In an idle moment this afternoon, I wrote a Stata command.

It was to create a light-weight implemention of the “percentogram” described at https://statmodeling.stat.columbia.edu/2023/04/13/the-percentogram-a-histogram-binned-by-percentages-of-the-cumulative-distribution-rather-than-using-fixed-bin-widths/, and I like the result, but it struck me that it is a good example of how practical and useful it can be to engage in Stata programming. Also, it’s a good example of how writing code in Stata (in a programmable command language) is very different from writing code in a stats-capable programming language like R, Python or Julia.

A very nice idea: search for evidence of COVID-19 RNA in municipal wastewater, as a cheap and fast form of public health surveillance. A pre-print shows that this works well, in a trial in Connecticut. I think the evidence is in their favour, but they commit two cardinal errors: first, they report a correlation (well, a squared correlation) between time-series and second, they do it on smoothed data. Autocorrelation means time-series may have vastly inflated and/or spurious correlations, and stripping the noise out of variables removes the noise from the comparison, making it seem, well, much less noisy than it is.

This is one of their key results: the smoothed RNA curve looks just like the smoothed hospital admissions curve, with a lead of about 3 days:

They report an R² of 0.99 for this relationship.

However, they also show the data. Given there are 2 series for 44 days, we can pick this off the graph without too much effort:

(This is prompted by @lycraolaoghaire’s tweets: https://twitter.com/lycraolaoghaire/status/1265251252239286272?s=20).

It turns out that the correlation between the RNA measurement and hospital admissions is 0.357 (R² = 0.13). If we lag by one day, the R² rises to a very respectable 0.45, but declines again if we lag by 2 (0.22) or 3 (0.22) days. In other words, there is a real signal here, but it is vastly overstated by R² = 0.99, and the lead it gives is not as big as claimed.

Predicting hospital admissions using lagged RNA values, with lags of 1 to 5, and then all five lags together (green line) looks like this:

This is a much less impressive graph than the original, but it is picking up something. Most of the work is done by the one-day lag, which has a clear effect, and the combined 5-lag model isn’t better (by LR-test) than the L1 model only. However, using this technique very widely as a passive surveillance technique is going to pick up unexpected large shifts in disease RNA, which is much more important than being able to predict moderate changes in hospitalisation from moderate changes in RNA presence in sewage sludge.

Screen-picked data available here, no warranties.

Mike Tomlinson has created a certain amount of controversy by asserting that Northern Ireland’s COVID-19 death rate is disproportionate with that of the Republic (see article). In particular, he notes that the per capita rate of deaths in hospital settings (which is all that is reported for NI) is higher than that for the Republic (which normally reports all deaths, but for which the hospital deaths figure is also available). For instance, yesterday’s data says the cumulative figure for hospital deaths in the RoI is 386, while for NI it was 250.

Scaling by the relative populations, that suggests an expected NI hospital death rate of 386 * 1.891 / 4.904 = 148.8. 250 is a lot more than 149, even allowing for some incomparability in how the stats are collected.
Continue reading COVID-19 deaths: NI and IRL compared →

For years I have taught students to read printed statistical tables: the Standard Normal Distribution, the t-Distribution, the chi-square Distribution. I want them to do certain tasks (e.g., construct a confidence interval) “by hand” a few times, rather than in Stata, so that they understand what it is doing. I also want them to be able to do it with no more than a calculator, in the final exam.

For the past few years I’ve been working with R-Shiny to develop web-apps, which allow exploration of a concept, self-learning exercises and self-marking assessments. I also use it increasingly in class to demonstrate ideas. I’ve been tempted to replace the paper distribution tables with online versions, but have been holding back because of the pen and paper exam.

Continue reading Shiny apps for distributions →