Figure 7. Effect of distributing 100 paleomagnetic samples over a varying number of paleomagnetic sites on the a) angular distance of the acquired paleopole to the reference pole with all sites with n=7 (or 6, in case of Antarctica), and the b) A95, c) K, and d) S of the acquired paleopole. Every datapoint represents 1000 runs of a distribution, with error bars indicating the 95% confidence interval of these 1000 runs.
5 Discussion and conclusions
We first evaluate whether measuring multiple samples per site to constrain within-site scatter allows to objectively filter unreliable VGPs. In our reinterpretation, we closed our ‘expert eye’ and included sites that were excluded in the original studies because of subjective criteria: for instance, our dataset contains sites with erratic demagnetization behavior or very high intensities that were originally interpreted as lightning struck. Our analysis illustrates that objective statistical filters, such as a k-cutoff, a minimum number of samples per site, or eliminating the farthest outliers per site, are incapable of filtering these unreliable VGPs. The subjective judgement of the expert thus determines which samples and sites are included in the calculation of the paleopole (and makes it important that all data are made available following the FAIR principles (Wilkinson et al., 2016)).
Our analysis next shows that for the three datasets with high between-site scatter and low within-site scatter (i.e., Turkey, Mongolia, and Antarctica), applying the within-site scatter-based criteria leads to marginal increase in data clustering (i.e., an increase in K and decrease in S). For the Norwegian dataset, which has the rare combination of high within-site scatter and low between-site scatter, these criteria even lead to significantly higher clustering of the data. This lends support to using these data filters for PSV studies.
But applying these filters when calculating paleopoles in all cases leads to a decrease in pole precision (i.e., an increase in A95). This is the result of the decrease in N that is caused by applying data filters. Our experiments thus show that the effect of between-site scatter on pole precision is far dominant over within-site scatter. And because the accuracy of paleopoles also decreases with decreasing N (Vaes et al., 2021), applying within-site scatter filters systematically decreases pole accuracy as well as precision (see also Figure 7).
Averaging multiple samples per site, without applying filters, does not negatively influence pole accuracy or precision. In case only few sites are available, or there is reason to question data quality, collecting multiple samples per site, as recommended by Meert et al. (2021), may still be a useful strategy. It may also be useful to demonstrate for a selection of lava sites in a study that within-site precision (k ) is high, as a field test equivalent to a fold, reversal, or conglomerate test. But, our analysis shows that within-site scatter does not significantly influence paleopole accuracy or precision. The extra efforts put into collecting multiple samples per site are thus more effectively spent collecting more single-sample sites.
Acknowledgements
BV and DJJvH acknowledge NWO Vici grant 865.17.001 to DJJvH.
Data Availability Statement
All uninterpreted paleomagnetic data are publicly available on Paleomagnetism.org for the Mongolia, Norway, and Turkey dataset and in the MagIC database for the Antarctica dataset: https://earthref.org/MagIC/17076. The Python code used for our analysis will be made public on GitHub upon acceptance of this manuscript.