## Archive for the ‘Statistics’ Category

### Ethics, logistic regression, and 0-1 loss

December 27, 2016

Andrew Gelman and David Madigan wrote a paper on why 0-1 loss is so problematic:

This is related to the issue of whether one should be training on an artificial gold standard. Suppose we have a bunch of annotators and we don’t have perfect agreement on items. What do we do? Well, in practice, machine learning evals tend to either (1) throw away the examples without agreement (e.g., the RTE evals, some biocreative named entity evals, etc.), or (2) go with the majority label (everything else I know of). Either way, we are throwing away a huge amount of information by reducing the label to artificial certainty. You can see this pretty easily with simulations, and Raykar et al. showed it with real data.

Yet 0/1 corpora and evaluation remain the gold standard (pun intended) in most machine learning evals. Kaggle has gone largely to log loss, but even that’s very flat around the middle of the range, as Andrew discusses in this blog post:

The problem is that it’s very hard to train a model that’s well calibrated if you reduce the data to an artificial gold standard. If you don’t know what I mean by calibration, check out this paper:

It’s one of my favorites. Once you’ve understood it, it’s hard not to think of evaluating models in terms of calibration and sharpness.

### Limits of Floating Point and the Asymmetry of 0 and 1

December 19, 2013

If we are representing probabilities, we are interested in numbers between 0 and 1. It turns out these have very different properties in floating-point arithmetic. And it’s not as easy to solve as just working on the log scale.

### Smallest Gaps between Floating Point Numbers of Different Magnitudes

The difference between 0 and the next biggest number representable in double-precision floating point (Java or C++ double) is on the order of 1e-300. In constrast, the difference between 1 and the next smallest number representable is around 1e-15. The reason is that to exploit the maximum number of significant digits, the mantissa for the number near 1 has to be scaled roughly like 0.999999999999999 or 1.0000000000001 and the exponent has to be scaled like 0.

### CDFs and Complementary CDFs

This is why there are two differently translated error functions in math libs, erf() and erfc(), which are rescaled cumulative and complementary cumulative distribution functions. That is, you can use erf() to calculate the cumulative unit normal distribution function (cdf), written Phi(x); erfc() can be used to calculate the complementary cumulative unit normal distribution function (ccdf), written (1 – Phi(x)).

### Log Scale no Panacea

Switching to the log scale doesn’t sove the problem. If you have a number very close to 1, you need to be careful in taking its log. If you write log(1-x), you run into the problem that x can only be so close to 1. That’s why standard math libraries provide a log1p() function defined by log1p(x) = log(1 + x). This gives you back the precision you lose by subtraction two numbers close to each other (what’s called “catastrophic cancellation” in the arithmetic processing literature).

### A Little Quiz

I’ll end with a little quiz to get you thinking about floating point a little more closely. Suppose you set the bits in a floating point number randomly, say by flipping a coin for each bit.

Junior Varsity:  What’s the approximate probability that the random floating point number has an absolute value less than 1?

Varsity:  What is the event probability of drawing a number in the range (L,U). Or, equivalently, what’s the density function to which the draws are proportional?

### Fokkens et al. on Replication Failure

September 25, 2013

After ACL, Becky Passonneau said I should read this:

You should, too. Or if you’d rather watch, there’s a video of their ACL presentation.

### The Gist

The gist of the paper is that the authors tried to reproduce some results found in the literature for word sense identification and named-entity detection, and had a rather rough go of it. In particular, they found that every little decision they made impacted the evaluation error they got, including how to evaluate the error. As the narrative of their paper unfolded, all I could do was nod along. I loved that their paper was in narrative form — it made it very easy to follow.

The authors stress that we need more information for reproducibility and that reproducing known results should get more respect. No quibble there. A secondary motivation (after the pedagogical one) of the LingPipe tutorials was to show that we could reproduce existing results in the literature.

### The Minor Fall

Although they’re riding one of my favorite hobby horses, they don’t quite lead it in the direction I would have. In fact, they stray into serious discussion of the point estimates of performance that their experiments yielded, despite the variation they see under cross-validation folds. So while they note that rankings of approaches changes based on what seem like minor details, they stop short of calling into question the whole idea of trying to assign a simple ranking. Instead, they stress getting to the bottom of why the rankings differ.

