Visualizing asylum statistics

Note: of potential interest to R users for the dynamic Google chart generated via googleVis in R and discussed towards the end of the post. Here you can go directly to the graph.

02alessandro-penso
An emergency refugee center, opened in September 2013 in an abandoned school in Sofia, Bulgaria. Photo by Alessandro Penso, Italy, OnOff Picture. First prize at World Press Photo 2013 in the category General News (Single).

The tragic lives of asylum-seekers make for moving stories and powerful photos. When individual tragedies are aggregated into abstract statistics, the message gets harder to sell. Yet, statistics are arguably more relevant for policy and provide for a deeper understanding, if not as much empathy, than individual stories. In this post, I will offer a few graphs that present some of the major trends and patterns in the numbers of asylum applications and asylum recognition rates in Europe over the last twelve years. I focus on two issues: which European countries take the brunt of the asylum flows, and the link between the application share that each country gets and its asylum recognition rate.

Asylum applications and recognition rates
Before delving into the details, let’s look at the big picture first. Each year between 2001 and 2012, 370,000 people on average have applied for asylum protection in one of the member states of the European Union (plus Norway and Switzerland). As can be seen from Figure 1, the number fluctuates between 250,000 and 500,000 per year, and there is no clear trend. Altogether, during this 12-year period, approximately 4.5 million people have applied for asylum, which makes slightly less than one percent of the total EU population. Of course, this figure only tracks people who have actually made it to the asylum centers and filed an application – all potential refugees who have perished on the way, or have arrived but been denied the right of formal application, or have remained clandestine are not counted.

asylum_applications_small

Figure 1 also shows the annual number of persons actually recognized as ‘refugees’ under the terms of the Geneva Convention by the European governments: a status which grants considerable rights and protection. This number is quite lower with an average of around 40.000 per year (in the EU+ as a whole) which makes for less than half-a-million in total for the 12 years between 2001 and 2012. While the overall recognition rate remains between 7% and 14%, there is considerable variation between the different European states both in the share from the asylum flows they receive, and in the national asylum recognition rates.

Who takes the brunt of the asylum burden?
Both the asylum flows and the recognition rates are in fact distributed highly unequally across the continent, and in a way that cannot be completely accounted for by the wealth of destination countries, former (colonial) ties between asylum sources and destinations, nor geographical distance. To compare the shares of the total European pool of asylum applications and recognitions that a destination country gets, I create the so-called ‘burden coefficient’. The ‘burden coefficient’ compares the actual share of asylum applications a country received in a year to its ‘fair’ share which is defined as its relative share of the annual  total EU+ GDP. Simply put, if a country accounts for 10% of the European GDP, it would have been expected to receive 10% of all asylum applications filed in Europe that year. Taking account of GDP adjusts the raw asylum application shares in view of the expectation that richer and more populous countries should bear a proportionally higher share of the total European asylum ‘burden’ than poorer and smaller states.

asylum_applications_burden

Figure 2 shows the (logged) burden coefficient for asylum application shares for each EU+ country, averaged over the period 2010-2012. The solid line at zero indicates an asylum applications share perfectly proportional to a  country’s GDP share (a ‘fair’ burden). Countries with positive values receive a higher share of all applications than implied by their GDP level, and countries with negative values receive a lower than their implied share. (The dotted lines show where a country that is doing twice as much / twice as little as expected would be). Clearly, Spain, Portugal, Italy and many (but not all) of the East European countries underdeliver while Cyprus, Malta, Greece, and several West European states (notably Sweden, Belgium, and Norway) take a disproportionately high  share of the total pool of asylum applications filed in Europe over the last few years. Note that these comparisons already take into account (correct for) the fact that most of the Southern and Eastern European countries are poorer (have lower GDP) than the ones in the Western and Northern parts of the continent.

asylum_recognitions_burden

The picture does not change much when we focus on actual asylum recognitions (under the terms of the Geneva Convention) instead of applications. Figure 3 shows the burden coefficient (again averaged over 2010-2012) for full status refugee recognitions in Europe. The country ranking is similar with a few important exception – Greece grants much fewer asylum recognitions than expected even after we account for the state of its economy; Austria and Switzerland join the ranks of states which do much more than their implied share; and, sadly, many more countries in fact underdeliver when it comes to full refugee status grants. (Note that some states offer alternative protection to those denied the full ‘Geneva Convention’ status but the forms and level of this protection differs significantly across the continent).

