# The ethics of modelling in a world where normality no longer exists

(this article was originaly writen in French – part one and two – and published in Risques) The mechanism for covering natural disasters, in France, was created to compensate “direct uninsurable material damage caused by the abnormal intensity of a natural agent” (article L. 125-1 paragraph 3 of the Insurance Code). Still on the legal level, the Court of Cassation formulated, in November 1986, a principle according to which “no one must cause others an abnormal neighbourhood disturbance”. And in order to be entitled to compensation following pre-trial detention, it is necessary for the plaintiff to prove that the detention caused him “manifestly abnormal and particularly serious harm” (Article 149 of the Code of Civil Procedure). But what does this “abnormality” in all these articles mean? According to the dictionary, abnormality is defined as “contrary to the usual order of things” (one could see there an empirical, statistical notion), “contrary to the just order of things” (this notion of “just” probably calls for a normative definition) but also “not in conformity with the model”. Defining a standard is already not simple if we are only interested in the descriptive, empirical aspect, as actuaries can do when they construct a model (especially in large dimensions, where, as we shall see, normality no longer exists), but if we also integrate a dimension of justice and ethics, we wonder if the task is not impossible…

# The average man from Quetelet and Galton

In the 19th century, if several astronomers measured the speed of the same celestial object, they obtained (often) several different measurements. In order to know which one to use in their calculations, the idea of using the “averages method” was quickly imposed – as Stahl recalls, and especially Sheynin – this average having a greater precision than any other quantity (or would now say statistically). From a set of observations $\{x_1,...,x_n\}$, we set
$$\bar x=\frac{(x_1+⋯+x_n)}{n}=\frac{1}{n}\sum_{i=1}^nx_i$$We can note that this size is also solution of the optimization problem
$$\bar x=\text{argmin}\lbrace\sum_{i=1}^n(x_i-m)^2\rbrace$$which shows the importance of “least squares”. Adolphe Quételet was, it seems, the first to apply this calculation of averages to human measures, introducing his famous concept of the “average man”. If we define the mean using a quadratic error minimization, we have an interpretation in terms of forecast: the mean size is the size that a randomly drawn person should measure (up to a random – and unpredictable – variation). In 1846, in a letter Adolphe Quételet used the image of the gladiator statue to explain what the average man might be:

Suppose a thousand statues were used to copy the gladiator with all the care imaginable. Your Highness certainly does not think that the thousand copies that will have been made will each reproduce exactly the model, and that by measuring them successively, the thousand measures that I would obtain would be as concordant as if I had taken them all on the statue of the gladiator himself. The first chances of error would be joined by the inaccuracies of the copyists; so that the probable error would perhaps be very great. Despite this, if the copyists have not worked with preconceived ideas, exaggerating or reducing certain proportions according to school prejudices, and if their inaccuracies are only accidental, the thousand measures, grouped in order of magnitude, will still present a remarkable regularity and will follow one another in the order assigned to them by the law of possibility. I see Your Highness smiling; she will no doubt tell me that such assertions will not compromise me, since we will not be willing to try the experiment. And why not? Perhaps I will surprise him by saying that the experience is ready-made. Yes really, more than a thousand copies of a statue have been measured, which I will not guarantee to be that of the gladiator, but which, in any case, is not far from it: these copies were even alive, so that the measures were taken with every possible chance of error: I would add, moreover, that the copies could have been distorted by a host of accidental causes. One must therefore expect, here, to find a very sensitive probable error.

This average man liked a lot at the time, especially within the English eugenicist school, directed by Francis Galton, even if the latter is mainly interested in deviations from this norm (upward deviation and downward deviation). As Bulmer recalls, “the deviations from that average – upwards towards genius, and downwards towards stupidity – must follow the law that governs deviations from all true averages”. Galton’s work was aimed at understanding these deviations. If Florence Nightingale stated that “the average man is God’s will”, Galton was more interested in the hereditary character of the deviation than in the average. But does that mean anything to this average man?