### A Slightly Different Take on the Issue

What I would’ve liked to have seen in this paper is more emphasis on two key statistical concepts:

1. the underlying sample variation problem, and
2. the overfitting problem with evaluations.

Sample Variation

The authors talk about sample variation a bit when they consider different folds, etc., in cross-validation. For reasons I don’t understand they call it “experimental setup.” But they don’t put it in terms of estimation variance. That is, independently of getting the features to all line up, the performance and hence rankings of various approaches vary to a large extent because of sample variation. The easiest way to see this is to run cross-validation.

I would have stressed that the bootstrap method should be used to get at sample variation. The bootstrap differs from cross-validation in that it samples training and test subsets with replacement. The bootstrap effectively evaluates within-sample sampling variation, even in the face of non-i.i.d. samples (usually there is correlation among the items in a corpus due to choosing multiple sentences from a single document or even multiple words from the same sentence). The picture is even worse in that the data set at hand is rarely truly representative of the intended out-of-sample application, which is typically to new text. The authors touch on these issues with discussions of “robustness.”

Overfitting

The authors don’t really address the overfitting problem, though it is tangential to what they call the “versioning” problem (when mismatched versions of corpora such as WordNet produce different results).

Overfitting is a huge problem in the current approach to evaluation in NLP and machine learning research. I don’t know why anyone cares about the umpteenth evaluation on the same fold of the Penn Treebank that produces a fraction of a percentage gain. When the variation across folds is higher than your estimated gain on a single fold, it’s time to pack the papers in the file drawer, not in the proceedings of ACL. At least until we also get a journal of negative results to go with our journal of reproducibility.

### The Major Lift

Perhaps hill-climbing error on existing data sets is not the best way forward methdologically. Maybe, just maybe, we can make the data better. It’s certainly what Breck and I tell customers who want to build an application.

If making the data better sounds appealing, check out

I find the two ideas so compelling that it’s the only area of NLP I’m actively researching. More on that later (or sooner if you want to read ahead).

(My apologies to Leonard Cohen for stealing his lyrics.)

### All Bayesian Models are Generative (in Theory)

May 23, 2013

[This post is a followup to my previous post, Generative vs. discriminative; Bayesian vs. frequentist.]

I had a brief chat with Andrew Gelman about the topic of generative vs. discriminative models. It came up when I was asking him why he didn’t like the frequentist semicolon notation for variables that are not random. He said that in Bayesian analyses, all the variables are considered random. It just turns out we can sidestep having to model the predictors (i.e., covariates or features) in a regression. Andrew explains how this works in section 14.1, subsection “Formal Bayesian justification of conditional modeling” in Bayesian Data Analysis, 2nd Edition (the 3rd Edition is now complete and will be out in the foreseeable future). I’ll repeat the argument here.

Suppose we have predictor matrix $X$ and outcome vector $y$. In regression modeling, we assume a coefficient vector $\beta$ and explicitly model the outcome likelihood $p(y|\beta,X)$ and coefficient prior $p(\beta)$. But what about $X$? If we were modeling $X$, we’d have a full likelihood $p(X,y|\beta,\psi)$, where $\psi$ are the additional parameters involved in modeling $X$ and the prior is now joint, $p(\beta,\psi)$.

So how do we justify conditional modeling as Bayesians? We simply assume that $\psi$ and $\beta$ have independent priors, so that $p(\beta,\psi) = p(\beta) \times p(\psi)$.

The posterior then neatly factors as $p(\beta,\psi|y,X) = p(\psi|X) \times p(\beta|X,y)$.

Looking at just the inference for the regression coefficients $\beta$, we have the familiar expression $p(\beta|y,X) \propto p(\beta) \times p(y|\beta,X)$.

Therefore, we can think of everything as a joint model under the hood. Regression models involve an independence assumption so we can ignore the inference for $\psi$. To quote BDA,

The practical advantage of using such a regression model is that it is much easier to specify a realistic conditional distribution of one variable given $k$ others than a joint distribution on all $k+1$ variables.

We knew that all along.