Are asylum application shares responsive to the recognition rate?
Given these rather significant discrepancies across Europe in how many asylum applications countries get, and how much protection they offer, it is natural to ask whether the applications shares and the recognition rates are in fact related. Do asylum seekers flock at the gates of the European states which are most generous in their recognition policy? Do low recognition rates deter potential refugees from applying in certain countries? Can the strictness of asylum policy be an effective policy tool shaping future application flows? A comprehensive statistical analysis shows that while application shares and recognition rates are associated, their responsiveness to each other is rather weak. Simply put, manipulating the recognition rates is unlikely to have big practical effects on the asylum application share a country receives, and changes in the applications rates only weakly affect state recognition rates. The details of the analysis are rather technical and can be found here, but a dynamic visualization can help illustrate the patterns.

The dynamic interactive chart linked here shows the relationship between asylum applications and asylum recognition rates for each EU+ country over the last 12 years (the chart cannot be embedded in this post due to WordPress policy, but there is a screenshot below). When you press ‘Play’ each dot traces the experience of one country over time. You can choose to observe all, select a single state to focus upon, or tick a couple to compare their experiences.

dynamic-asylum-1

A movement of a dot (and the trace in leaves) in a horizontal direction means that the number of asylum applications received by a country increases while the recognition rates remains the same. Similarly, a vertical move implies a change in the recognition rate but a stable asylum application flow. A trajectory that follows a diagonal suggests a link between applications and recognition rates.

When paused, the state of the chart at each year shows the cross-sectional association between applications and recognition rates: it is easy to see that there is a (rather stable) weakly-strong positive relationship. But the trajectories of individual countries over time do not suggest that there is a temporal link between the two aspects of asylum policy for particular countries. For example, in the UK between 2001 and 2004 both the recognition rates and the applications fall, which would suggest strong responsiveness, but then the recognition rate moves up from 4% to almost 30% without any significant increase in applications. The trajectory of Denmark (try it out) exhibits something close to a dynamic link with rates depressing applications initially but then when they rise again, applications seem to pick up as well. Of course, asylum flows are driven by many other factors as well, so while suggestive, the patterns in the chart should be interpreted with care.

dynamic-asylum-2

More comprehensive analyses of asylum policy in Europe addressing these questions and more are available in my published articles accessible here and here. The original data comes from the UNHCR annual reports. The dynamic chart is generated using Google Chart Tools through the googleVis library in R, you can find the code here. I found it useful to generate a simple version, adjust the settings manually, and then copy the final settings via the Google Chart’s Advanced Panel back to R.

Predicting movie ratings with IMDb data and R

It’s Oscars season again so why not explore how predictable (my) movie tastes are. This has literally been a million dollar problem and obviously I am not gonna solve it here, but it’s fun and slightly educational to do some number crunching, so why not. Below, I will proceed from a simple linear regression to a generalized additive model to an ordered logistic regression analysis. And I will illustrate the results with nice plots along the way. Of course, all done in R (you can get the script here).

Data
The data for this little project comes from the IMDb website and, in particular, from my personal ratings of 442 titles recorded there. IMDb keeps the movies you have rated in a nice little table which includes information on the movie title, director, duration, year of release, genre, IMDb rating, and a few other less interesting variables. Conveniently, you can export the data directly as a csv file.