# Looking for the “average” person

Rose presents two examples in her book Tyrany of the NormThe End of Average . The first is drawn from problems encountered by the US military in the 1950s. When designing the cockpits of fighter aircraft, engineers had used the dimensions of more than 4,000 pilots to optimally position the seat relative to the pedals, the joystick, the height of the windscreen, but also the shape of the seat, the helmet, etc. These measurements made it possible to calculate the measurements of the “median” pilot in about ten dimensions. For example, the average pilot size was 179 cm, which allowed the average pilot size to be defined between 175 and 185 cm. While a majority of the pilots were medium in size, none of the 4,000 pilots was “average” in all dimensions. As Daniels stated, “designing a cockpit for the average pilot was in fact not designing one for any pilot.

The second example is linked to two statues, those of Norma and Normann (historically on display in Cleveland, now in the Harvard Library). The artist Abram Belskie and the obstetrician Robert Latou Dickinson made these statues together in 1943. Their particularity is that no model has been represented. In fact, it was to represent a woman and a man who had the average measurements of the time (from measurements made on thousands of subjects). Once these statues were made, a contest was held to find out who these statues could represent. Several thousand people from Ohio sent their measurements, but none matched those of the statues. Of course, several hundred were the same size. Several hundred had the same chest circumference. But none had all the right measurements. Because as Todd Rose explains, man is not unidimensional: it is on several dimensions that we measure it in several dimensions. And trying to summarize it in a one-dimensional size is far too reductive. This is what he shows in his book on intelligence tests, for example, where the same IQ can be associated with two very different people. It makes no sense to focus on a single indicator when deciding to recruit someone. The concern when working in a multivariate context is that the average loses its meaning. In fact, from a probabilistic point of view, being average can be extraordinary.

# The curse of the dimensionality

In fact, this problem is well known to statisticians as the “scourge of the dimension”. Let’s take a simple example: suppose that a quantity of interest follows a normal law $N(\mu,\sigma^2)$, for example weight, height, chest circumference, etc.. One could say that being in the norm is being in an interval $[\mu\pm1.5\sigma]$. If we have a normal law, this situation occurs in 85% of cases. And the 15% that are not in this range will be seen as “abnormal”. The sizes can be abnormally small, or abnormally large. It is the drawing of figure 1, on the top. We can now look at two dimensions, weight and height, for example. The norm here would be that in both dimensions, we are in the interval $[\mu\pm1.5\sigma]$. If the quantities are independent, the probability that both quantities are “normal” is 75%, since $0.85^2\sim 0.75$

In other words, in dimension two, 75% of the observations are globally normal, and 25% are then abnormal. In dimension 3, we pass to 65 %, that is to say more than a third of abnormal observations (on the bottom on figure 1, the red points being the abnormal points).   Figure 1 Proportion of “average” individuals in dimensions 1, 2 and 3

In dimension five, we go below 50%, in other words, being in the norm in the five dimensions is no longer the case of the majority. And in dimension twenty, those which are normal are rather atypical, with a proportion of the order of 5 %. Thus, in large dimensions, normality is no longer associated with the idea of a majority. This is the problem that actuaries face today when using very large data, in pricing for example: it becomes very difficult to characterize a rate class (by saying what the average insured in that class looks like).

# Normality, statistics and standards

From an empirical, descriptive point of view, being within the norm means nothing other than being within the average, not getting too far from that average. We will then tend to define the norm as the frequency of what happens most often, as the attitude most frequently encountered or the preference most regularly expressed. But this normality is not normativity, and “to be in the norm”, to be exemplary, is then a different dimension, which this time no longer relates to a description of reality but to an identification of what it should tend towards. So we move from the register of being to that of being, from “is” to “ought” to use Hume’s terminology. It is indeed difficult to envisage the model (or normality) without sliding towards the second meaning that can be found in the concept of standard, which in turn has a strictly normative dimension. This vision leads to confusion between norms and laws, even if not all normativity is exhausted by laws. Hume thus notes that, in all moral systems, authors move from statements of fact, that is, statements of the “there is” type, to proposals that include a normative expression, such as “one must”, “one must”. What Hume disputes is the shift from one type of statement to another: for him, these are two types of statements that have nothing to do with each other, and that cannot therefore be logically linked with each other, in particular from an empirical norm to a normative rule. For Hume, an assertion that is not normative cannot give rise to a normative conclusion. Hume’s assertion has given rise to numerous comments and interpretations, particularly because, as it stands, it seems to be an obstacle to any attempt at naturalization of morality – as McIntyre or Rescher detail. In this sense, there is a strong distinction between the norm in regularity (normality) and the rule (normativity).

