How is Vesuvius like predicting the U.S. presidential election? Image: Antonsusi / Wikimedia Commons Unless you've been living on the far side of the moon for the past two years, you'll know that the U.S. presidential election is next week. Now, I'm by no means going to get into politics here, but I am going to talk a little bit about the controversy regarding the analysis of polls/data to project who might win the election. I have been a fan of Nate Silver, the head of 538, for many years now, starting back when he was working for Baseball Prospectus creating the PECOTA projection system for baseball statistics. Now he's headlong into examining how polling and previous electoral, demographic and economic data can be used to model the outcome of an election - and this has rubbed pundits who use, well, beats me, to try to "predict" the outcome of the election, the wrong way. Let's think about what I just said - we have someone collecting and analyzing data to look at the potential outcomes of an election versus people who just think they know what is going to happen. Sound familiar? Volcanology, and in particular, modern volcanic hazard assessment, does exactly the same thing that Nate Silver does at 538. Geologists collect data of what the volcano has done in the past (similar to previous voting behavior), look at what volcanoes like the one in question do when an eruption occurs (demographic), assess the current state of the volcano through gas monitoring, seismometers, tilt meters - both for recent and long term trends (polling and economic). Once all that data is in hand, the volcanologists trying to assess the potential hazard will construct a probability/event tree (see below), which is a map to what may occur at a given volcano when it begins to show signs of eruption (election). This tree doesn't tell what the volcano is going to do -- that is, it doesn't predict the behavior of the volcano during a specific period of unrest. What it does is offer the probability of certain events, that one type of eruption is more or less likely than another. This way, hazard planners can have a sense of what to prepare for as the most likely events, while keeping in mind what some of the least likely events might be. Let's look at this example event tree for Vesuvius in Italy (published in Neri et al, 2008; see below). If you start on the left, you can follow different paths that have different probabilities of occurrence. The red value listed under the event is their model's most likely probability, while the two black numbers on each side are the ranges of values that come out of their model. So, if Vesuvius shows unrest, there is a 99.97% chance that NO sector collapse will happen. This doesn't rule out a sector collapse as we will have that 0.03% chance, but it is highly unlikely. Moving right, the unrest has a 40% chance of producing an eruption, so with any given unrest, we are more likely to see no eruption rather than eruptive activity. Keep on heading right and we have a 77% chance of an explosive eruption IF an eruption does occur. The probabilities are cumulative, so you can start with eruption/no eruption and if there is an eruption, you have a ~3% chance, overall, of it producing a Plinian column (4% of the total 77% chance of an explosive eruption).
An event tree with probability envelopes for potential activity at Vesuvius in Italy. Figure: Neri et al., 2008, Journal of Volcanology and Geothermal Research You can look at the election same way - a series of stacked events that all have different probabilities - all built off of data and observations of the Mt. Election (as it were). Just like Vesuvius, you can say that if there is magmatic unrest, then we have a ~40% chance of an eruption, but this doesn't mean we won't have one merely because it is the less likely event. All events on the tree are possible, just some are more likely than others. However, the real challenge (as we've seen with the recent l'Aquila trial) is trying to convey what path you as a hazard geologist (or election modeler) think the eruption/election will take. This is where interpretation meets scientific data. Unless the data you suggests otherwise, you want to follow the highest probability path -- for Vesusius (assuming an eruption), it would be violent strombolian eruptions. However, as you collect more data and examine the monitoring (polling), you might begin to see that the data is pointing towards a lower level event, like an effusive lava flow. However, it takes special circumstances and data to back it up to follow those other paths. It is the job of the geologists to determine what the data is trying to tell us about the future activity. So, let's say you make the call about the volcano's future activity based on the probability and you're WRONG. Do you throw the whole model out the window? No - what this means is either you might have missed a key component of the data or your model needs adjustments. Models are human constructs to try to represent natural behavior, so they can (and likely will) be flawed to some extent. You have to go back and see where the model deviated from the actual events so next time, it works better. That is how scientific modeling works, so if Mt. Election doesn't do what 538's model suggests, then we know that it needs some work or the data had unforeseen issues. You go back and try to fix it based on more observations and data. However, if you're the type who just likes to prognosticate without any data - maybe you like to predict earthquakes based on the existence of some unseen dark solar companion or you're a politic pundit paid to talk - then you can sit back and make predictions without fear. This is because, unlike a hazard event tree or 538's model, where the evidence is there to see how you made the decision and how the model is built, you leave no evidence. So, if you didn't make the right call, well, that's life. Anyone can make a prediction, and that is the real danger. Predictions and probability should not be confused, which is why something like 538 or an event tree is fundamentally different than the talking heads on television. An 80% chance of something happening is NOT the same as saying that the something WILL happen, and many people inside and outside the mainstream media don't seem to understand (or want to understand) this - both for volcanic unrest and elections.