Outcome variable
The outcome variable that I want to predict is my personal movie rating. IMDb lets you score movies with one to ten stars. Half-points and other fractions are not allowed. It is a tricky variable to work with. It is obviously not a continuous one; at the same time ten ordered categories are a bit too many to treat as a regular categorical variable. Figure 1 plots the frequency distribution (black bars) and density (red area) of my ratings and the density of the IMDb scores (in blue) for the 442 observations in the data.

figure1

The mean of my ratings is a good 0.9 points lower than the IMDb scores, which are also less dispersed and have a higher peak (can you say ‘kurtosis’).

Data-generating process
Some reflection on how the data is generated can highlight its potential shortcomings. First, life is short and I try not to waste my time watching bad movies. Second, even if I get fooled to start watching a bad movie, usually I would not bother rating it on IMDb.There are occasional two- and three-star scores, but these are usually movies that were terrible and annoyed me for some reason or another (like, for example, getting a Cannes award or featuring Bill Murray). The data-generating process leads to a selection bias with two important implications. First, the effective range of variation of both the outcome and the main predictor variables is restricted, giving the models less information to work with. Second, because movies with a decent IMDb ratings which I disliked have a lower chance of being recorded in the dataset, the relationship we find in the sample will overestimate the real link between my ratings and the IMDb ones.

Take one: linear regression
Enough preliminaries, let’s get to business. An ordinary linear regression model is a common starting point for analysis and its results can serve as a baseline. Here are the estimates that lm provides for regressing my ratings on IMDb scores:

summary(lm(mine~imdb, data=d))

Coefficients:
            Estimate Std. Error t value Pr(>|t|)    
(Intercept)  -0.6387     0.6669  -0.958    0.339    
imdb          0.9686     0.0884  10.957   ***
---
Residual standard error: 1.254 on 420 degrees of freedom
Multiple R-squared: 0.2223,	Adjusted R-squared: 0.2205

The intercept indicates that on average my ratings are more than half a point lower. The positive coefficient of IMDb score is positive and very close to one which implies that one point higher (lower) IMDb rating would predict, on average, one point higher (lower) personal rating. Figure 2 plots the relationship between the two variables (for an interactive version of the scatter plot, click here):

figure2

The solid black line is the regression fit, the blue one shows a non-parametric loess smoothing which suggests some non-linearity in the relationship that we will explore later.

Although the IMDb score coefficient is highly statistically significant that should not fool us that we have gained much predictive capacity. The model fit is rather poor. The root mean squared error is 1.25 which is large given the variation in the data. But the inadequate fit is most clearly visible if we plot the actual data versus the predictions. Figure 3 below does just that. The grey bars show the prediction plus/minus two predictive standard errors. If the predictions derived from the model were good, the dots (observations) would be very close to the diagonal (indicated by the dotted line). In this case, they are not. The model does a particularly bad job in predicting very low and very high ratings.

figure3

We can also see how little information IMDb scores contain about (my) personal scores by going back to the raw data. Figure 4 plots to density of my ratings for two sets of values of IMDb scores – from 6.5 to 7.5 (blue) and from 7.5- to 8.5 (red). The means for the two sets differ somewhat, but the overlap in the density is great.

figure4

In sum, knowing the IMDb rating provides some information but on its own doesn’t get us very far in predicting what my score would be.

Take two: adding predictors
Let’s add more variables to see if things improve. Some playing around shows that among the available candidates only the year of release of the movie and dummies for a few genres and directors (selected only from those with more than four movies in the data) give any leverage.

 summary(lm(mine~imdb+d$comedy +d$romance+d$mystery+d$"Stanley Kubrick"+d$"Lars Von Trier"+d$"Darren Aronofsky"+year.c, data=d))