# Statistical laws, from micro to macro

The statistical law is about what “is” because it has been observed (for example, “men are taller than dogs”). Human law (divine, or judicial) is what “is” because it has been decreed, and therefore “must be” (“Men are free and equal” or “Man is good”). Finally, the physical law is about what “is” because we can show it (“The planets are attracted to each other”), often within the framework of hypotheses. We see that the three concepts can be linked. For example, Kepler’s law was historically established using observations (and historically fell into the first category), before being demonstrated in the Copernican model (and then moved on to the third). A concept of balance can also be associated with this law, this “norm”. However, as Hilpinen points out, however, probabilistic laws ask many questions, one need only think of dice throws or expectations: what is meant by “it is normal to wait five minutes for the bus to stop”, or more ethically disturbing, “it is normal for a person remanded in custody to be imprisoned for eighteen months”?

The norm can be seen as a regularity of cases, observed using frequencies (or averages), for example, on the size of individuals, the duration of sleep, in other words the data that constitute the description of individuals. Anthropometric data have thus made it possible to define an average size of individuals in a given population, according to their age; compared to this average size, a difference of 20% more or less determines gigantism or dwarfism. If we think of road accidents, it can be considered abnormal to have a road accident in a given year, at an individual (micro) level, because the majority of drivers do not have an accident. Nevertheless, from the insurer’s (macro) point of view, the norm is that 10% of drivers have an accident. It would therefore be abnormal for no one to have an accident.

Nevertheless, from the insurer’s (macro) point of view, the norm is that 10% of drivers have an accident. It would therefore be abnormal for no one to have an accident. This is the argument found in Durkheim. From the singular act that is suicide, if it is considered from the point of view of the individual who commits it, Durkheim tries to see it as a social act, then a real regularity, within a given society. From then on, according to Durkheim, suicide became a normal phenomenon. Statistics then make it possible to quantify the tendency to suicide in a given society, as soon as we no longer observe the irregularity that appears in the singularity of an individual story, but a social normality of suicide.

# Standard, convention and ethical aspects

If we take an evolutionary view, what is normal is what is most capable of adapting, of responding to needs, of providing a model for the resolution of situations (nature making abnormality disappear), and normality tends towards normativity, and it becomes difficult to distinguish between the two aspects. In fact, David Hume addresses this point in the well-known example of rowers, who get into the same boat to cross a river and row in rhythm (this example is discussed at length in Mackie). The two rowers gradually adjust their rowing strokes, one in relation to the other, and it is not necessary to obtain an explicit agreement (which would formulate the standard) that they would respect. The law, which consists in imposing a standard can be useful in case of conflict (if one of the rowers refuses to row, or two rowers of very different physical capacities), but very often, it is not necessary to formulate explicitly this standard inherent to their conduct. The external observer will observe a regularity (when the cruising rhythm is reached) that he can model, but this normal observed rhythm is not necessarily imposed by a law. In the case of rowers, we find the notion of balance mentioned previously. To build a model is to extract the signal from the noise (to use Silver’s distinction), it is to look for a standard, in the statistical sense. But this goes further if a predictive model is constructed, and reality must then conform to the model, as actuaries often hope.

Patrick Blackburn, Maarten de Rijke & Yde Venema, Modal Logic, Cambridge University Press, 2002.

Bulmer M., Francis Galton: Pioneer of Heredity and Biometry. Johns Hopkins University Press, 2004.

Daniels G., “The Average Man”, Air Force Aerospace Medical Research Lab, vol. 53, n° 7, 1952.

Durkheim E., Le suicide, 1897.

Hilpinen R., Deontic Logic: Introductory and Systematic Readings, 1971, Dordrecht, D. Reidel Publishing Company.

Hume D., Traité de la nature humaine. Tome III : de la morale, 1739.

McIntyre D.C., “Hume on ‘Is’ and ‘Ought’”, The Philosophical Review, vol. 68, n° 4, 1959, pp. 451-468, Duke University Press.

