Video killed the radio stars

Francisco tells me that they have uploaded my talk (which I gave last week in ULPGC). I haven't seen it all, but the bit I did see is not too bad, I thought... Check it out!

 

Nice & weird people in the Canary island (oh: I went there for work too!)

The past one has been a very interesting week, which I've spent visiting the University of Las Palmas, in the Canary Island. Since it was the last week on maternity leave for Marta, we all went. I knew the weather would be good, but we didn't expect it to be so nice $-$ like proper summer weather!

We were really lucky to start with, as while we were driving to the airport last Saturday, I accidentally got the wrong turn before entering the motorway and by doing that, we were able to see a massive queue ahead. So we decided to ask our phone to drive us to the airport through the Surrey back roads. We got a bit scared because the navigator kept telling us to turn into small roads that looked nothing like going to the airport, but in the end we actually got there on time (unlike 15 people who actually missed the flight).

Of all the family, XY has enjoyed the week the most. He has spent most of the time eating, walking around and being chatted up by the nice, friendly people in the streets of Las Palmas. At first he was a bit puzzled by people speaking in Spanish (I guess it sounded close enough to Italian, but not quite Italian, so it freaked him out a bit). But then he got used to it and started waiving at people when they were stopping us in the streets to say that he was so "precioso", or "lindo", or "guapo".

Speaking of which, I found it amazing how we could relatively easily get by with our Spanitalian. In fact, I did study a bit Spanish, which was helpful, but for most things we could speak to people without resorting to hand-waiving or English (of course that's not ground-breaking news $-$ but it was nice).

Which brings me to the actual reason of the trip, which was a week of meeting with Miguel Negrín and Francisco Vázquez Polo. We mainly talked about some work in health economics, specifically related to the application of utility functions different from the standard net benefit (which is defined as $ke - c$, where $e,c$ are the variables of clinical effectiveness and cost, respectively, and $k$ is a willingness-to-pay parameter, used to rescale the benefits and put them on a monetary scale). In the beginning I wasn't quite convinced about their argument, but then we got to talk and, while I'm still not 100% sure of all the implications, I think this is an interesting problem and definitely worth investigating. So we'll try and work on this together, which is a good outcome of the visit $-$ and I'll post about this, once we (I) have clearer ideas/material. I also gave a seminar on the structural zero model, expanding on the talk I gave at GSK a few days back. I think this is really interesting and I seem to get useful comments, which I'll definitely use in my revision of the paper.

The last day we rented a car and decided to tour the island of Gran Canaria. We'd never been there before, so we didn't really know much, but Miguel and Francisco made a couple of suggestions. While we were walking around in Maspalomas, a weird Italian lady stopped us and started to blab. She looked interested in us, but in fact asked us how old we were 3 times in the space of 5 minutes and told us pretty much the same things about 20 times. As it turned out, she was working for a holiday resort (apparently they have taken some sand from Bahamas and brought it there before building the resort $-$ I am still annoyed with myself that I didn't turned around and left as she actually said that, as if it was some amazing thing which was supposed to make me want to divorce Marta and marry her instead). Anyway, she wanted us to get some sort of scratch'n'win ticket. We weren't really interested but she started scratching a couple and apparently we won either an iPad, a free one-week holiday (in one of their "amazing resorts"), or €500 in cash. To get the prize we "only" had to go and see the resort with the Bahamas beach, which we could only do by going with a cab that she had to call (and there was no other way, "because those are the rules"). We went back to the car and actually visited the nice little town of Teror (not very big and not very much to do, but still quite nice and much less hot!).

You stole my idea!

Earlier today, Gareth has showed me a recent, interesting paper by Michael Sweeting (and colleagues). In the paper, Micheal et al describe their work on a R package to extend on the framework of the Continual Reassessment Method (the original paper by John O'Quigley and colleagues is here), which is a particular design that can be used in dose-finding studies.

No one has really stolen anything (just to be clear!) but I did have the same idea just a couple of months ago $-$ Gareth was working on a similar problem and asked me to have a chat about how one can apply the CRM, which is essentially a Bayesian method.

In its standard form, essentially, the CRM uses a very simple model based on very simple priors (typically exponentials or Gamma). While we were discussing it, I thought it would be nice to expand this and may be make a package that would link to JAGS or BUGS and allow you to select the prior from a wider range. Which, as it turns out, is just what Michael et al have done! 