### Generative vs. Discriminative; Bayesian vs. Frequentist

April 12, 2013

[There’s now a followup post, All Bayesian models are generative (in theory).]

I was helping Boyi Xie get ready for his Ph.D. qualifying exams in computer science at Columbia and at one point I wrote the following diagram on the board to lay out the generative/discriminative and Bayesian/frequentist distinctions in what gets modeled.

To keep it in the NLP domain, let’s assume we have a simple categorization problem to predict a category z for an input consisting of a bag of words vector w and parameters β.

 Frequentist Bayesian Discriminative p(z ; w, β) p(z, β ; w) = p(z | β ; w) * p(β) Generative p(z, w ; β) p(z, w, β) = p(z, w | β) * p(β)

If you’re not familiar with frequentist notation, items to the right of the semicolon (;) are not modeled probabilistically.

Frequentists model the probability of observations given the parameters. This involves a likelihood function.

Bayesians model the joint probability of data and parameters, which, given the chain rule, amounts to a likelihood function times a prior.

Generative models provide a probabilistic model of the predictors, here the words w, and the categories z, whereas discriminative models only provide a probabilistic model of the categories z given the words w.

### Mean-Field Variational Inference Made Easy

March 25, 2013

I had the hardest time trying to understand variational inference. All of the presentations I’ve seen (MacKay, Bishop, Wikipedia, Gelman’s draft for the third edition of Bayesian Data Analysis) are deeply tied up with the details of a particular model being fit. I wanted to see the algorithm and get the big picture before being overwhelmed with multivariate exponential family gymnastics.

### Bayesian Posterior Inference

In the Bayesian setting (see my earlier post, What is Bayesian Inference?), we have a joint probability model $p(y,\theta)$ for data $y$ and parameters $\theta$, usually factored as the product of a likelihood and prior term, $p(y,\theta) = p(y|\theta) p(\theta)$. Given some observed data $y$, Bayesian predictive inference is based on the posterior density $p(\theta|y) \propto p(\theta, y)$ of the the unknown parameter vector $\theta$ given observed data vector $y$. Thus we need to be able to estimate the posterior density $p(\theta|y)$ to carry out Bayesian inference. Note that the posterior is a whole density function—we’re not just after a point estimate as in maximum likelihood estimation.

### Mean-Field Approximation

Variational inference approximates the Bayesian posterior density $p(\theta|y)$ with a (simpler) density $g(\theta|\phi)$ parameterized by some new parameters $\phi$. The mean-field form of variational inference factors the approximating density $g$ by component of $\theta = \theta_1,\ldots,\theta_J$, as $g(\theta|\phi) = \prod_{j=1}^J g_j(\theta_j|\phi_j)$.

I’m going to put off actually defining the terms $g_j$ until we see how they’re used in the variational inference algorithm.

### What Variational Inference Does

The variational inference algorithm finds the value $\phi^*$ for the parameters $\phi$ of the approximation which minimizes the Kullback-Leibler divergence of $g(\theta|\phi)$ from $p(\theta|y)$, $\phi^* = \mbox{arg min}_{\phi} \ \mbox{KL}[ g(\theta|\phi) \ || \ p(\theta|y) ]$.

The key idea here is that variational inference reduces posterior estimation to an optimization problem. Optimization is typically much faster than approaches to posterior estimation such as Markov chain Monte Carlo (MCMC).

The main disadvantage of variational inference is that the posterior is only approximated (though as MacKay points out, just about any approximation is better than a delta function at a point estimate!). In particular, variational methods systematically underestimate posterior variance because of the direction of the KL divergence that is minimized. Expectation propagation (EP) also converts posterior fitting to optimization of KL divergence, but EP uses the opposite direction of KL divergence, which leads to overestimation of posterior variance.