Mackie J.L., Hume’s Moral Theory, Routledge & Kegan Paul Books, 1980.

Rescher N., “How Wide Is the Gap Between Facts and Values?”, Philosophy and Phenomenological Research, vol. 50, 1990, pp. 297-319.

Silver N., The Signal and the Noise: Why So Many Predictions Fail – But Some Don’t, Penguin Press, 2015.

Rose T., The End of Average: How We Succeed in a World That Values Sameness, HarperOne, 2016.

Sheynin O., “Mathematical Treatment of Astronomical Observations (A Historical Essay)”. Archive for History of Exact Sciences, vol. 11, 1973, pp. 97-126.

Stahl S., “The Evolution of the Normal Distribution”, Mathematics Magazine, vol. 79, 2006, pp. 96-113.

# p-hacking, or cheating on a p-value

Yesterday evening, I discovered some interesting slides on False-Positives, p-Hacking, Statistical Power, and Evidential Value, via  ‘s post on Twitter. More precisely, there was this slide on how cheating (because that’s basically what it is) to get a ‘good’ model (by targeting the p-value) As mentioned by @david_colquhoun  one should be careful when reading the slides : some statistician might have a heart attack when they read But still, there are interesting points in that slide.

# Bias of Hill Estimators

In the MAT8595 course, we’ve seen yesterday Hill estimator of the tail index. To be more specific, we did see see that if , with , then Hill estimators for are given by

for . Then we did say that satisfies some consistency in the sense that if , but not too fast, i.e. (under additional assumptions on the rate of convergence, it is possible to prove that ). Further, under additional technical conditions

In order to illustrate this point, consider the following code. First, let us consider a Pareto survival function, and the associated quantile function

> alpha=1.5
> S=function(x){ifelse(x>1,x^(-alpha),1)}
> Q=function(p){uniroot(function(x) S(x)-(1-p),lower=1,upper=1e+9)$root} The code here is obviously too complicated, since this power function can easily be inverted. But later on, we will consider a more complex survival function. Here are the survival function, and the quantile function, > u=seq(0,5,by=.01) > plot(u,Vectorize(S)(u),type="l",col="red") > u=seq(0,99/100,by=.01) > plot(u,Vectorize(Q)(u),type="l",col="blue",ylim=c(0,20))  Here, we need the quantile function to generate a random sample from this distribution, > n=500 > set.seed(1) > X=Vectorize(Q)(runif(n)) Hill plot is here > library(evir) > hill(X) > abline(h=alpha,col="blue") We can now generate thousands of random samples, and see how those estimators behave (for some specific ‘s). > ns=10000 > HillK=matrix(NA,ns,10) > for(s in 1:ns){ + X=Vectorize(Q)(runif(n)) + H=hill(X,plot=FALSE) + hillk=function(k) H$y[H$x==k] + HillK[s,]=Vectorize(hillk)(15*(1:10)) + } and if we compute the average, > plot(15*(1:10),apply(HillK,2,mean) we do get a series of estimators that can be considered as unbiased. So far, so good. Now, recall that being in the max-domain of attraction of the Fréchet distribution does not mean that , with , but is means that for some slowly varying function , not necessarily constant! In order to understand what could happen, we have to be slightly more specific. And this can be done only by looking at second order regular variation property of the survival function. Assume, here that there is some auxilary function such that This (positive) constant is – somehow – related to the speed of convergence of the ratio of the survival functions to the power function (see e.g. Geluk et al. (2000) for some examples). To be more specific, assume that then, the second order regular variation property is obtained using , and then, if goes to infinity too fast, then the estimator will be biased. More precisely (see Chapter 6 in Embrechts et al. (1997)), if , then, for some , The intuitive interpretation of this result is that if is too large, and if the underlying distribution is not exactly a Pareto distribution (and we do have this second order property), then Hill’s estimator is biased. This is what we mean when we say • if is too large, is a biased estimator • if is too small, is a volatile estimator (the later comes from properties of a sample mean: the more observations, the less the volatility of the mean). Let us run some simulations to get a better understanding of what’s going on. Using the previous code, it is actually extremly simple to generate a random sample with survival function > beta=.5 > S=function(x){+ ifelse(x>1,.5*x^(-alpha)*(1+x^(-beta)),1) } > Q=function(p){uniroot(function(x) S(x)-(1-p),lower=1,upper=1e+9)$root}

