Bayesians, puzzle me this:

If you are trying to learn the p parameter of a binomial distribution, the least informative prior is the uniform distribution, i.e. Beta(1,1)

If you are fitting the same data using a logistic regression, setting your prior to be a logistic distribution with s=1 (the same scale parameter as your generative model) is equivalent to assuming a uniform distribution over the p parameters of the binomial distribution.

But that prior is more restrictive than a logistic distribution with a larger scale parameter, and thus we could make it less informative by increasing the scale. Back in the world of binomial distributions, that would be like a Beta prior with the probability mass pushed up against the edges.

How can informativeness/restrictiveness of the prior be dependent on how you parameterize the hypothesis space?

A couple days ago, as I was wandering the interwebz, I ran across the Academic Productivity blog, and from there, this article about increasing research productivity. It’s intended audience appears to be administrators at research institutions, but the advice is fairly applicable to individual researchers too.

First on the list is the advice of writing 15 to 30 minutes every day. I had heard advice before about writing on a regular schedule, but the advice I’d seen before was more like to make 2 hour blocks of time. Personally, I can’t imagine making much progress in 15 minute blocks, since it seems to me like it takes 15 minutes just to remember where I was in whatever I’m writing, but I’m going to try it.

Some of the time, what I’m writing will also appear here, as “leaving your door open” is also supposed to be a good practice (though not while you’re trying to write).

Oh, and I finally figured out all the details of migrating this blog off of godaddy to 5gigs hosting!

This is a summary of my research project, written for a class.

The tone sandhi systems of Wu Chinese, like those of many Min and Gan dialects, offer rich ground for understanding the phonology of tone. With typically six to eight citation tones and complex sandhi patterns, each language offers a challenge to analysis, and there is a high degree of geographic and generational variation. Significant generational variation has been recorded for at least a century, and ultimately the great regional variation must have arisen from generational variation, but with China’s economic development, the regional differences have experienced leveling while the generational differences have likely increased. In Jinhua, a city in the southern Wu zone, Mandarin is replacing Wu even in the home and marketplace, so that few people born after 1990 speak it all, and those born in the previous decade have limited vocabularies and strong Mandarin influence in their Wu. In view of these things, it is important to document the language as well as interesting to examine the phonology and generational differences.

Previous phonological documentation of the dialects in the Jinhua area consist of one large dictionary, one dialectology survey of the region that included Jinhua and one village nearby, and one acoustic study of citation tones. For each of these studies, the tone data comes from just one informant, and the two sets of sandhi data differ considerably. Using methods similar to other acoustic sandhi studies, I recorded six speakers born in the 1980s and four speakers born before 1950, who all spoke the dialect of urban Jinhua, and five speakers (three young, two older) from a village about 10 km outside the city. I expect to compile average sandhi contours for each group and compare the inter-group variation. I also intend to provide a phonological analysis of the sandhi patterns and try to identify the causes of intra-group variation.