Coefficients:
                      Estimate Std. Error t value Pr(>|t|)    
(Intercept)           1.074930   0.651223   1.651  .  
imdb                  0.727829   0.087238   8.343  ***
d$comedy             -0.598040   0.133533  -4.479  ***
d$romance            -0.411929   0.141274  -2.916  ** 
d$mystery             0.315991   0.185906   1.700  .  
d$"Stanley Kubrick"   1.066991   0.450826   2.367  *  
d$"Lars Von Trier"    2.117281   0.582790   3.633  ***
d$"Darren Aronofsky"  1.357664   0.584179   2.324  *  
year.c                0.016578   0.003693   4.488  ***
---
Residual standard error: 1.156 on 413 degrees of freedom
Multiple R-squared: 0.3508,	Adjusted R-squared: 0.3382

The fit improves somewhat. The root mean squared error of this model is 1.14. Moreover, looking again at the actual versus predicted ratings, the fit is better, especially for highly rated movies – no surprise given that the director dummies pick these up.

figure5

The last variable in the regression above is the year of release of the movie. It is coded as the difference from 2014, so the positive coefficient implies that older movies get higher ratings. The statistically significant effect, however, has no straightforward predictive interpretation. The reason is again selection bias. I have only watched movies released before the 1990s that have withstood the test of time. So even though in the sample older films have higher scores, it is highly unlikely that if I pick a random film made in the 1970s I would like it more than a random film made after 2010. In any case, Figure 6 below plots the year of release versus the residuals from the regression of my ratings on IMDb scores (for the subset of films after 1960). We can see that the relationship is likely nonlinear (and that I really dislike comedies from the 1980s).

figure6

So far both regressions assumed that the relationship between the predictors and the outcome is linear. Needless to say, there is no compelling reason why this should be the case. Maybe our predictions will improve if we allow the relationships to take any form. This calls for a generalized additive model.

Take three: generalized additive model (GAM)
In R, we can use the mgcv library to fit a  GAM. It doesn’t make sense to hypothesize non-linear effects for binary variables, so we only smooth the effects of IMDb rating and year of release. But why stop there, perhaps the non-linear effects of IMDb rating and release year are not independent, why not allow them to interact!

library(mgcv)
summary(gam(mine ~ te(imdb,year.c)+d$"comedy " +d$"romance "+d$"mystery "+d$"Stanley Kubrick"+d$"Lars Von Trier"+d$"Darren Aronofsky", data = d)) 

PParametric coefficients:
                     Estimate Std. Error t value Pr(|t|)    
(Intercept)           6.80394    0.07541  90.225   ***
d$"comedy "          -0.60742    0.13254  -4.583   ***
d$"romance "         -0.43808    0.14133  -3.100   ** 
d$"mystery "          0.32299    0.18331   1.762   .  
d$"Stanley Kubrick"   0.83139    0.45208   1.839   .  
d$"Lars Von Trier"    2.00522    0.57873   3.465   ***
d$"Darren Aronofsky"  1.26903    0.57525   2.206   *  
---
Approximate significance of smooth terms:
                  edf Ref.df     F p-value    
te(imdb,year.c) 10.85  13.42 11.09

Well, the root mean squared error drops to 1.11 and the jointly smoothed (with a full tensor product smooth) variables are significant, but the added predictive value is minimal in this case. Nevertheless, the plot below shows the smoothed terms are more appropriate than the linear ones, and that there is a complex interaction between the two:

figure7

Take four: models for categorical data
So far we treated personal movie ratings as if they were a continuous variable, but they are not – taking into account that they are essentially an ordered categorical variable might help. But ten categories, while possible to model, would make the analysis rather unwieldy, so we recode the personal ratings into five categories without much loss of information: 5 and less, 6,7,8,9 and more.

We can first see a nonparametric conditional destiny plot of the newly created categorical variable as a function of IMDb scores:
figure8

The plot shows the observed density for each category of the outcome variable along the range of the predictor. For example, for a film with an IMDb rating of ’6′, about 35% of the personal scores are ’5′, a further 50% are ’6′, and the remaining 15% are ’7′. Remember that the plot is based on the observed conditional frequencies only (with some smoothing), not on the projections of a model. But the small ups and downs seem pretty idiosyncratic. We can also fit an ordered logistic regression model, which would be appropriated for the categorical outcome variable we have, and plot its predicted probabilities given the model.

