen, we find no evidence of a higher VGP scatter during the Matuyama as previously suggested. When comparing our VGP scatter to the two commonly used models for VGP dispersion: Model G and TK03, we find a good fit for all 1–7 Ma VGP scatter data SB(1–7) to Model G, whereas SB(1–7) is not fitted by TK03, even when considering the uncertainty of SB(1–7). We also find that all VGP scatter estimates, except that for the Gilbert subset, are consistent with Model G, while the discrepancy with TK03 is generally larger.