This story from RTE provides welcome news that the suicide rate has fallen. It contains the following remark "Today's report shows a continued reduction in suicides each year since 2003. However, it says it is still too early to say whether this trend is significant." But why so hesitant:why is it too early to say? If they mean significance in the normal sense of the word, then the only thing that matters is the size of the effect, surely?

If they mean statistical significance, I am not sure that makes sense either: we are talking about changes in a population parameter here so statistical inference doesn't come into it? That is, if the rate falls then it falls, end of story. I suppose this is related to the hoary old question of what t statistics mean when your data is the population & not a sample.

Actually the news is even better: as the article notes "Given the population growth, the rate of suicide is now the lowest since 1993, when suicide was decriminalised".

http://www.rte.ie/news/2009/0902/suicide.html

## 8 comments:

Good news indeed.

Re t statistics here is a beautiful paper published in J Econ Persp on significance, t student and the art of brewing, in a word Guinnessometrics:

http://faculty.roosevelt.edu/Ziliak/doc/Guinnessometrics_Ziliak%20article%20JEP%202008.pdf

Mirko

Yes nice to have some good news for a change.

There are folks who have a big problem with t stats, McCloskey (who works with Ziliak) on this for example. I don't personally.

I presume they mean significant in the sense of whether the last few years have been "randomly" good in terms of reductions in numbers, and we can expect a return to previous levels in the future.

The alternative being that underlying factors have permanently reduced suicide rates

That may well be what they mean but isn't that statistical significance? I suppose you could estimate a time series model to see is it trending down. You could infer then whether innovations are likely to have long run effects. Maybe thats what they did 'though I doubt it.

Kevin, I share you're view that questioning statistical significance here is misguided.

The article clearly highlights both poorly considered application and interpretation of useful adn important statistical concepts.

I think it would be beneficial for all if we had a session on this. Or at least considered compiling the key readings in the area. Liam has already indicated a willingness to do something like this and i think Mark might also be interested.

One specific area that i'm particularily interested in is the Bonferroni 'correction' which seems to be en vogue at the moment.

Bonferroni is about 100 years old! Its too conservative & there are many better alternatives. I can suggest a few readings - I have looked at this a bit.

this is definitely one for a future discussion. kevin and wen's presentation was nice last year. bonferroni is usually thought of as too conservative as Kevin says. the bonferoni-holm is commonly used and there are a number of similar ones. The debate that i was having with pete rested on the conditions under which you need to use these corrections.

If the comparisons are independent (say on different sub-samples) its relatively easy. If not it gets tricky. Given a particular condition, "positive regression dependence", you can use certain methods.

For a start see:

FORMALIZED DATA SNOOPING BASED ON GENERALIZED ERROR RATES

Joseph P. Romano, Azeem M. Shaikh and Michael Wolf Econometric Theory (2008), 24 : 404-447

Post a Comment