First, here is the output of the model:

library(MASS)
summary(polr(as.factor(mine.c) ~ imdb+year.c,  Hess=TRUE, data = d)
Coefficients:
        Value Std. Error t value
imdb   1.4103   0.149921   9.407
year.c 0.0283   0.006023   4.699

Intercepts:
    Value   Std. Error t value
5|6  9.0487  1.0795     8.3822
6|7 10.6143  1.1075     9.5840
7|8 12.1539  1.1435    10.6289
8|9 14.0234  1.1876    11.8079

Residual Deviance: 1148.665 
AIC: 1160.665

The coefficients of the two predictors are significant. The plot below shows the predicted probability of the outcome variable – personal movie rating – being in each of the five categories as a function of IMDb rating and illustrates the substantive scale of the effect.

figure9

Compared to the non-parametric conditional density plot above, these model-based predictions are much smoother and have ‘disciplined’ the effect of the predictor to follow a systematic pattern.

It is interesting to ponder which of the two would be more useful for out-of-sample predictions. Despite the fact that the non-parametric one is more faithful to the current data, I think I would go for the parametric model projections. After all, is it really plausible that a random film with an IMDb rating of 5 would have lower chance a getting a 5 from me than a film with an IMDb rating of 6, as the non-parametric conditional density plot suggests? I don’t think so. Interestingly, in this case the parametric model has actually corrected for some of the selection bias and made for more plausible out-of-sample predictions.

Conclusion
In sum, whatever the method, it is not very fruitful to try to predict how much a person (or at least, the particular person writing this) would like a movie based on the average rating the movie gets and covariates like the genre or the director. Non-linear regressions and other modeling tricks offer only marginal predictive improvements over a simple linear regression approach, but bring plenty of insight about the data itself.

What is the way ahead? Obviously, one would want to get more relevant predictors, but, unfortunately, IMDb seems to have a policy against web-scrapping from its database, so one would either have to ask for permission or look at a different website with a more liberal policy (like Rotten Tomatoes perhaps). For me, the purpose of this exercise has been mostly in its methodological educational value, so I think I will leave it at that. Finally, don’t forget to check out the interactive scatterplot of the data used here which shows a user’s entire movie rating history at a glance.

Endnote
As you would have noted, the IMDb ratings come at a greater level of precision (like 7.3) than the one available for individual users (like 7). So a user who really thinks that a film is worth 7.5 has to pick 7 or 8, but its average IMDb score could well be 7.5. If the rating categories available to the user are indeed too coarse, this would show up in the relationship with the IMDb score: movies with an average score of 7.5 would be less predictable that movies with an average score of either 7 or 8. To test this conjecture, a rerun the linear regression models on two subsets of the data: one comprising the movies with an average IMDb rating between 5.9 and 6.1, 6.9 an 7.1, etc., and a  second one comprising those with an average IMDb rating between 5.4 and 5.6, 6.4 and 6.6, etc. The fit of the regression for the first group was better than for the second (RMSE of 1.07 vs. 1.11), but, frankly, I expected a more dramatic difference. So maybe ten categories are just enough.

Music Network Visualization

Note: probably of interest only to the intersection of the readers who are into niche music genres and those interested in network visualization.

My music interests have always been rather, hmm…, eclectic. Somehow IDM, ambient, darkwave, triphop, acid jazz, bossa nova, qawali, Mali blues and other more or less obscure genres have managed to happily co-exist in my music collection. The sheer diversity always invited the question whether there is some structure to the collection, or each genre is an island of its own. Sounds like a job for network visualization!

Now, there are plenty of music network viz applications on the web. But they don’t show my collection, and just seem unsatisfactory for various reasons. So I decided to craft my own visualization using R and igraph.