If we use the code above. Here, with

> n=500
> set.seed(1)
> X=Vectorize(Q)(runif(n))

the Hill plot becomes

> library(evir)
> hill(X)
> abline(h=alpha,col="blue") But it’s based on one sample, only. Again, consider thousands of samples, and let us see how Hill’s estimator is behaving, so that the (empirical) mean of those estimator is # Likelihood Based Methods, for Extremes

This week, in the MAT8595 course, we will start the section on inference for extreme values. To start with something simple, we will use maximum likelihood techniques on a Generalized Pareto Distribution (we’ve seen Monday Pickands-Balkema-de Hann theorem).

• Maximum Likelihood Estimation

In the context of parametric models, the standard technique is to consider the maximum of the likelihood (or the log-likelihod).Let denote the parameter (with ). Given some – stnardard – technical assumptions, such as , or  on some neighbourhood of , then

where denotes Fisher information matrix (see any textbook for mathematical statistics courses). Consider here some i.i.d. sample, from a Generalized Pareto Distribution, with parameter , so that

If we solve (numerically) the first order condition of the maximum likelihood, we get an estimator  which satisfies

The idea of this asymptotic normality is the following : if the true distribution of the sample is a GPD with parameter , then, if  is large enough, then will have a joint normal distribution. So if we generate a lot of sample (sufficently large, say 200 observations), then the scatterplot of the estimator should the same as the scatterplot of a Gaussian distribution,

> library(evir)
> n=200
> param=matrix(NA,1000,2)
> for(s in 1:1000){
+ x=rgpd(n,xi=1/1.5,beta=1)
+ param[s,]=gpd(x,0)$par.ests + } > m=apply(param,2,mean) > S=var(param) > library(mnormt) > x=seq(min(param[,1])-.05,max(param[,1])+.05,length=101) > y=seq(min(param[,2])-.05,max(param[,2])+.05,length=101) > vx=rep(x,each=length(y)) > vy=rep(y,length(x)) > vz=dmnorm(cbind(vx,vy),m,S) > z=matrix(vz,length(y),length(x)) > COL=rev(heat.colors(100)) > image(x,y,z,col=COL) > points(param) and to get a 3d representation > x=seq(min(param[,1])-.05,max(param[,1])+.05,length=31) > y=seq(min(param[,2])-.05,max(param[,2])+.05,length=31) > vx=rep(x,each=length(y)) > vy=rep(y,length(x)) > vz=dmnorm(cbind(vx,vy),m,S) > z=matrix(vz,length(y),length(x)) > persp(x,y,t(z),shade=TRUE,col="green",theta=-30,phi=20,ticktype="detailed", + xlab="xi",ylab="sigma")  With 200 observations, if the true underlying distribution is a GPD, then, indeed, the joint distribution of seems to be normal. That would be interesting to generate some confidence intervals for instance, or define some tests. To go further, see any standard textbook on statistical mathematics, e.g. Casella & Berger (2002). • Delta Method Another important property is the so called delta-method (we’ve seen Monday in class that it was obtained easily using a first order Taylor expansion). The idea is that if is asymptotically normal, and if is sufficently smooth, then will also be asymptotically Gaussian. More precicely (see also the header of this blog) From this property, we can get the normality of (which is another parametrization used in extreme value models), or on any quantile, . Let us run some simulation, one more time to check that we actually have a joint normality. > library(evir) > n=200 > param=riskm=matrix(NA,1000,2) > for(s in 1:1000){ + x=rgpd(n,xi=1/1.5,beta=1) + param[s,]=gpd(x,0)$par.ests
+ xihat=param[s,1]
+ shat=param[s,2]
+ q=shat * (.01^(-xihat) - 1)/xihat
+ tvar=q+(shat + xihat * q)/(1 - xihat)
+ riskm[s,]=c(1/xihat,q)
+ }
> m=apply(riskm,2,mean)
> S=var(riskm)
> library(mnormt)
> x=seq(min(riskm[,1])-.05,max(riskm[,1])+.05,length=101)
> y=seq(min(riskm[,2])-.05,max(riskm[,2])+.05,length=101)
> vx=rep(x,each=length(y))
> vy=rep(y,length(x))
> vz=dmnorm(cbind(vx,vy),m,S)
> z=matrix(vz,length(y),length(x))
> image(x,y,t(z),col=COL)
> points(riskm)

