Interesting Links for 3-10-2011

[andrewducker]

• This is how they gerrymander homeless children
  (tags: uk homeless housing politics)


• Why the world envies Mr Osborne
  (tags: GeorgeOsborne finance politics uk)


• Rick Perry's hunting ranch is called....Niggerhead.
  (tags: usa RickPerry racism politics)


• Homemade GPS Receiver
  (tags: GPS Electronics)


• Judge rules against Bayesian probability
  (tags: statistics law uk)


• Why This Year's Doctor Who Finale Was (Mostly) Better than Last Year's
  (tags:)


• The Pinboard people talk about the fallout from how badly Delicious fucked up.
  (tags: avos bookmarking fanfiction del.icio.us pinboard taxonomy)


• Cameron wants to scrap the Human Rights Act but Clegg won't let him
  (tags: libdem Conservatives politics rights uk)


• Historians Politely Remind Nation To Check What's Happened In Past Before Making Any Big Decisions
  (tags: TheOnion satire funny history)


• USA now assassinating its own citizens without due process
  (tags: usa law assassination murder)


• How cat videos get onto the internet
  (tags: cats videos youtube funny comic)

Comments

[purplecthulhu.livejournal.com]

This all gets interesting when you're not in a position to have a parent population to draw your sample from. That's why Bayesianism has taken over cosmology - we only have the one universe.

I'm now working in a nest of Bayesians and what it says about priors and posteriors makes a lot of sense to me. But maybe that's the indoctrination :-)

[steer.livejournal.com]

Heh... yes, it puts frequentists in an awkward philosophical position but it's not actually uncommon. For example, in road transport work, due to the nature of the systems you really don't want to repeat an experiment (for example the "experiment" you are studying is a massive road closure -- you only get one). So philosophically a frequentist is more troubled by the notion of talking of the probability of an event which has definitely happened in a one off unrepeatable experiment... but the mathematics still works even if the philosophy is broken. I'm extremely uncomfortable with the notion of the prior distribution -- especially as the notion of an "uninformative prior" is completely inconsistent for many situations (anywhere where there's no symmetry in your outcomes).

In a legal situation (which is where this started) you might be feeding into your model a prior probability of guilt (or a prior probability of the blood sample being the defendants). Now obviously a competent statistician checks that the prior does not greatly affect the posterior. Nonetheless it could lead to some awkward situations. I agree that mathematically the whole situation makes sense -- indeed the mathematics is uncontroversial once the prior is asserted (OK, it can get rough and you need sampling but I'm 100% fine with that). However, I'm not at all comfortable with asserting an ex nihilo belief and then that being an important part of the model.

[purplecthulhu.livejournal.com]

Indeed, the prior is where I've had philosophical and technical issues with Bayesianism. But where there is good reason for a prior (a previous set of data for example), or where a variety of uniform priors don't affect the posterior, then I don't have a problem. They are in effect representing your ignorance of the system.

The prior, likelihood, posterior approach also explains why the same set of data can be taken as evidence for utterly different conclusions - eg. tea partyists and liberals. They're working from utterly different priors, which are there no matter what. Even if you're a frequentist.

[steer.livejournal.com]

is good reason for a prior (a previous set of data for example)

Where did that get its prior? :-)

here there is good reason for a prior (a previous set of data for example), or where a variety of uniform priors don't affect the posterior, then I don't have a problem

Absolutely -- in a symmetrical situation (the ball is under one of three cups) a uniform prior is perfect.

They're working from utterly different priors, which are there no matter what. Even if you're a frequentist.

Sure -- but having different outcomes from the same set of data (depending on prior) would not happen in frequentist analysis (and, a good Bayesian would say that the conclusions are not meaningful unless there was a reason to prefer one prior).

[purplecthulhu.livejournal.com]

You might be interested in this discussion on Prof Peter Coles' blog...

http://telescoper.wordpress.com/2011/10/06/bayes-in-the-dock/#entry

[steer.livejournal.com]

Interesting -- thanks for that.

My reading of the original redacted court document was that it was Bayesian statistics, not Bayes Theorem which the judge ruled against. The Guardian article muddies the water completely by using the two interchangably. The original court document is heavily redacted so it's unclear.

If the Prof you link to is right, it's a stupid ruling but I *think* he's working from the Guardian article not the court document. The rest of his column is interesting though.

[purplecthulhu.livejournal.com]

Do comment on Peter's blog if you wish - he likes to have new people visit!

[andrewducker]

Cheers, that was fascinating. And the conversation between you and Steer, likewise.