Bayes beats baying banalities

Are people sick of politics, sick of the way it is practised, sick of the way it is reported in the media or some combination of all of the above? For the blog it’s mainly media coverage and the obsessive and largely irrelevant speculation and prediction which accompanies it. Indeed, about the only reason for reading most political reportage is to identify who inspired it, why, and whether as a result it is likely to be wrong or right.

But the most depressing aspect of print media coverage for anyone still interested in politics is undoubtedly the obsession with polls and predictions about what might or might not happen in the next day, week, month or year. As most of the predictions are braying banalities, wrong or wishful thinking they confirm the Tetlock research on expert views (see various blogs over the past year or so).

Chatting to a friend recently whose political judgement and experience the blog respects we fell to discussing the prospects for Victoria’s Napthine Government. This was before recent opinion polls, without either of us being privy to recent internal party research and was really only a quiet reflection rather than a sustained analysis. We went over the impacts of the recent electoral re-distribution and the recent Federal budget but the general agreement was that either we didn’t know or we leaned towards the possibility of an ALP win.

Shortly after the chat the AC Neilsen poll indicated anything from a 56/44 to 59/41 split in favour of the ALP depending on how conservative the preference allocation assumptions were.  Premier Napthine’s inevitable comments about the only poll that counts is on election day and how much he was looking forward to that was an indication of how bad it was.

But what does the poll actually mean? Since Nate Silver’s triumph over the 2012 US Presidential election we have a much better idea. Silver (see his The Signal and the Noise or the blog seriatum) uses Bayesian methodologies to work out probabilities. For instance, in his book he uses data from polls and election results for US Senatorial elections from 1998 to 2008 to create a table from which you can estimate the probability of a candidate winning, based on size of lead in polling average and time before the election. Now there have to be some provisos – the US doesn’t have compulsory voting and the polling data is denser and more extensive than, difficult as it is believe, Australian data particularly when it comes to State elections.

Nevertheless, with the Victorian State election five months away you can extrapolate from Silver’s table, derived from his FiveThirtyEight models, that the probability of an ALP win is somewhere around 69% to 83%

This led the blog to do some more research on Australian research on how good polls are in predicting results. What is surprising is that there is little directly comparable Australian data, partly because of the lack of polling data density and the fact that the claims and counter-claims of the polling companies accuracy and successes are irrelevant and/or selective.

One of the most thorough pieces of work was done by Justin Wolfers of Stanford and Andrew Leigh, then at Harvard, which suggests Australian polls had no predictive power beyond a few months before an election See http://users.nber.org/~jwolfers/Papers/GSB1723-ThreeTools.pdf  Wolfers and Leigh also looked at two other predictive tools – betting markets and combinations of economic factors. A popular version of the latter for a while was the ‘misery index’ created by Arthur Okun and calculated by adding the inflation and unemployment rates. Wolfers and Leigh found economic indicators (not the misery index) as such were quite good indicators and the betting markets were not bad either.

There is also an Australian strategy and corporate advisory firm, Pottinger, which uses the same Bayesian techniques as Silver. They did an exercise on the 2013 election which was interesting. See http://www.pottinger.com/uploads/1/9/5/1/19512909/tr_poll_tracker_v4.pdf They got the result right – which was hardly surprising – but the techniques they used are an interesting insight into Bayesian probability theories as they apply to polls and politics.

Simon Jackman, a Stanford political science professor, (http://jackman.stanford.edu/) has also done work on polling and looking at tools to interpret them, encompassing things like ‘house effects’ and how particular polling organisation’s methodologies produce particular results. He is an excellent guide to how one should interpret opinion polls and, while he often says he is not talking about how the media interprets them, the sub text is fairly damning. If you visit his site he has a section on Bayes. He has also written a book, Bayesian Analysis for the Social Sciences, which among other things, guides the reader through some of the software which helps employ Bayesian analysis.

But the third tool Wolfers and Leigh use – the betting markets – are probably better than economic data, which is relatively sparse at the State level, or polls as a way of predicting State elections. Current betting on the Victorian election shows odds of Libs 2.75 to 1 and ALP 1.4 to one. The odds are not generous – even in a two horse race – but the blog’s SP bookmaking grandfather, who had never heard of Bayes, would probably be edging the ALP numbers even closer to even money on the latest data.

The other interesting thing about this is, of course, Andrew Leigh, who is now an ACT MHR and probably one of the best things to happen to the Federal ALP Parliamentary Party in recent years. He is a genuinely deep thinker and researcher with a gift for campaigning and a willingness to explore policies and issues beyond sound bites. His new book, The Economics of Just About Everything, is due out late July and should be as well worth reading as his book on inequality, Battlers and Billionaires. In a small way Leigh is following the Whitlam path of analysis and policy development as part of the plan to win.

And then there is internal polling. Now the blog is deeply suspicious of claims about internal polls and is generally not prepared to believe them unless they come with a certified copy of the results and the raw data on which they are based. But there are hints from here and there that internal polling shows that Tony Abbott and his budget are a significant part of the Victorian Liberals problem. If the claims are correct they will be confirmed by whether Abbott gets any closer to the Victorian election campaign than Darwin or some conveniently organised overseas jaunt.

And speaking of Gough and our current much-loved PM, when, sadly, Gough Whitlam leaves us will Abbott will be as churlish in his comments as he was when Margaret Whitlam died? It would not be unexpected from a man who elevates sound bites above policy analysis and evidence and the question will be: will he be able to muster the necessary grace? If my grandfather could miraculously re-appear from his betting shop in a Carlton back lane he might give you long odds on it.