As a first step I collected for all artists in my last.fm library the artists that the site classifies as similar. So I piggyback on last.fm for the network similarity measures. I also get info on the most-often used tag for the artist and the number of plays it has on the site. The rest is pretty straightforward as can be seen from the code.

# Load the igraph and foreign packages (install if needed)
require(igraph)
require(foreign)
lastfm<-read.csv("http://www.dimiter.eu/Data_files/lastfm_network_ad.csv", header=T,  encoding="UTF-8") #Load the dataset

lastfm$include<-ifelse(lastfm$Similar %in% lastfm$Artist==T,1,0) #Index the links between artists in the library
lastfm.network<-graph.data.frame(lastfm, directed=F) #Import as a graph

last.attr<-lastfm[-which(duplicated(lastfm$Artist)),c(5,3,4) ] #Create some attributes
V(lastfm.network)[1:106]$listeners<-last.attr[,2]
V(lastfm.network)[107:length(V(lastfm.network))]$listeners<-NA
V(lastfm.network)[1:106]$tag<-last.attr[,3]
V(lastfm.network)[107:length(V(lastfm.network))]$tag<-NA #Attach the attributes to the artist from the library (only)
V(lastfm.network)$label.cex$tag<-ifelse(V(lastfm.network)$listeners>1200000, 1.4, 
                                    (ifelse(V(lastfm.network)$listeners>500000, 1.2,
                                            (ifelse(V(lastfm.network)$listeners>100000, 1.1,
                                                   (ifelse(V(lastfm.network)$listeners>50000, 1, 0.8))))))) #Scale the size of labels by the relative popularity

V(lastfm.network)$color<-"white" #Set the color of the dots
V(lastfm.network)$size<-0.1 #Set the size of the dots
V(lastfm.network)$label.color<-NA
V(lastfm.network)[1:106]$label.color<-"white" #Only the artists from the library should be in white, the rest are not needed

E(lastfm.network)[ include==0 ]$color<-"black" 
E(lastfm.network)[ include==1 ]$color<-"red" #Color edges between artists in the library red, the rest are not needed

fix(tkplot) #Add manually to the function an argument for the background color of the canvas and set it to black (bg=black)

tkplot(lastfm.network, vertex.label=V(lastfm.network)$name, layout=layout.fruchterman.reingold,
       canvas.width=1200, canvas.height=800) #Plot the graph and adjust as needed

I plot the network with the tkplot command which allows for the manual adjustments necessary because many artist names get on top of each other in the initial plot. Because the export options of tkplot are limited I just took a print screen ( I know, I know, that’s kind of cheating ;-)), added the tittle in Photoshop and, voila, it’s done!

[click to enlarge and explore]
my-music-netowrk

Knowing intimately the artists in the graph, I can certify that the network definitely makes a lot of sense. I love the small clusters (Flying Louts, Andy Stott, Extrawelt and Claro Intelecto [minimal/dub], or Anouar Brahem and Rabih Abou-Khalil [ethno jazz]) loosely connected to the rest of the network. And I love the fact that the boundary spanners are immediately obvious (e.g. Pink Martini between acid jazz and world music [what a stupid label by the way!], or Cesaria Evora between African and Caribbean music, or Portishead between brit-pop, trip-hop and darkwave, or Amon Tobin between trip-hop, electro and IDM). Even the different world music genres are close to each other but still unconnected. And somehow Banco De Gaya, the most ethno of all electronica in the library, ended up closest to the world/ethno clusters. There are a few problems, like Depeche Mode, which get to be pulled from the opposite sides of the graph, but these are very few.

Altogether, I have to admit I feel like a teenage dream of mine has finally been realized. But I realize the network is a rather personal thing (as it was meant to be) so I don’t expect many to get overly excited about it. Still, I would be glad to hear your comments or suggestions for extensions and improvements. And, if you were a good boy/girl during the year, I could also consider visualizing your last.fm network as a present for the new year!