Simon N Wood
Chair of Computational Statistics, School of Mathematics, James Clerk Maxwell Building,
University of Edinburgh, EH9 3FD
simon.wood
ed.ac.uk
Note: since July email to my old bristol address has not been delivered to me, and generated no warning of this.
I work as a professor of statistics. Statistics is about the honest interpretation of data. It's not always a popular subject: honest interpretation of data is difficult, and much less appealing than less honest interpretation. Disraeli is often supposed to have said that there are lies, damned lies and statistics. Gladstone was president of the Royal Statistical Society. I tend to think both were right.
Round figure economic deprivation estimates based on Marmot review (e.g. figures 2.3 and 2.5). The figure for post 2008 increase is right at the lower end of what is implied by Marmot. Round figure Covid estimates based on ONS lifetable and covid death by age data, a 0.6% IFR, 80% herd immunity fraction and 1 life year co-morbidity adjustment, and the assumption of minimal mitigation only to avoid complete health system collapse: since the latter assumes less mitigation than anyone is proposing the figure is an upper bound. See this pre-print for further references.
Here is one reason why the above could be worrying: a comparison with the economic shock of the 2008 crisis, based on ONS data...
So, in words, we require the current economic shock, which is much larger than 2008, to result in much smaller life loss than was associated with 2008. Otherwise we will lose more life to the economic effects of Covid-19 suppression efforts than were ever likely to have been lost to Covid-19 itself. Of course the consequences of the 2008 crisis were amplified by the policies adopted thereafter, and perhaps those consequences could have been substantially alleviated by a more enlightened approach. But the historical record from the UK does not suggest a willingness to vote for such an approach, even if any sort of credible plan for avoiding the economic life loss were actually to be proposed. The 1945 election was perhaps the exception, but it's unclear that several months stuck at home on your sofa really leads to the same sort of cathartic re-evaluation of life's priorities as storming the beaches of Normandy.
Given that COVID-19 presents a rather low risk to most people in economically active age groups, and is generally a higher risk in retired age groups, it is questionable that one-size-fits-all measures requiring substantial economic contraction are a sensible policy response. Targetted measures based on people's vulnerability would give a better chance of saving the old from SARS-CoV-2 and the young from the very large health effects likely to be caused by historically large economic shocks: that is of minimising the covid crisis related life loss from all causes, in the long run.
From the data on life loss associated with economic hardship and economic shocks in the past, it is hard to see that a more balanced targetted approach would not save substantially more life than lockdowns (nuking your house will put an end to your rat infestation, no question, but a more nuanced approach might cause you fewer long term problems). If it is 'impossible' to differentially protect the old, in particular, then the age specific infection rates reported in the ONS infection survey data take some explaining (this is not cherry picking: the ONS data are the main source of direct measurement of infection). Similarly a scientific case that very stringent restrictions are the only way forward does need to also explain the Swedish data, rather than treating Sweden as some sort of discardable outlier (it needs a bit more than a lazy argument about population density, given the densities in Swedish cities).
It also seems that we are willing to pay much more to save a life from COVID-19 than from other causes. The usual NICE affordability threshold is around £30,000 per life year saved. OBR projected peak extra borrowing of £660B suggests at least 7 times that for COVID-19 (
here is an atempt to discuss this issue in more detail). Finally, there has been quite alot of slightly odd comparison with previous pandemics in the media. Here is a visualization comparing the severity of three pandemics.
I can't comment on the newly emerging concerns about `long covid' due to a lack of data on this. However it seems clear that these concerns should be weighed carefully against the effects of economic deprivation, insecurity and unemployment on health. There are data on these effects: they are very substantial.
Something that the Covid-19 crisis has emphasised is the fact that statisticians have not managed to adequately communicate how fundamental random sampling is to proper measurement of things like infection rates. That data somehow related to the thing we want to measure are not the same as data that actually measure it. Here is an attempt to explain random sampling and why it means that you should trust plots like the ones from the ONS Covid survey in preference to plots of cases, like the first one here.
A lockdown reading list
- Collapse (Jared Diamond) on how societies are destroyed, not by external forces, but by their failure to adapt their cultural norms to those forces.
- Thinking Fast and Slow (Daniel Kahneman) on the pitfalls of our intuitive reasoning, especially about risk and uncertainty.
- Mistakes were made, but not by me (Carol Tarvis and Elliot Aronson) on the psychology of sticking with bad decisions.
- The Parable of the Old Man and the Young by Wilfred Owen, on consequences of the above.
- Economics The User's Guide (Ha-Joon Chang) on what you really need to know about economics, and how it is not mainly about money, nor just a scaled up version of household accounting.
