I thought it would be fun to fit a simple model in Stan to estimate the abilities of the teams in the World Cup, then I could post everything here on the blog, the whole story of the analysis from beginning to end, showing the results of spending a couple hours on a data analysis.

It didn’t work so well, but I guess that’s part of the story too.

All the data and code are here.

**Act 1: The model, the data, and the fit**

My idea was as follows: I’d fit a simple linear item response model, using the score differentials as data (ignoring the shoot-outs). I also have this feeling that when the game is not close the extra goals don’t provide as much information so I’ll fit the model on the square-root scale.

The model is simple: if team i and team j are playing and score y_i and y_j goals, respectively, then the data point for this game is y_ij = sign(y_i-y_j)*sqrt|y_i-y_j|, and the data model is:

y_ij ~ normal(a_i-a_j, sigma_y),

where a_i and a_j are the ability parameters (as they say in psychometrics) for the two teams and sigma_y is a scale parameter estimated from the data. But then before fitting the model I was thinking of occasional outliers such as that Germany-Brazil match so I decided that a t model could make more sense:

y_ij ~ t_df (a_i-a_j, sigma_y),

setting the degrees of freedom to df=7 which has been occasionally recommended as a robust alternative to the normal.

(Bob Carpenter hates this sort of notation because I’m using underbar (_) in two different ways: For y_i, y_j, y_ij, a_i, and a_j, the subscripts represent indexes (in this case, integers between 1 and 32). But for sigma_y, the subscript “y” is a name indicating that this is a scale parameter for the random variable y. Really for complete consistency we should use the notation sigma_”y” to clarify that this subscript is the name, not the value, of a variable.)

[Ed. (Bob again): How about y[i] and y[i,j] for the indexing? Then the notation’s unambiguous. I found the subscripts in Gelman and Hill’s regression book very confusing.]

Getting back to the model . . . There weren’t so many World Cup games so I augmented the dataset by partially pooling the ability parameters toward a prior estimate. I used the well-publicized estimates given by Nate Silver from last month, based on something called the Soccer Power Index. Nate never says exactly how this index is calculated so I can’t go back to the original data, but in any case I’m just trying to do something basic here so I’ll just use the ranking of the 32 teams that he provides in his post, with Brazil at the top (getting a “prior score” of 32) and Australia at the bottom (with a “prior score” of 1). For simplicity in interpretation of the parameters in the model I rescale the prior score to have mean 0 and standard deviation 1/2.

The model for the abilities is then simply,

a_i ~ N(mu + b*prior_score_i, sigma_a).

(I’m using the Stan notation where the second argument to the normal distribution is the scale, not the squared scale. In BDA we use the squared scale as is traditional in statistics.)

Actually, though, all we care about are the relative, not the absolute, team abilities, so we can just set mu=0, so that the model is,

a_i ~ N(b*prior_score_i, sigma_a).

At this point we should probably add weakly informative priors for b, sigma_a, and sigma_y, but I didn’t bother. I can always go back and add these to stabilize the inferences, but 32 teams should be enough to estimate these parameters so I don’t think it will be necessary.

And now I fit the model in Stan, which isn’t hard at all. Here’s the Stan program worldcup.stan:

data { int nteams; int ngames; vector[nteams] prior_score; int team1[ngames]; int team2[ngames]; vector[ngames] score1; vector[ngames] score2; real df; } transformed data { vector[ngames] dif; vector[ngames] sqrt_dif; dif <- score1 - score2; for (i in 1:ngames) sqrt_dif[i] <- (step(dif[i]) - .5)*sqrt(fabs(dif[i])); } parameters { real b; realsigma_a; real sigma_y; vector[nteams] a; } model { a ~ normal(b*prior_score, sigma_a); for (i in 1:ngames) sqrt_dif[i] ~ student_t(df, a[team1[i]]-a[team2[i]], sigma_y); }

(Sorry all the indentation got lost. I’d really like to display my code and computer output directly but for some reason the *code* tag in html compresses whitespace. So annoying.)

[Ed. (aka, Bob): use the "pre" tag rather than the "code" tag. I just fixed it for this blog post.]

I didn’t get the model working right away—I had a few typos and conceptual errors (mostly dealing with the signed square root), which I discovered via running the model, putting in print statements, checking results, etc. The above is what I ended up with. The whole process of writing the model and fixing it up took 20 minutes, maybe?

And here’s the R script:

library("rstan") teams <- as.vector(unlist(read.table("soccerpowerindex.txt", header=FALSE))) nteams <- length(teams) prior_score <- rev(1:nteams) prior_score <- (prior_score - mean(prior_score))/(2*sd(prior_score)) data2012 <- read.table ("worldcup2012.txt", header=FALSE) ngames <- nrow (data2012) team1 <- match (as.vector(data2012[[1]]), teams) score1 <- as.vector(data2012[[2]]) team2 <- match (as.vector(data2012[[3]]), teams) score2 <- as.vector(data2012[[4]]) df <- 7 data <- c("nteams","ngames","team1","score1","team2","score2","prior_score","df") fit <- stan_run("worldcup.stan", data=data, chains=4, iter=2000) print(fit)

(I'm using stan_run() which is a convenient function from Sebastian that saves the compiled model and checks for it so I don't need to recompile each time. We're planning to incorporate this, and much more along these lines, into rstan 3.)

