I noticed this article in EOS recently (thanks to Jon Mound and Nick Swanson-Hysell on Twitter for the heads up), and thought I’d comment. Although I’m framing these as caveats, please don’t take the comments to be an attack on anyone, either the article’s author or the authors of the study it describes. I’m just trying to outline my thought process and the kinds of questions a paleomagnetist like me might have when I look at coverage of something in the popular press (which EOS is, sort of…).
Caveat 1: Be careful of the source. My first impulse when I see an article on paleomagnetism in EOS or another responsible publication is to look for the original study. As someone who works in this field, I want to know the details: how was the data analysis done? What dataset was used, exactly? How does it fit into the context of work that’s come before? (I can usually figure this out by myself, but maybe I’ve missed something.) Were there any checks to make sure the result is plausible versus being an illusion of how the data were processed or a bias in the dataset? The problem here is that the results reported in EOS were described in a talk, not a peer-reviewed paper. The talk was by Kirschner and co-authors at the European Geosciences Union (EGU) conference a few months ago. EGU is kind of analogous to the American Geophysical Union conference here. The abstract from the talk, which is all I have to go on, is here. There’s actually a lot in it that didn’t make it into the EOS article, but not the details I’m looking for. (Also, there were some other neat talks in that session at EGU that I wish I’d heard!)
This isn’t to say that science journalists should never write about talks – of course they should. Conferences are where we share current work in progress. But as a reader, when you see that a story is based on a talk, know that there are some questions about the work that might not be answered – or answerable – yet.
Caveat 2: The Data. Estimates of Earth’s magnetic field intensity are much harder to deal with than standard paleomagnetic data. This is in part because the intensity of a rock’s magnetization has a complicated relationship with the intensity of the magnetic field in which the rock was magnetized. You can, for example, collect samples from two basalts that cooled at the same time, and so were magnetized in the same field. The magnetizations that you measure from your two basalt samples might be vastly different for a number of reasons. For one thing, one basalt might have more magnetite in it than the other. (Titanium-bearing) magnetite is the mineral that is mostly responsible for the magnetic record in basalt. Alternatively, differently sized or shaped or aligned titano-magnetite particles may have led one of the basalts to record the magnetic field more efficiently than the other. Alternatively, one of the basalts may have had its magnetic record wiped clean (by being reheated, for example, or chemically changed), or may have been remagnetized by a lightning strike, or may just have lost part of its magnetic record by sitting around in changing magnetic fields for a long time (we call that “viscous decay”). Over the years, various techniques have been developed to screen for these effects, and in some cases to adjust for them. But techniques do matter, and sometimes applying the wrong techniques or not applying the proper adjustments may bias estimates of ancient magnetic fields.
This is relevant because in the EOS article, the study is described as using “all available data”. The PINT database contains published results from hundreds of studies that attempt to estimate Earth’s ancient magnetic field intensity. This includes some from studies in the 1950s that don’t check for many of the problems that we know exist, some results that use different kinds of screening techniques, different ways to estimate amounts of magnetic material or the efficiency of magnetization, and different corrections for the weird effects I described. So it’s sometimes difficult to compare data from one study – one data point in the PINT database – to another. The usual approach to comparing paleointensity estimates through time is to come up with a set of criteria (based on the type of intensity estimate, or on the number of checks for a sample’s “ideal” behavior, or on agreement between different types of estimates) and look at data that meet those criteria. In fact, that appears to be what the authors of the study in question did in addition to the analysis of the whole dataset (From the abstract: “Spectral analysis of all palaeointensity data and a quality-filtered dataset obtained from the palaeointensity database…”).
Now, even if you decided to compare only the same kind of estimates of magnetic field intensity through time, there would still be some issues with the data. For example, different rock types may require different checks or adjustments, or may respond differently to the same estimation technique… and those rocks may be more or less common through different parts of the geological time scale. Intensity estimates from one particularly time-constrained rock type, basaltic glass, are really only available for the past 180 million years. Basaltic glass records magnetic fields differently from basalt (mainly because the magnetic particles are much smaller, and glass cooled more rapidly than basalt), and different adjustments for cooling rate (if you agree with them, which not everyone does) may need to be applied. Geology matters!
Caveat 3: The analysis. OK, so even if the data are filtered so that they are all reliable and comparable in terms of technique, estimates of past magnetic field intensity are unevenly spaced in time. Not all rocks are appropriate to use in intensity experiments. The older the rocks are, the fewer usable ones there are that have withstood the ravages of weathering, metamorphism, reheating, lightning strikes, and tectonism. This makes for a dataset that unevenly samples the magnetic field through time. One of the assumptions of digital signal processing- which the authors of the study in question do – is that the data are (at least close to) evenly spaced in time. There are ways to get around this requirement, say by fitting a smooth curve to the data and re-sampling it – but those require some assumptions about the data and the underlying process. Because this study was reported in a talk, I’m not sure what those assumptions were. Nonetheless, any smoothing, fitting, or filtering done to the data could strongly influence the cyclic behavior that the authors describe. I’m not trying to imply that the study’s authors are signal processing newbies – they may very well know more than I do about it. I’m just highlighting a question that I’d ask if I were trying to evaluate the study.
By the way, more than the signal processing aspects, I’m interested in the authors’ work on the change in the distribution of intensity estimates through time. This isn’t mentioned in the EOS article, but is in the abstract – I’m not sure why. Here’s my summary: When we talk about a “distribution”, we’re imagining that the intensity estimates are like students’ grades in a class: a random bunch of numbers to an outsider. If you take all of those grades together, they fall within some range of a middle value, with some grades are more frequent than others. If you taught the class again the same way, but with a different set of students, the pattern of grades would probably look somewhat similar even though the specifics would be different. In statistical terms, the underlying pattern of numbers (grades or intensity estimates) is a probability distribution. The authors, in their abstract, note that they see a change in the probability distribution of intensity estimates around 1.3 billion years ago, ” coincident with the time that geologic and palaeogeographic evidence suggests the onset of quasiperiodic assembly and fragmentation of supercontinents.” It’s also within the time frame that recent work (if you believe it) has suggested that the inner core began to grow. So is the cause of the change external (tectonic-related), internal (inner-core-driven), or neither? I’m not sure, but it adds another intriguing possibility into the mix! (I guess that’s Caveat 4: The Interpretation.) The approach that the study’s authors take to look at changes in probability distributions isn’t described in the abstract, but it might work better than the signal processing approach with data that are unevenly scattered through time.
Caveat 5: Is it a New Result? The idea that tectonics influence Earth’s magnetic field has been popular at least since the late 1990s, when geodynamo simulations suggested that different patterns of heat flux at the core-mantle boundary – due to patterns of cold, subducting lithospheric slabs – could influence magnetic reversals (see Glatzmaier et al., Nature, 1999). Since then, there have been a number of studies looking for tectonic-related cycles in intensity data (see here and here for example; both studies’ authors are quoted in the EOS article). So the research problem isn’t new, but I’m not sure that anyone has succeeded at the signal processing approach.
I’m sure other people have other ideas or concerns about the study or the way it was reported in EOS. The things I’d like to know about the Kirschner et al. study represent my own perspective as someone who used to do this stuff. Maybe you have a different set of questions? Please comment below!