### Variational Inference Algorithm

Given the Bayesian model $p(y,\theta)$, observed data $y$, and functional terms $g_j$ making up the approximation of the posterior $p(\theta|y)$, the variational inference algorithm is:

• $\phi \leftarrow \mbox{random legal initialization}$
• $\mbox{repeat}$
• $\phi_{\mbox{\footnotesize old}} \leftarrow \phi$
• $\mbox{for } j \mbox{ in } 1:J$
• $\mbox{{} \ \ set } \phi_j \mbox{ such that } g_j(\theta_j|\phi_j) = \mathbb{E}_{g_{-j}}[\log p(\theta|y)]$.
• $\mbox{until } ||\phi - \phi_{\mbox{\footnotesize old}}|| < \epsilon$

The inner expectation is a function of $\theta_j$ returning a single non-negative value, defined by $\mathbb{E}_{g_{-j}}[\log p(\theta|y)]$ $\begin{array}{l} \mbox{ } = \int_{\theta_1} \ldots \int_{\theta_{j-1}} \int_{\theta_{j+1}} \ldots \int_{\theta_J} \\[8pt] \mbox{ } \hspace*{0.4in} g(\theta_1|\phi_1) \times \cdots \times g(\theta_{j-1}|\phi_{j-1}) \times g(\theta_{j+1}|\phi_{j+1}) \times \cdots \times g(\theta_J|\phi_J) \times \log p(\theta|y) \\[8pt] \mbox{ } \hspace*{0.2in} d\theta_J \cdots d\theta_{j+1} \ d\theta_{j-1} \cdots d\theta_1 \end{array}$

Despite the suggestive factorization of $g$ and the coordinate-wise nature of the algorithm, variational inference does not simply approximate the posterior marginals $p(\theta_j|y)$ independently.

### Defining the Approximating Densities

The trick is to choose the approximating factors so that the we can compute parameter values $\phi_j$ such that $g_j(\theta_j|\phi_j) = \mathbb{E}_{g_{-j}}[\log p(\theta|y)]$. Finding such approximating terms $g_j(\theta_j|\phi_j)$ given a posterior $p(\theta|y)$ is an art form unto itself. It’s much easier for models with conjugate priors. Bishop or MacKay’s books and the Wikipedia present calculations for a wide range of exponential-family models.

### What if My Model is not Conjugate?

Unfortunately, I almost never work with conjugate priors (and even if I did, I’m not much of a hand at exponential-family algebra). Therefore, the following paper just got bumped to the top of my must understand queue:

It’s great having Dave down the hall on sabbatical this year — one couldn’t ask for a better stand in for Matt Hoffman. They are both insanely good at on-the-fly explanations at the blackboard (I love that we still have real chalk and high quality boards).

### Bayesian Inference for LDA is Intractable

February 18, 2013

Bayesian inference for LDA is intractable. And I mean really really deeply intractable in a way that nobody has figured or is ever likely to figure out how to solve.

Before sending me a “but, but, but, …” reply, you might want to bone up on the technical definition of Bayesian inference, which is a bit more than applying Bayes’s rule or using a prior,

and on computational intractability, which is a bit more involved than a program being slow,

The reason Bayesian inference for LDA is intractable is that we can’t effectively integrate over its posterior. And I don’t just mean there’s no analytic solution; there rarely is for interesting Bayesian models. But for some models, you can use numerical techniques like MCMC to sample from the posterior and compute a posterior integral by averaging.

Posterior sampling doesn’t work for LDA. I know, I know, you’re going to tell me that lots of smart people have applied Gibbs sampling to LDA. You might even point out that LingPipe has a Gibbs sampler for LDA.

Yes, these systems return values. But they’re not sampling from the posterior according to the posterior distribution. The number of modes in the posterior makes comprehensive posterior sampling intractable.

And I’m not just talking about what people call “label switching”, which refers to the non-identifiability of the indexes. Label switching just means that you can permute the K indices corresponding to topics and get the same model predictions. It adds another layer of intractability, but it’s not the one that’s interesting — the real problem is multimodality.

So what does everybody do? Some use Gibbs sampling to take a single sample and then go with it. This is not a Bayesian approach; see the description of Bayesian inference above. It’s not even clear you can get a single effective sample this way (or at least it’s unclear if you can convince yourself you have an effective sample).

Others use variational inference to compute an approximate posterior mean corresponding to a single posterior mode. Variational inference approximates the posterior variance as well as the posterior mean. But alas, they don’t work that way for LDA, where there is no meaningful posterior mean due to lack of identifiability.