This model fits ok and gives reasonable estimates:

Inference for Stan model: worldcup. 4 chains, each with iter=2000; warmup=1000; thin=1; post-warmup draws per chain=1000, total post-warmup draws=4000. mean se_mean sd 2.5% 25% 50% 75% 97.5% n_eff Rhat b 0.46 0.00 0.10 0.26 0.40 0.46 0.53 0.66 1292 1 sigma_a 0.15 0.00 0.06 0.05 0.10 0.14 0.19 0.29 267 1 sigma_y 0.42 0.00 0.05 0.33 0.38 0.41 0.45 0.53 1188 1 a[1] 0.35 0.00 0.13 0.08 0.27 0.36 0.44 0.59 4000 1 a[2] 0.39 0.00 0.12 0.16 0.32 0.39 0.47 0.66 4000 1 a[3] 0.44 0.00 0.14 0.19 0.34 0.43 0.53 0.74 1101 1 a[4] 0.19 0.00 0.17 -0.19 0.10 0.22 0.31 0.47 1146 1 a[5] 0.29 0.00 0.14 0.02 0.20 0.29 0.37 0.56 4000 1 a[6] 0.30 0.00 0.13 0.06 0.22 0.29 0.37 0.56 4000 1 a[7] 0.33 0.00 0.13 0.09 0.24 0.32 0.41 0.62 1327 1 a[8] 0.16 0.00 0.14 -0.16 0.08 0.17 0.25 0.41 4000 1 a[9] 0.06 0.00 0.15 -0.29 -0.03 0.08 0.16 0.32 1001 1 a[10] 0.20 0.00 0.12 -0.02 0.12 0.19 0.27 0.46 4000 1 a[11] 0.27 0.01 0.14 0.04 0.17 0.25 0.36 0.58 746 1 a[12] 0.06 0.00 0.14 -0.23 -0.01 0.07 0.14 0.33 4000 1 a[13] 0.06 0.00 0.13 -0.21 -0.02 0.06 0.14 0.31 4000 1 a[14] 0.03 0.00 0.13 -0.26 -0.05 0.04 0.11 0.29 4000 1 a[15] -0.02 0.00 0.14 -0.31 -0.09 -0.01 0.07 0.24 4000 1 a[16] -0.06 0.00 0.14 -0.36 -0.14 -0.05 0.03 0.18 4000 1 a[17] -0.05 0.00 0.13 -0.33 -0.13 -0.04 0.03 0.21 4000 1 a[18] 0.00 0.00 0.13 -0.25 -0.08 -0.01 0.07 0.27 4000 1 a[19] -0.04 0.00 0.13 -0.28 -0.11 -0.04 0.04 0.22 4000 1 a[20] 0.00 0.00 0.13 -0.24 -0.09 -0.02 0.08 0.29 1431 1 a[21] -0.14 0.00 0.14 -0.43 -0.22 -0.14 -0.06 0.14 4000 1 a[22] -0.13 0.00 0.13 -0.37 -0.21 -0.13 -0.05 0.15 4000 1 a[23] -0.17 0.00 0.13 -0.45 -0.25 -0.17 -0.09 0.09 4000 1 a[24] -0.16 0.00 0.13 -0.40 -0.24 -0.16 -0.08 0.12 4000 1 a[25] -0.26 0.00 0.14 -0.56 -0.34 -0.25 -0.17 0.01 4000 1 a[26] -0.06 0.01 0.16 -0.31 -0.17 -0.08 0.05 0.28 658 1 a[27] -0.30 0.00 0.14 -0.59 -0.38 -0.29 -0.21 -0.03 4000 1 a[28] -0.39 0.00 0.15 -0.72 -0.48 -0.38 -0.29 -0.12 1294 1 a[29] -0.30 0.00 0.14 -0.57 -0.39 -0.31 -0.22 -0.02 4000 1 a[30] -0.41 0.00 0.14 -0.72 -0.50 -0.40 -0.32 -0.16 1641 1 a[31] -0.25 0.00 0.15 -0.50 -0.35 -0.27 -0.15 0.08 937 1 a[32] -0.40 0.00 0.14 -0.69 -0.48 -0.39 -0.31 -0.13 4000 1 lp__ 64.42 0.86 12.06 44.83 56.13 62.52 71.09 93.28 196 1

Really, 2000 iterations are overkill. I have to get out of the habit of running so long straight out of the box. In this case it doesn't matter---it took about 3 seconds to do all 4 chains---but in general it makes sense to start with a smaller number such as 100 which is still long enough to reveal major problems.