Well $-$ at least I can see that was a good idea!

My talk @ GSK

This Thursday I'll give a talk at the GSK Statistics Forum. Erika (with whom I shared a train journey to the 2012 BayesPharma and a group walk in Oxfordshire a few years back) now works at GSK and invited me. I will talk about the model for cost-effectiveness with structural zero costs, on which I am working at the moment.

In fact, I managed to get hold of some new and more interesting data, which I will use to produce a better, more articulated example (although I think for the talk, I'll stick with the original acupuncture example, which is part of the working paper). 

I'll post my slides later, in case anyone's interested...

Parum PI

A couple of weeks ago, the MRC-funded research project on the Regression Discontinuity Design (of which I'm the Principal Investigator) has officially started, so I thought I wrote a few lines of update about it, after the couple of posts (here and here) referring to presentations we've given on (very preliminary) work we've done.

Unfortunately, I don't think I can quite claim to have the physique du role to be a Magnum PI [hence, and to show off that I did study Latin in school $-$ although you may argue that I could have easily used Google Translate... but I haven't: promised! $-$ the title of the post], if only for the fact that we're having a terrible September, weather-wise, here in London...

Let me be clear that I'm not complaining and of course I am very happy that we got the grant. But I must say that being PI is at the same time a very exciting and exhausting role (I'll pretend that it hasn't occurred to me that the project is just started). Today I put on my most Hawaiian shirt and spent most of the time trying to sort out a few admin things and sending emails. 

Hopefully, we'll shortly have something a bit more substantial to report about $-$ the signs are all there, luckily...

Attendance in parliament (in Italy)

Earlier today, I found some interesting data on Italian MPs voting records in the current parliament (which was opened last April). The data are available for both Houses (the Italian system has a lower house, called Camera $-$ effectively the same as the House of Commons in the UK $-$ and an upper house, the Senate). The Camera is made by 630 MPs, while there are "only" 321 senators (some of whom are appointed for life by the President of the Republic). For each MP, data are recorded on their political affiliation, constituency, number of votes attended and total number of votes in the current parliament. I thought I tried some relatively simple model to see whether there are some substantial differences in terms of party or constituency. 

I think the two houses can be modelled separately, at least as a first approximation. So, for example, say I concentrate on the lower House and define $y_i \sim \mbox{Binomial}(\pi_i,n_i)$ where for $i=1,\ldots,630$, $y_i$ is the number of recorded attended votes, $n_i$ is the total number of votes and $\pi_i$ is the individual probability of attendance. [NB: not all MPs are "exposed" to the same number of votes, as some have been appointed after by-elections, ie after parliament was opened last April. Also, the system does not record the reason for absence and so some of them may be legitimate (eg illness, or due to other institutional engagements). Nevertheless, the data give some good indication of the actual activity of the MPs].

I used a very simple model to define the attendance probability 
$$ \mbox{logit}(\pi_i) = \alpha + \beta_{p_i} + \gamma_{c_i}$$where the index $p_i$ indicates the party to which the $i-$th MP is affiliated and the index $c_i$ indicate their constituency. Of course, parties and constituencies will be replicated within the dataset (because MPs cluster within them). Also, other individual- or party- or constituency-specific variables may be relevant to explain away the different attitudes to participating to the parliamentary work $-$ but I'll keep the model simple, mainly because I'm a bit lazy and am not spending time to find these other variables!

The model is completed specifying vague prior for the overall average attendance rate (on the logit scale), $\alpha$ and structured priors for $\beta_j \sim \mbox{Normal}(0,\tau_p)$ and $\gamma_k \sim \mbox{Normal}(0,\tau_c)$, with $\sigma_p = \frac{1}{\sqrt{\tau_p}} \sim \mbox{Uniform}(0,10)$ and $\sigma_c = \frac{1}{\sqrt{\tau_c}} \sim \mbox{Uniform}(0,10)$. 

With this specification, the coefficients $\beta_j$ define the incremental effect of the $j=1,\ldots,N_p$ parties and $\gamma_k$ is the incremental effect of the $k=1,\ldots,N_c$ constituencies on the propensity of each MP to attend the votes. Negative values for the coefficients means that a given party/constituency decrease the chance of attendance.