As we can see bellow, with samples of size 200, we cannot use this asymptotical result: it looks like we do not have enought data. But if we run the same code with

> n=5000  We get the joint normality of and . This is what we can get from this result, called delta-method in statistical textbooks. See again Casella & Berger (2002) for more details.

• Profile Likelihood

Another interesting tool is the concept of profile likelihood. This would be interesting here since the main interest is the tail index , being here some kind of auxilary parameter. See Venzon & Moolgavkar (1988) for more details. Here, we will plot But more generally, it is possible to consider where is the set of interesting parameters. Then (under standard suitable conditions) we can prove that which can be used to derive confidence intervals. In the GPD case, for each , we have to find an optimal . We compute the (profile) likelihood i.e. . And we can compute the maximum of this profile likelihood. This two-stage optimization is, in general, not equivalent with the (global) maximization of the likelihood, as computed below

>  n=500
>  set.seed(1)
>  x=rgpd(n,xi=1/1.5,beta=1)
>  loglikelihood=function(xi,beta){
+  sum(log(dgpd(x,xi,mu=0,beta))) }
>  XIV=(1:300)/100;L=rep(NA,300)
>  for(i in 1:300){
+  XI=XIV[i]
+  profilelikelihood=function(beta){
+  -loglikelihood(XI,beta) }
+  L[i]=-optim(par=1,fn=profilelikelihood)$value } > plot(XIV,L,type="l") > XIV[which.max(L)]  0.67 > gpd(x,0)$par.ests
xi      beta
0.6730145 0.9725483 We are not far away. Actually, if we want to compute the maximum of the profile likelihood (and not only compute the values of the profile likelihood on a grid, as before), we use

>  PL=function(XI){
+  profilelikelihood=function(beta){
+  -loglikelihood(XI,beta) }
+  return(optim(par=1,fn=profilelikelihood)$value)} > (OPT=optimize(f=PL,interval=c(0,3)))$minimum
 0.6731025

$objective  822.5574 Observe that, indeed, we are not far away from the maximum likelihood estimator of (I believe that it’s mainly a computational issue here, and theat the two are similar, here… actually, I’d be glad to hear about cases where maximum of the profile likelihood is not the same as the maximum of the likelihood). The interesting point is that we can use this technique to compute a confidence interval, and even visualize it on a graph > up=OPT$objective
>  abline(h=-up)
>  abline(h=-up-qchisq(p=.95,df=1),col="red")
>  I=which(L>=-up-qchisq(p=.95,df=1))
>  lines(XIV[I],rep(-up-qchisq(p=.95,df=1),length(I)),
+  lwd=5,col="red")
>  abline(v=range(XIV[I]),lty=2,col="red") The vertical lines are the lower and the upper bound of a 95% confidence interval for parameter .

To go further, see Murphy, S.A & van der Vaart, A.W. (2000). On Profile Likelihood.

# Pariwise and global dependence

I was just contacted by some researchers willing to test if a multivariate copula is – or not – Gaussian. They use a test proposed in an unpublished paper by Malevergne and Sornette, stating that one should simply test for pairwise normality. In this paper, the following result is mentioned In other words, it is stated that (or at least something closed to this)  Unfortunately, this result is (probably) not correct (and if it is valid, it is nontrivial). It should be possible to construct a counterexample (thanks to Roger Nelsen (web) for the idea) by letting the correlation in each pair be close to -1 (i.e. global pairwise contercomonotonicity). Then the correlation matrix of the triplet would fail to be positive definite (which is a requierment for Gaussian vectors). This idea can be related to the “compatibility” condition in Joe (1997).  My colleague Mihai Gradinaru (web) mentioned the book by Wlodzimierz Bryc, Normal distribution: characterizations with applications, but I did not find any result in it (I looked very quickly). Anyway, I am still working on this… any comment is welcome !