In any case, with hierarchical models it's better general practice to use the Matt trick so I recode the model as worldcup_matt.stan, which looks just like the model above except for the following:

parameters { real b; realsigma_a; real sigma_y; vector[nteams] eta_a; } transformed parameters { vector[nteams] a; a <- b*prior_score + sigma_a*eta_a; } model { eta_a ~ normal(0,1); for (i in 1:ngames) sqrt_dif[i] ~ student_t(df, a[team1[i]]-a[team2[i]], sigma_y); }

We fit this in R:

fit <- stan_run("worldcup_matt.stan", data=data, chains=4, iter=100) print(fit)

And here's what we get:

Inference for Stan model: worldcup_matt. 4 chains, each with iter=100; warmup=50; thin=1; post-warmup draws per chain=50, total post-warmup draws=200. mean se_mean sd 2.5% 25% 50% 75% 97.5% n_eff Rhat b 0.46 0.01 0.10 0.27 0.39 0.45 0.53 0.65 200 1.01 sigma_a 0.13 0.01 0.07 0.01 0.08 0.13 0.18 0.28 55 1.05 sigma_y 0.43 0.00 0.05 0.35 0.38 0.42 0.46 0.54 136 0.99 eta_a[1] -0.24 0.07 0.92 -1.90 -0.89 -0.23 0.35 1.81 200 0.99 eta_a[2] 0.10 0.06 0.83 -1.48 -0.32 0.17 0.65 1.52 200 1.00 eta_a[3] 0.55 0.06 0.82 -0.99 0.00 0.55 1.08 2.06 200 1.00 eta_a[4] -0.54 0.08 1.10 -2.36 -1.33 -0.58 0.18 1.61 200 1.01 eta_a[5] 0.05 0.07 0.99 -1.64 -0.68 0.05 0.70 1.80 200 0.99 eta_a[6] 0.18 0.06 0.79 -1.36 -0.32 0.17 0.75 1.51 200 0.99 eta_a[7] 0.52 0.07 0.94 -1.30 -0.05 0.62 1.10 2.25 200 0.98 eta_a[8] -0.27 0.06 0.89 -1.84 -0.95 -0.26 0.36 1.45 200 0.98 eta_a[9] -0.72 0.06 0.79 -2.25 -1.28 -0.74 -0.17 0.82 200 0.99 eta_a[10] 0.18 0.06 0.86 -1.60 -0.38 0.21 0.81 1.78 200 0.99 eta_a[11] 0.71 0.06 0.90 -0.98 0.08 0.79 1.29 2.54 200 1.00 eta_a[12] -0.39 0.06 0.89 -2.16 -0.97 -0.36 0.19 1.27 200 0.99 eta_a[13] -0.18 0.06 0.90 -1.77 -0.78 -0.18 0.44 1.55 200 1.00 eta_a[14] -0.25 0.06 0.90 -1.86 -0.87 -0.28 0.32 1.52 200 0.99 eta_a[15] -0.33 0.07 0.99 -2.27 -1.06 -0.33 0.42 1.58 200 0.99 eta_a[16] -0.43 0.06 0.83 -1.78 -1.06 -0.50 0.22 1.13 200 1.00 eta_a[17] -0.16 0.07 0.94 -1.90 -0.84 -0.20 0.51 1.65 200 0.98 eta_a[18] 0.18 0.05 0.76 -1.20 -0.28 0.10 0.69 1.59 200 0.99 eta_a[19] 0.12 0.07 0.94 -1.63 -0.47 0.07 0.73 2.10 200 1.00 eta_a[20] 0.41 0.07 0.92 -1.47 -0.24 0.40 0.90 2.19 200 1.01 eta_a[21] -0.08 0.06 0.85 -1.58 -0.68 -0.10 0.49 1.58 200 0.98 eta_a[22] 0.03 0.06 0.82 -1.58 -0.58 0.05 0.58 1.55 200 1.00 eta_a[23] -0.12 0.06 0.83 -1.86 -0.58 -0.14 0.49 1.40 200 0.99 eta_a[24] 0.11 0.07 0.94 -1.62 -0.54 0.12 0.78 1.91 200 0.99 eta_a[25] -0.25 0.07 1.01 -2.09 -0.94 -0.27 0.37 1.99 200 0.99 eta_a[26] 0.98 0.07 0.97 -0.99 0.37 1.04 1.61 2.87 200 1.00 eta_a[27] -0.14 0.07 0.97 -2.05 -0.79 -0.16 0.55 1.78 200 0.98 eta_a[28] -0.61 0.07 0.99 -2.95 -1.20 -0.53 0.05 1.18 200 0.99 eta_a[29] 0.07 0.06 0.89 -1.55 -0.46 0.09 0.64 1.84 200 0.99 eta_a[30] -0.38 0.06 0.89 -1.95 -1.05 -0.37 0.25 1.38 200 1.00 eta_a[31] 0.65 0.07 0.92 -1.22 0.01 0.76 1.35 2.26 200 1.00 eta_a[32] -0.09 0.06 0.81 -1.72 -0.61 -0.12 0.46 1.36 200 0.99 a[1] 0.35 0.01 0.13 0.09 0.27 0.35 0.42 0.59 200 1.00 a[2] 0.38 0.01 0.11 0.16 0.30 0.38 0.44 0.59 200 1.00 a[3] 0.42 0.01 0.14 0.20 0.33 0.39 0.48 0.73 154 1.01 a[4] 0.21 0.01 0.19 -0.22 0.12 0.25 0.33 0.51 200 1.02 a[5] 0.28 0.01 0.14 0.03 0.20 0.29 0.36 0.59 200 0.99 a[6] 0.29 0.01 0.12 0.07 0.21 0.28 0.36 0.51 200 0.99 a[7] 0.32 0.01 0.15 0.08 0.22 0.30 0.40 0.63 200 0.99 a[8] 0.17 0.01 0.13 -0.13 0.08 0.18 0.24 0.39 200 0.99 a[9] 0.08 0.01 0.12 -0.21 0.01 0.09 0.16 0.28 161 1.00 a[10] 0.19 0.01 0.11 -0.06 0.12 0.18 0.24 0.45 200 0.99 a[11] 0.25 0.01 0.14 0.04 0.15 0.22 0.33 0.61 94 1.02 a[12] 0.05 0.01 0.13 -0.26 -0.02 0.07 0.13 0.26 200 1.00 a[13] 0.06 0.01 0.12 -0.20 -0.01 0.07 0.13 0.26 200 0.99 a[14] 0.03 0.01 0.12 -0.21 -0.03 0.04 0.09 0.31 200 1.00 a[15] -0.02 0.01 0.13 -0.32 -0.10 0.00 0.07 0.20 200 0.99 a[16] -0.05 0.01 0.12 -0.31 -0.12 -0.04 0.03 0.14 200 1.00 a[17] -0.04 0.01 0.14 -0.32 -0.11 -0.02 0.04 0.22 200 0.98 a[18] -0.01 0.01 0.11 -0.22 -0.07 -0.03 0.04 0.26 200 0.99 a[19] -0.04 0.01 0.13 -0.31 -0.12 -0.05 0.01 0.24 200 1.01 a[20] -0.02 0.01 0.14 -0.26 -0.10 -0.04 0.05 0.35 200 1.01 a[21] -0.13 0.01 0.12 -0.38 -0.20 -0.12 -0.05 0.10 200 0.99 a[22] -0.13 0.01 0.12 -0.35 -0.20 -0.13 -0.06 0.10 200 1.00 a[23] -0.18 0.01 0.13 -0.44 -0.24 -0.17 -0.11 0.03 200 1.00 a[24] -0.17 0.01 0.13 -0.41 -0.25 -0.18 -0.09 0.07 200 1.00 a[25] -0.25 0.01 0.13 -0.51 -0.33 -0.24 -0.16 -0.02 200 0.99 a[26] -0.08 0.02 0.17 -0.33 -0.20 -0.11 0.02 0.29 82 1.02 a[27] -0.29 0.01 0.14 -0.54 -0.36 -0.28 -0.19 -0.06 200 0.98 a[28] -0.37 0.01 0.16 -0.81 -0.46 -0.35 -0.27 -0.11 200 1.00 a[29] -0.29 0.01 0.12 -0.55 -0.38 -0.30 -0.23 -0.05 200 0.99 a[30] -0.39 0.01 0.14 -0.71 -0.47 -0.38 -0.29 -0.17 180 1.00 a[31] -0.25 0.01 0.15 -0.47 -0.36 -0.27 -0.17 0.06 200 1.01 a[32] -0.40 0.01 0.14 -0.69 -0.48 -0.40 -0.31 -0.14 200 1.00 lp__ -1.06 0.94 5.57 -12.58 -4.46 -0.73 2.64 9.97 35 1.08