I ran this model and here're the results, in the form of coefficient plots, reporting the posterior interval estimation for the effects $\beta$ and $\gamma$. For the Camera, the MPs affiliated with the "Movimento 5 Stelle" party and those with "Sinistra Ecologia Libertà" show positive propensity to participate in votes, while FdI ("Fratelli d'Italia") and PdL ("Popolo delle Libertà", Berlusconi's party) are substantially associated with negative propensity to attend.

Constituency also seem to have some differential effect; MPs from Valle d'Aosta, Liguria, Sardinia, Friuli Venezia Giulia, some provinces in Lombardia (but not including that of Milan) and some provinces in Sicily (including Palermo) have positive propensity to attend the votes.

In the senate, the situation is kind-of-different, if only for the fact that by the nature of the Italian system, some parties are not even the same. All the "positive" party effects disappear and no party is significantly associated with higher propensity of attendance. However, "PdL" and "Scelta Civica" (the party of former Prime Minister, Mario Monti) do show substantially negative values, indicating their senators have a lower propensity to show up at parliamentary votes.

BCEA in UseR!

In a recent post, I had hinted at big news for BCEA $-$ I thought it was pretty much a done deal, but because it wasn't yet set in stone, I didn't want to jinx it...

But now I've sorted all the details with Springer, who have asked me to write a book on the R package (which I originally wrote to accompany BMHE) and so it's official: BCEA is going to feature in the Use R! series!

We ("we" being Andrea and myself, who will co-author the book) are very excited about this. We are still working on the next release of the package, which will include the code to run the multiparameter analysis of the expected value of partial information using the algorithm developed by Strong & Oakley and based on non-parametric regression.

But at the same time, Andrea and I will need to crack on the actual write up. The tentative table of contents is this:

1. Bayesian analysis in health economics.
1. Very brief introduction to the Bayesian approach, with particular reference to MCMC computations.
2. Basics of health economic evaluation, specifically under a Bayesian approach.
3. Probabilistic sensitivity analysis through simulations.
2. BCEA – worked examples to describe all the functions
1. Basic analysis
1. Cost-effectiveness plane
2. Expected incremental benefit
3. Contour plots 
4. Summary tables
2. Probabilistic sensitivity analysis
1. Cost-effectiveness acceptability curve
2. Expected value of information 
3. Expected value of partial information (2 stage MCMC and approximation methods)
3. Advanced methods
1. Mixed strategy
2. Including risk-aversion in the utility functions
3. Comparison of multiple interventions
3. Graphical options in BCEA
1. Brief introduction to ggplot2 and its use in BCEA
2. Differences between the base and ggplot options in BCEA
4. Conclusions

The idea is to have lots of worked examples to show how to do the analysis using BCEA. We have some already, but we'll find new ones as well.

Biostatistics seminar

As part of the activities of the UCL Biostatistics Network, we organise regular seminars, to which we invite (usually relatively local $-$ for budget reasons only!) speakers. 

September is the start of the new term, and we'll kick off with what (in my opinion) is a very interesting topic. Alexina Mason (Imperial College) will speak about full Bayesian methods to deal with missing data. The seminar will be on September 18th at 4pm in the department of Statistical Science. 

Title: A general strategy for dealing with missing data using Bayesian methods
Abstract: Bayesian full probability modelling provides a flexible approach for analysing data with missing values, and offers an alternative to standard multiple imputation.  Plausible models allowing for missing responses and/or missing covariates can be built, which incorporate realistic assumptions about the missingness mechanism.  Additionally, the Bayesian approach lends itself naturally to sensitivity analysis, which is crucial when the missingness mechanism is unknown.  These strengths will be demonstrated by presenting a general strategy for a "statistically principled" investigation of data which contain missing values. The first part of this strategy entails constructing a "base model" by selecting an analysis model, then adding a sub-model to impute the missing covariates followed by a sub-model to allow informative missingness in the response.  The second part involves running a series of sensitivity analyses to check the robustness of the conclusions.  An antidepressant trial comparing the effects of three treatments will be used as an illustrative example throughout.  In particular, we will focus on missing responses assuming a non-ignorable missingness mechanism.