free range statistics

Persian Monarchs described by P. G. Wodehouse in one of his funniest novels is an extremely simple fictional card game, but the gambling makes it a game of skill, and we can even construct plausible different strategies for winning. A good strategy involving card-counting beats a non-counting alternative by about 4% and random wagering by 36%.

It turns out that about 10 billion people were born in total in the twentieth century.

I have a play with counting how often particular digits turn up in numbers, starting with page numbers of a book based on a training exercise, and moving on to the so-called Benford's law or first-digit law.

A more systematic comparison of different ways of dealing with cells in a cross tab that have been supressed for confidentiality. For the particular model tested here, the best thing to do is the simple method of replacing all suppressed cells with 5; this works even better than using the original unsuppressed data which is very unstable when many cell counts are near zero.

I experiment with some different ways of handling counts in tables that have been suppressed for confidentiality, and come up in favour of multiple imputation. The mice R package helpfully lets you define your own imputation algorithm.

I play around with simulating some dice games.

Cross-sectional country-level data will show a relationship between income inequality and life expectancy even if inequality itself has no direct impact on life exectancy; so long as there is changing marginal impact of individual income on individual life space (as of course there is).

Sri Lanka has a rapidly growing tourism industry, two international tourism seasons, and seasonality patterns in arrivals that vary according to country of origin.

Rents in Melbourne have on average grown fastest in suburbs that were the cheapest in 2000; at least for two and three bedroom flats and for two bedroom houses. Also, scatterplots are awesome.

Simulating complex survey data in order to fit slightly mis-specified relative risk models, we find that confidence intervals' coverage is pretty much as advertised if we use appropriate methods that adjust for the complex survey data, but under-perform if the data is treated naively as coming from a simple random sample.