All this output isn't so bad: the etas are the team-level residuals and the a's are team abilities. The group-level error sd sigma_a is estimated at 0.13 which is a small value, which indicates that, unsurprisingly, our final estimates of team abilities are not far from the initial ranking. We can attribute this to a combination of two factors: first, the initial ranking is pretty accurate; second, there aren't a lot of data points here (indeed, half the teams only played three games each) so it would barely be possible for there to be big changes.

Now it's time to make some graphs. First a simple list of estimates and standard errors of team abilities. I'll order the teams based on Nate's prior ranking, which makes sense for a couple reasons. First, this ordering is informative, there's a general trend from good to bad so it should be easy to understand the results. Second, the prior ranking is what we were using to pull toward in the multilevel model, so this graph is equivalent to a plot of estimate vs. group-level predictor, which is the sort of graph I like to make to understand what the multilevel model is doing.

Here's the code:

colVars <- function(a) {n <- dim(a)[[1]]; c <- dim(a)[[2]]; return(.colMeans(((a - matrix(.colMeans(a, n, c), nrow = n, ncol = c, byrow = TRUE)) ^ 2), n, c) * n / (n - 1))} sims <- extract(fit) a_sims <- sims$a a_hat <- colMeans(a_sims) a_se <- sqrt(colVars(a_sims)) library ("arm") png ("worldcup1.png", height=500, width=500) coefplot (rev(a_hat), rev(a_se), CI=1, varnames=rev(teams), main="Team quality (estimate +/- 1 s.e.)\n", cex.var=.9, mar=c(0,4,5.1,2), xlim=c(-.5,.5)) dev.off()

And here's the graph:

At this point we could compute lots of fun things such as the probability that Argentina will beat Germany in the final, etc., but it's clear enough from this picture that the estimate will be close to 50% so really the model isn't giving us anything for this one game. I mean, sure, I could compute such a probability and put it in the headline above, and maybe get some clicks, but what's the point?