So what to do? The output of LDA sure does look cool. And clients do love reading the tea leaves, even if you obfuscate them in a word cloud.

I only implemented Gibbs for LDA in LingPipe because it was all that I understood when I did it. But now that I understand Bayesian inference much better, I agree with the experts: just do variational inference. You’re going to get a lousy point-based representation of the posterior no matter what you do, but variational inference is fast.

Specifically, I recommend Vowpal Wabbit package’s online LDA algorithm, which is based on Matt Hoffman’s stochastic variational LDA algorithm. It’ll also be much much more scalable than LingPipe’s sampling-based approach because it works online and doesn’t need to store all the data in memory, much like stochastic gradient descent.

The algorithm’s clean and it’d be easy enough to add such an online LDA package to LingPipe so that we could compute topic models over dozens of gigabytes of text really quickly (as Matt did with Wikipedia in the paper). But I don’t know that anyone has the time or inclination to do it.

October 30, 2012

### Bernoulli Model

Consider the following very simple model of drawing the components of a binary random N-vector y i.i.d. from a Bernoulli distribution with chance of success theta.

data {
int N;  // number of items
int y[N];  // binary outcome for item i
}
parameters {
real theta;  // Prob(y[n]=1) = theta
}
model {
theta ~ beta(2,2); // (very) weakly informative prior
for (n in 1:N)
y[n] ~ bernoulli(theta);
}


The beta distribution is used as a prior on theta. This is the Bayesian equivalent to an “add-one” prior. This is the same model Laplace used in the first full Bayesian analysis (or as some would have it, Laplacian inference) back in the Napoleonic era. He used it to model male-vs.-female birth ratio in France, with N being the number of samples and y[n] = 1 if the baby was male and 0 if female.

### Beta-Binomial Model

You can get the same inferences for theta here by replacing the Bernoulli distribution with a binomial:

model {
theta ~ beta(2,2);
sum(y) ~ binomial(N,theta);
}


But it doesn’t generalize so well. What we want to do is let the prediction of theta vary by predictors (“features” in NLP speak, covariates to some statisticians) of the items n.

### Logistic Regression Model

A roughly similar model can be had by moving to the logistic scale and replacing theta with a logistic regression with only an intercept coefficient alpha.

data {
int N;
int y[N];
}
parameters {
real alpha;  // inv_logit(alpha) = Prob(y[n]=1)
}
model {
alpha ~ normal(0,5);  // weakly informative
for (n in 1:N)
y[n] ~ bernoulli(inv_logit(alpha));
}


Recall that the logistic sigmoid (inverse of the logit, or log odds function) maps $\mbox{logit}^{-1}:(-\infty,\infty)\rightarrow(0,1)$

by taking $\mbox{logit}^{-1}(u) = 1 / (1 + \mbox{exp}(-u))$.

The priors aren’t quite the same in the Bernoulli and logistic models, but that’s no big deal. In more flexible models, we’ll move to hierarchical models on the priors.

Now that we have the inverse logit transform in place, we can replace theta with a regression on predictors for y[n]. You can think of the second model as an intercept-only regression. For instance, with a single predictor x[n], we could add a slope coefficient beta and write the following model.

data {
int N;  // number of items
int y[N];  // binary outcome for item n
real x[N];  // predictive feature for item n
}
parameters {
real alpha;  // intercept
real beta;  // slope
}
model {
alpha ~ normal(0,5);  // weakly informative
for (n in 1:N)
y[n] ~ bernoulli(inv_logit(alpha + beta * x[n]));
}


### Stan

I used Stan’s modeling language — Stan is the full Bayesian inference system I and others have been developing (it runs from the command line or from R). For more info on Stan, including a link to the manual for the modeling language, see:

Stan’s not a competitor for LingPipe, by the way. Stan scales well for full Bayesian inference, but doesn’t scale like LingPipe’s SGD-based point estimator for logistic regression. And Stan doesn’t do structured models like HMMs or CRFs and has no language-specific features like tokenizers built in. As I’ve said before, it’s horses for courses.

### High Kappa Values are not Necessary for High Quality Corpora

October 2, 2012

I’m not a big fan of kappa statistics, to say the least. I point out several problems with kappa statistics right after the initial studies in this talk on annotation modeling.