- The Great Crash 1929 (Galbraith) a delightful disection of economic hubris (and the need for stabilizing controls that we long since did away with).
- The Rise and Fall of the Third Reich (William Shirer) detailing exactly how things went wrong in Germany after the Great Depression.
- Witch hunting in Scotland (Brian Levack) on the Scottish experience of the great European witchcraft panic (James I/VI wrote a treatise on Witchcraft).
- Wood and Thomas paper on the problems of prediction with disease models in the absense of direct validation data (the least impressive item here).
(from March 2020)
Notes on debugging R code and C code from R
Apologies if I have not replied to you on an mgcv related query: I've got
rather behind on mgcv email, especially on the interesting stuff that
requires thought.
Boring career stuff about me: state comprehensive school educated; BSc in Physics from Manchester; PhD on biological modelling (Dept Applied Physics Strathclyde); short stint (1989-90) as a civil service bioeconomic modeller MAFF; 4 years postdoc at Imperial (biology) on biological dynamic models. Lecturer - Reader in statistical ecology, St Andrews (maths), also RSS graduate diploma in statistics; Reader - Prof in statistics (Glasgow, Bath, Bristol).
Books
Core Statistics (2015) is a short textbook in the CUP IMS textbook series. The idea is to offer a concise coverage of the essentials that anyone starting a statistics PhD ought to know, in the form of a brief introduction to statistics for the numerate. A pdf version is here (A5 format - ok for e-reading). Try this version for less wasteful printing on A4. Comments (including typo and error reports) very welcome. Here is the errata list and the algae and urchin datasets. (e.g. alg <- read.table("http://www.maths.bris.ac.uk/~sw15190/data/algae.txt") to read directly into R.). If you find the free download useful please consider buying the book (click top right to change location).
Generalized Additive Models: An Introduction with R (2nd ed) (2017) provides an introduction to linear (mixed) models, generalized linear (mixed) models, generalized additive models and their mixed model extensions. The second edition has a completely revised structure, with greater emphasis on mixed models and the equivalence of smooths and Gaussian random fields. A greatly enhanced range of smoothers is covered, along side a thorough upgrading of the chapter on GAM theory, and many new examples including functional data analysis, survival analysis, location-scale modelling and more.
I work as a professor in the
statistics group at the university of Edinburgh. I think Brexit was probably a rather poor decision. I kind of liked being in a block that produces roughly 100% of its own food, as opposed to an island that produces roughly 60%, and it seemed like a sensible thing to be part of a group large enough to stand up to China and the USA. But then I also tend to think of the British Empire as a rather racist enterprise responsible for the death of millions and the subjugation of far more, rather than, say, an indication of the country's former greatness. I'm currently joint editor of JRSSB and have two main research interests.
- Smoothing. In particular methods
for generalized additive modelling and applications of generalized
additive models (GAMs). I am especially interested in smoothness
selection, and low rank spline smoothing, and have written an R package called
mgcv
which implements GAMs. Some recent example smoothing papers are
- Wood, SN, N Pya and B Saefken (2016) Smoothing parameter and model selection for general smooth models (with discussion). Journal of the American Statistical Association.
- Wood, SN, Z Li, G Shaddick and NH Augustin (2017) Generalized additive models for gigadata: modelling the UK black smoke network daily data Journal of the American Statistical Association. Here are the data set black_smoke.RData and its description.
- Wood, SN and M Fasiolo (2017) A Generalized Fellner-Schall Method for Smoothing Parameter Optimization with Application to Tweedie Location, Scale and
Shape Models Biometrics.
- Wood, SN (2019) Simplified Integrated Nested Laplace Approximation in press Biometrika.
- Wood, SN (2020) Inference and computation with generalized additive
models and their extensions (with discussion from Greven, Scheipl, Kneib and Eilers) TEST
- Statistical Ecology. In particular using ecological dynamic
models as
statistical models to help understand ecological mechanisms, and
ecological applications of nonlinear random effects models and smooth
models, as part of NCSE. Some recent
example statistical ecology papers are
Fuller lists of papers are at
researcherid and
google scholar .
Here is a 2014 BIRS talk on inference for ecological dynamic models.
I am interested in taking on PhD students working on any area related to my research interests. Here are a couple of example projects: GAMs for big data and GAMs for multivariate data. The department has funding for strong students.
Advisees:
- Emiko Dupont, works on spatial modelling.
- Bertrand Nortier, works on quantile GAM methodology.
Here is a selection of talks. It's not exhaustive, but hopefully gives
some idea of what I work on.
This year I'm teaching Theory of Inference. Here are a couple of examples of previous courses.