One thing I *am* curious about, though, is how much of these estimates are coming from the prior ranking? So I refit the model. First I create a new Stan model, "worldcup_noprior_matt.stan", just removing the parameter "b". Then I run it from R:

fit_noprior <- stan_run("worldcup_noprior_matt.stan", data=data, chains=4, iter=100) print(fit_noprior)

Convergence after 100 iterations is not perfect---no surprise, given that with less information available, the model fit is less stable---so I brute-force it and run 1000:

fit_noprior <- stan_run("worldcup_noprior_matt.stan", data=data, chains=4, iter=1000) print(fit_noprior)

Now time for the plot. I could just copy the above code but this is starting to get messy so I encapsulate it into a function and run it:

worldcup_plot <- function (fit, file.png){ sims <- extract(fit) a_sims <- sims$a a_hat <- colMeans(a_sims) a_se <- sqrt(colVars(a_sims)) library ("arm") png (file.png, height=500, width=500) coefplot (rev(a_hat), rev(a_se), CI=1, varnames=rev(teams), main="Team quality (estimate +/- 1 s.e.)\n", cex.var=.9, mar=c(0,4,5.1,2), xlim=c(-.5,.5)) dev.off() } worldcup_plot(fit_noprior, "worldcup1_noprior.png")

I put in the xlim values in that last line to make the graph look nice, after seeing the first version of the plot. And I also played around a bit with the margins (the argument "mar" to coefplot above) to get everything to look good. Overall, though, coefplot() worked wonderfully, straight out of the box. Thanks, Yu-Sung!

Similar to before but much noisier.

**Act 2: Checking the fit of model to data**

OK, fine. Now let's check model fit. For this we'll go back to the model worldcup_matt.stan which includes the prior ranking as a linear predictor. For each match we can compute an expected outcome based on the model and a 95% predictive interval based on the t distribution. This is all on the signed sqrt scale so we can do a signed square to put it on the original scale, then compare to the actual game outcomes.

Here's the R code (which, as usual, took a few iterations to debug; I'm not showing all my steps here):

expected_on_sqrt_scale <- a_hat[team1] - a_hat[team2] sigma_y_sims <- sims$sigma_y interval_975 <- median(qt(.975,df)*sigma_y_sims) signed_square <- function (a) {sign(a)*a^2} lower <- signed_square(expected_on_sqrt_scale - interval_975) upper <- signed_square(expected_on_sqrt_scale + interval_975) png ("worldcup2.png", height=1000, width=500) coefplot (rev(score1 - score2), sds=rep(0, ngames), lower.conf.bounds=rev(lower), upper.conf.bounds=rev(upper), varnames=rev(paste(teams[team1], "vs.", teams[team2])), main="Game score differentials\ncompared to 95% predictive interval from model\n", mar=c(0,7,6,2)) dev.off ()

In the code above, I gave the graph a height of 1000 to allow room for all the games, and a width of 500 to conveniently fit on the blog page. Also I went to the trouble to explicitly state in the title that the intervals were 95%.

And here's what we get:

Whoa! Some majorly bad fit here. Even forgetting about Brazil vs. Germany, lots more than 5% of the points are outside the 95% intervals. My first thought was to play with the t degrees of freedom, but that can't really be the solution. And indeed switching to a t_4, or going the next step and estimating the degrees of freedom from the data (using the gamma(2,0.1) prior restricted to df>1, as recommended by Aki) didn't do much either.

**Act 3: Thinking hard about the model. What to do next?**

So then I got to thinking about problems with the model. Square root or no square root, the model is only approximate, indeed it's not a generative model at all, as it makes continuous predictions for discrete data. So my next step was to take the probabilities from my model, use them as predictions for each discrete score differential outcome (-6, -5, -4, . . ., up to +6) and then use these probabilities for a multinomial likelihood.

I won't copy the model in this post but it's included in the zipfile that I've attached here. The model didn't work at all and I didn't get around to fixing it.

Indeed, once we move to directly modeling the discrete data, a latent-data approach seems like it might be more reasonable, perhaps something a simple as an ordered logit or maybe something more tailored to the particular application.

Ordered logit's in the Stan manual so it shouldn't be too difficult to implement. But at this point I was going to bed and I knew I wouldn't have time to debug the model the next day (that is, today) so I resolved to just write up what I *did* do and be open about the difficulties.

These difficulties are not, fundamentally, R or Stan problems. Rather, they represent the challenges that arise when analyzing any new problem statistically.

**Act 4: Rethinking the predictive intervals**

But then, while typing up the above (it's taken me about an hour), I realized that my predictive intervals were too narrow because they're based on point estimates of the team abilities. How could I have been so foolish??