I just got back from another talk on annotation where I was ranting again about the uselessness of kappa. In particular, this blog post is an attempt to demonstrate why a high kappa is not necessary. The whole point of building annotation models a la Dawid and Skene (as applied by Snow et al. in their EMNLP paper on gather NLP data with Mechanical Turk) is that you can create a high-reliability corpus without even having high accuracy, much less acceptable kappa values — it’s the same kind of result as using boosting to combine multiple weak learners into a strong learner.

So I came up with some R code to demonstrate why a high kappa is not necessary without even bothering with generative annotation models. Specifically, I’ll show how you can wind up with a high-quality corpus even in the face of low kappa scores.

The key point is that annotator accuracy fully determines the accuracy of the resulting entries in the corpus. Chance adjustment has nothing at all to do with corpus accuracy. That’s what I mean when I say that kappa is not predictive. If I only know the annotator accuracies, I can tell you expected accuracy of entries in the corpus, but if I only know kappa, I can’t tell you anything about the accuracy of the corpus (other than that all else being equal, higher kappa is better; but that’s also true of agreement, so kappa’s not adding anything).

First, the pretty picture (the colors are in honor of my hometown baseball team, the Detroit Tigers, clinching a playoff position). What you’re looking at is a plot of the kappa value vs. annotator accuracy and category prevalence in a binary classification problem. (It’s only the upper-right corner of a larger diagram that would let accuracy run from 0 to 1 and kappa from 0 to 1. Here’s the whole plot for comparison. Note that the results are symmetric in both accuracy and prevalence, because very low accuracy leads to good agreement in the same way that very high accuracy does.)

How did I calculate the values? First, I assumed accuracy was the same for both positive and negative categories (usually not the case — most annotators are biased). Prevalence is defined as the fraction of items belonging to category 1 (usually the “positive” category).

Everything else follows from the definitions of kappa, to result in the following definition in R to compute expected kappa from binary classification data with a given prevalence of category 1 answers and a pair of annotators with the same accuracies.

kappa_fun = function(prev,acc) {
agr = acc^2 + (1 - acc)^2;
cat1 = acc * prev + (1 - acc) * (1 - prev);
e_agr = cat1^2 + (1 - cat1)^2;
return((agr - e_agr) / (1 - e_agr));
}


Just as an example, let’s look at prevalence = 0.2 and accuracy = 0.9 with say 1000 examples. The expected contingency table would be

 Cat1 Cat2 Cat1 170 90 Cat2 90 650

and the kappa coefficient would be 0.53, below anyone’s notion of “acceptable”.

The chance of actual agreement is the accuracy squared (both annotators are correct and hence agree) plus one minus the accuracy squared (both annotators are wrong and hence agree — two wrongs make a right for kappa, another of its problems).

The proportion of category 1 responses (say positive responses) is the accuracy times the prevalence (true category is positive, correct response) plus one minus accuracy times one minus prevalence (true category is negative, wrong response).

Next, I calculate expected agreement a la Cohen’s kappa (which is the same as Scott’s pi in this case because the annotators have identical behavior and hence everything’s symmetric), which is just the resulting agreement from voting according to the prevalences. So that’s just the probability of category 1 squared (both annotators respond category 1) and the probability of a category 2 response (1 minus the probability of a category 1 response) squared.

Finally, I return the kappa value itself, which is defined as usual.

Back to the plot. The white border is set at .66, the lower-end threshold established by Krippendorf for somewhat acceptable kappas; the higher-end threshold of acceptable kappas set by Krippendorf was 0.8, and is also indicated on the legend.

In my own experience, there are almost no 90% accurate annotators for natural language data. It’s just too messy. But you need well more than 90% accuracy to get into acceptable kappa range on a binary classification problem. Especially if prevalence is high, because as prevalence goes up, kappa goes down.

I hope this demonstrates why having a high kappa is not necessary.

I should add that Ron Artstein asked me after my talk what I thought would be a good thing to present if not kappa. I said basic agreement is more informative than kappa about how good the final corpus is going to be, but I want to go one step further and suggest you just inspect a contingency table. It’ll tell you not only what the agreement is, but also what each annotator’s bias is relative to the other (evidenced by asymmetric contingency tables).