Now I'm excited. (And I'm writing here in real time, i.e. I have not yet cleaned up my code and recomputed those predictive error bounds.) Maybe everything will work out. That would make me so happy! The fractal nature of scientific revolutions, indeed.

OK, let's go to it.

*Pause while I code*

OK, that wan't too bad. First I refit the model using a zillion iterations so I can get the predictive intervals using simulation without worrying about anything:

fit <- stan_run("worldcup_matt.stan", data=data, chains=4, iter=5000) print(fit)

Now the code:

sims <- extract (fit) a_sims <- sims$a sigma_y_sims <- sims$sigma_y nsims <- length(sigma_y_sims) random_outcome <- array(NA, c(nsims,ngames)) for (s in 1:nsims){ random_outcome_on_sqrt_scale <- (a_sims[s,team1] - a_sims[s,team2]) + rt(ngames,df)*sigma_y_sims[s] random_outcome[s,] <- signed_square(random_outcome_on_sqrt_scale) } sim_quantiles <- array(NA,c(ngames,2)) for (i in 1:ngames){ sim_quantiles[i,] <- quantile(random_outcome[,i], c(.025,.975)) } png ("worldcup3.png", height=1000, width=500) coefplot (rev(score1 - score2), sds=rep(0, ngames), lower.conf.bounds=rev(sim_quantiles[,1]), upper.conf.bounds=rev(sim_quantiles[,2]), varnames=rev(paste(teams[team1], "vs.", teams[team2])), main="Game score differentials\ncompared to 95% predictive interval from model\n", mar=c(0,7,6,2)) dev.off ()

And now . . . the graph:

Hmm, the uncertainty bars are barely wider than before. So time for a brief re-think. The problem is that the predictions are continuous and the data are discrete. The easiest reasonable way to get discrete predictions from the model is to simply round to the nearest integer. So we'll do this, just changing the appropriate line in the above code to:

sim_quantiles[i,] <- quantile(round(random_outcome[,i]), c(.025,.975))

Aaaaand, here's the new graph:

Still no good. Many more than 1/20 of the points are outside the 95% predictive intervals.

This is particularly bad given that the model was fit to the data and the above predictions are not cross-validated, hence we'd actually expect a bit of overcoverage.

One last thing: Instead of plotting the games in order of the data, I'll order them by expected score differential. I'll do it using difference in prior ranks so as not to poison the well with the estimates based on the data. Also I'll go back to the continuous predictive intervals as they show a bit more information:

a_hat <- colMeans(a_sims) new_order <- order(a_hat[team1] - a_hat[team2]) for (i in 1:ngames){ sim_quantiles[i,] <- quantile(random_outcome[,i], c(.025,.975)) } png ("worldcup5.png", height=1000, width=500) coefplot ((score1 - score2)[new_order], sds=rep(0, ngames), lower.conf.bounds=sim_quantiles[new_order,1], upper.conf.bounds=sim_quantiles[new_order,2], varnames=paste(teams[team1[new_order]], "vs.", teams[team2[new_order]]), main="Game score differentials\ncompared to 95% predictive interval from model\n", mar=c(0,7,6,2)) dev.off ()

Hey, this worked on the first try! Cool.

But now I'm realizing one more thing. Each game has an official home and away team but this direction is pretty much arbitrary. So I'm going to flip each game so it's favorite vs. underdog, with favorite and underdog determined by the prior ranking.

So here it is, with results listed in order from expected blowouts to expected close matches:

Again, lots of outliers, indicating that my model sucks. Maybe there's also some weird thing going on with Jacobians when I'm doing the square root transformations. Really I think the thing to do is to build a discrete-data model. In 2018, perhaps.

Again, all the data and code are here. (I did not include last night's Holland/Brazil game but of course it would be trivial to update the analyses if you want.)

**P.S.** In the 9 June post linked above, Nate wrote that his method estimates that "a 20 percent chance that both the Netherlands and Chile play up to the higher end of the range, or they get a lucky bounce, or Australia pulls off a miracle, and Spain fails to advance despite wholly deserving to." I can't fault the estimate---after all, 1-in-5 events happen all the time---but I can't see why he was so sure that if Spain loses, that they still wholly deserved to win. They lost 5-1 to Holland and 2-0 to Chile so it doesn't seem like just a lucky bounce. At some point you have to say that if you lost, you got what you deserved, no?

P.P.S. The text in those png graphs is pretty hard to read. Would jpg be better? I'd rather not use pdf because then I'd need to convert it into an image file before uploading it onto the blog post.

P.P.P.S. I found a bug! The story is here.

Thanks for this postmortem. It’s nice to be reminded that I’m not the only one who has to run the same basic model a bunch of times to deal with bugs large and small.

What is the “Matt trick”?

And — this question probably comes from my naivete about sports, but it seems like modeling score differentials carries the implicit assumption that total points scored doesn’t matter. Is that reasonable? To elaborate — suppose Team A plays B and wins 1-0, and team C plays D and wins 7-6. Do those final scores tell you the same thing about the relative abilities of Teams A and C?

For the full details, see the Betancourt and Girolami paper on hierarchical models in HMC:

http://arxiv.org/abs/1312.0906

We called the non-centered parameterization the “Matt trick” because Matt Hoffman derived it from first principles while working on Stan. There are more details on how to use it in Stan in the modeling language manual’s chapter on variable transforms.

Does that have anything to do with the ideas in Yu and Meng’s ancillarity-sufficiency interweaving strategy paper?

Not directly. Yu and Meng selects from two different parameterizations (well, one parameterization and one parameter expansion), depending on the conditioning variables, to ease Gibbs sampling, where as here we’re just choosing a single optimal parameterization because HMC jointly updates all of the parameters at once. The underlying reasons for the improved performance are also different.

Your graphs seem fine. The text is readable. I don’t think you need to change!

But, why are the names in French?

I assume that they are official FIFA names, and FIFA is a French institution.

No, it’s just that my data-entry team typed in the match results from a French book.

Great post!

Imagine all methods sections were this detailed and honest.

Why did you model the difference of scores, and not the scores themselves ? Each one could be a Poisson (or negative binomial, or a gamma distribution), with a strength_y parameter by team.

The short answer is that the number of goals scores depends on the opponent’s defense. By modeling differential, you roll everything together into one model.

Andrew:

Can you post the two text files? soccerpowerindex.txt & worldcup2012.txt?

As predictors, what all finally went into this model? The Nate silver rankings for teams & 2012 world cup score differentials? Is that it? Or is there more stuff?

Rahul:

I posted all the files; see the link near the end of the post.

Thanks. Stupid me missed it on first read.

For the code in html, you need to prefix a “pre” tag in front of the “code” tag to preserve the formatting.

For the graphs, png is a better default than jpeg. Png does not cause any loss in the image, whereas jpeg does. This loss in jpegs is good for raster like images (e.g. photographs or maps with satellite imagery) on the web because they are smaller in size. For statistical graphics that are just filled with vector elements png is fine.

(You could export to svg for the web, but that is probably overkill. For detailed graphs I typically save the pngs as very large, and my wordpress theme automatically shrinks the images to fit the blog post, but if someone opens them in a new tab they can see the large image.)

What Andy W. said. I just updated the post to use “pre” tags instead of “code” tags.

Hello, Bob,

since you seem to use WordPress, you may want to read up the documentation concerning code-Tag:

https://codex.wordpress.org/Writing_Code_in_Your_Posts

I’m using a WordPress.com blog, where Automattic provides a nice shortcut with syntax highlighting:

http://en.support.wordpress.com/code/posting-source-code/

Read the end of that article to learn, how to use it in WordPress.org blogs.

Ehm, Andrew is the author. I’m sorry.

PNG vs JPEG: PNG is a lossless compression, whereas JPEG basically does some kind of fourier transform and truncates the high frequencies. Because of this, text, which has lots of high frequency edges, looks like CRAP, so no, PNG is definitely the right way to put graphs and text on the web.

I have more to say about this post, which I think is fabulous, but have to go get breakfast for my kids.

There are probably much easier ways to do this without the manual labor by fiddling with a few knobs on your blog host software, but until then I might reccomend using this easy service which already understands how to format Stan:

http://pygments.org/demo/462830/

(Sorry my demo example sadly also lost the benefit of proper indentations due to it being a copy and paste from your blog)

Scroll to bottom of page to paste in and format more examples for yourself(without having to permabebtly save a demo link like I did above unless you want to). Don’t forget to select “Stan” from the drop down menu on top right. After it runs you also get a choice between many different color “themes” in yet another drop down if you don’t like the defaults.

I’d be happy to set something up for your blog so you don’t have to worry about it anymore, feel free to inquire at obvious places based on what my email address reveals.

Wow, someone’s already written a syntax highlighting template for Stan?

Yeah, though that’s not very surprising to me. People writing and maintaing software of various forms these days tend to have a fetish for fancy colors and it’s not at all hard to define new grammar rules to format them for something like pygments. (which can be used in many more places than html output for web content…)

Everyone loves pretty colors and pretty graphs. =)

Hey, we’ve even got a slick emacs mode that Jeffrey Arnold’s been keeping up to date.

Skellam distribution is sometimes used to model the difference of (discrete) scores in sports, link: http://en.wikipedia.org/wiki/Skellam_distribution

Would that work in Stan? (I guess modified Bessel function might be the problem?)

We have Bessel functions, but I don’t think they’re this particular one. No idea how easy it would be to implement.

Given that the Skellam distro is the difference of two Poissons with different means, we could model the Poissons directly given that we have the raw data, not just the differences.

Rating Australia last? Come down Under and say that….

That was Nate. Andrew’s model put Honduras and Cameroon lower (having observed the actual games).

AG, can’t you hear the thunder? You better run, you better take cover.

Andrew, there’s no Jacobian needed — it’s just a data transform.

Once you start doing arbitrary transforms like this, there’s no stopping. Why not a log basis (with something dealing with the 0 case)?

Bob:

Ultimately we’d want to analyze the discrete data directly. Any of these transformations is just a way to keep as much of the data information as possible using a convenient model. In this case, convenience is not so much in programming as in having an understandable model.

Isn’t the real problem with this that you model a team with a single ability parameter rather than an ability and a variability? If you modeled variability per team, saying that the ability they brought to a game was their mean plus or minus some variability value, then the model would probably fit Brazil as highly skilled, but also highly variable, and the Germany match would look like less of an outlier.

Bob:

This could be done but I don’t think there’s enough information in the data to estimate team-level variability. I mean, sure, if you fit such a model to these data you’d find that Germany and Brazil have high estimated variabilities, but I don’t know that this would really give you anything useful.

Perhaps many team parameters (e.g. defense versus offense?) would be required to generate data like this and just taking the differences may not catch an important commonness separately from those.

The major problem seems to be that the predictive intervals for goal differences usually only cover three numbers and sometimes even only two. That’s way too small. I guess Andrew is right that this has to do with not using proper discrete modelling. I also guess that these difference distributions given the ability difference could look somewhat nonstandard. Probably a large number of exactly zero differences contaminates the estimated distribution if the shape is assumed to be t or normal (perhaps even worse so for t with low df than for the normal).

This takes me back fifteen years, when I fit a modified Bradley-Terry model to results of Ultimate Frisbee games in order to create a ranking for teams. I had a similar issue: too many games were decided by much larger margins than the model predicted. And that’s in Ultimate, where games are played to a fixed number of points rather than a fixed amount of time, so even a team that is down by a large margin still has a chance to come back; even so, I think that once a game seemed out of reach — a team behind 12-7 in a game to 15, say — there’s a tendency for the margin to widen rather than narrow. (All I had was the score for the complete game, so I couldn’t use the score history).

In soccer a team that trails by two goals (or more) in the second half faces a rather dismal situation: to have any chance of winning they have to play in a way that increases their own chance of scoring, compared to the way they played at the start of the game, but that increases their opponents’ chance of scoring even more. Their probability of winning increases, but so does the magnitude of their expected loss. (If this seems confusing or impossible, consider this example: you are playing a game with dice. If you roll a 1 your opponent gets a point, if you roll a 6 you get a point, and any other roll is neutral. You are trailing by one point and you have a single roll remaining in the game. You can choose to roll a conventional die, or to roll a die that has four 1′s and two 6′s. If you pick the conventional die, you have 1/6 chance of tying the game and 1/6 chance of losing by 2. If you pick the other die, you have a 1/3 chance of tying the game, but if you don’t tie the game you are guaranteed to lose by 2: you’ve substantially increased your chance of winning while simultaneously increasing your chance of losing by a larger margin. In more general terms, you’ve decreased your expected victory margin but increased the variance of the outcomes).

Andrew says “My first thought was to play with the t degrees of freedom, but that can’t really be the solution.” I don’t see why that can’t be the solution. But Andrew says he tried it and it didn’t help, so I agree that it _isn’t_ the solution. (But how could it not help?) I do think a wide-tailed distribution makes sense, though, so if I had to pick a number of degrees of freedom I’d be inclined to use a number like 3 or 4 instead of 7. It may not help much but it still seems better to me.

If having a wider distribution for p( y_ij | a_i – a_j) doesn’t fix the problem then I suppose one has to look to more complicated models. One thing that comes to mind is that team abilities change with time — Brazil was missing their best offensive player and one of their best defensive players in the loss to Germany, for example. Some of the group round games are essentially meaningless to one or both teams — they’ve already been eliminated — which one would expect to lead to less effort (or lower “ability”) in those games. Data on injuries and disqualification due to penalties are available, and of course it’s possible to determine whether a game is meaningful or not, so at least those factors could be incorporated in the model. But I would have expected that simply using a wider-tailed distribution would help a lot and I’m still surprised that it doesn’t.

Phil: I think the problem is rather with estimating the sigma than with how heavy the tails are. The “outliers” here are not really that extreme. The largest observed difference was 6, sqrt(6)=2.45. Given that most values were +/- 0, 1, sqrt(2), this is not really so far off that a t_3 or something would be needed to model it.

There is a literature on “conflict success functions” that might be relevant here.

I love seeing this kind of analysis, thinking through an applied statistical problem with code, exploring the data, and iterating on the model. Something I’d like to see more of on statistics blogs.

The game score difference is going to be heavy tailed, and I don’t think it captures the difference in team skills.

I’d prefer to treat each goal as a binary variable, and run logistic regression trying to predict which team scored it. It seems like 1-0 games should tell you less about team skills than 6-5 games, which this would do. In the 6-5 case, you are pretty sure that the teams are almost evenly matched. In the 1-0 case, you would think the team that won is better, but with very little confidence. This matches better how I think about teams after a game.

[…] The other day I fit a simple model to estimate team abilities from World Cup outcomes. I fit the model to the signed square roots of the score differentials, using the square root on the theory that when the game is less close, it becomes more variable. […]

[…] and commonsensical (but hardly routine) data analysis. Those of you who read my recent posts (here and here) on the World Cup might be interested in this, because it has lots of practical details of […]