In case anyone’s interested, here’s the R code I then used to generate the fancy plot:

pos = 1;
K = 200;
prevalence = rep(NA,(K + 1)^2);
accuracy = rep(NA,(K + 1)^2);
kappa = rep(NA,(K + 1)^2);
for (m in 1:(K + 1)) {
for (n in 1:(K + 1)) {
prevalence[pos] = (m - 1) / K;
accuracy[pos] = (n - 1) / K;
kappa[pos] = kappa_fun(prevalence[pos],accuracy[pos]);
pos = pos + 1;
}
}
library("ggplot2");
df = data.frame(prevalence=prevalence,
accuracy=accuracy,
kappa=kappa);
kappa_plot =
ggplot(df, aes(prevalence,accuracy,fill = kappa)) +
labs(title = "Kappas for Binary Classification\n") +
geom_tile() +
scale_x_continuous(expand=c(0,0),
breaks=c(0,0.25,0.5,0.75,1),
limits =c(0.5,1)) +
scale_y_continuous(expand=c(0,0),
breaks=seq(0,10,0.1),
limits=c(0.85,1)) +
low="orange", mid="white", high="blue",
breaks=c(1,0.8,0.66,0));


### More on the Terminology of Averages, Means, and Expectations

June 21, 2012

I missed Cosma Shalizi’s comment on my first post on averages versus means. Rather than write a blog-length reply, I’m pulling it out into its own little lexicographic excursion. To borrow stylistically from Cosma’s blog, I’ll warn you, the reader, that this post is more linguistics than statistics.

### Three Concepts and Terminology

Presumably everyone’s clear on the distinctions among the three concepts,

1. [arithmetic] sample mean,

2. the mean of a distribution, and

3. the expectation of a random variable.

The relations among these concepts is very rich, which is what I conjecture is causing their conflation.

Let’s set aside the discussion of “average”, as it’s less precise terminologically. But even the very precision of the term “average” is debatable! The Random House Dictionary lists the meaning of “average” in statistics as the arithmetic mean. Wikipedia, on the other hand, agrees with Cosma, and lists a number of statistics of centrality (sample mean, sample median, etc.) as being candidates for the meaning of “average”.

### Distinguishing Means and Sample Means

Getting to the references Cosma asked for, all of my intro textbooks (Ash, Feller, Degroot and Schervish, Larsen and Marx) distinguished sense (1) from senses (2) and (3). Even the Wikipedia entry for "mean" leads off with

In statistics, mean has two related meanings:

the arithmetic mean (and is distinguished from the geometric mean or harmonic mean).

the expected value of a random variable, which is also called the population mean.

Unfortunately, the invocation of the population mean here is problematic. Random variables aren’t intrinsically related to populations in any way (at least under the Bayesian conception of what can be modeled as random). Populations can be characterized by a set of (conditionally) independent and identically distributed (i.i.d.) random variables, each corresponding to a measureable quantity of a member of the population. And of course, averages of random variables are themselves random variables.

This reminds me to the common typological mistake of talking about “sampling from a random variable” (follow the link for Google hits for the phrase).

### Population Means and Empirical Distributions

The Wikipedia introduces a fourth concept, population mean, which is just the arithmetic mean of a given population. This is related to the point Cosma brought up in his comment that you can think of a sample mean as the mean of a distribution with the same distribution as the empirically observed distribution. For instance, if you observe three heads and a tail in three coin flips, you create a discrete random variable $X$ with $p_X(1) = 0.75$ and $p_X(0) = 0.25$, then the average number of heads is equal to the expectation of $X$ or the mean of $p_X$.

### Conflating Means and Expectations

I was surprised that like the Wikipedia, almost all the sources I consulted explicitly conflated senses (2) and (3). Feller’s 1950 Introduction to Probability Theory and Applications, Vol 1 says the following on page 221.

The terms mean, average, and mathematical expectation are synonymous. We also speak of the mean of a distribution instead of referring to a corresponding random variable.

The second sentence is telling. Distributions have means independently of whether we’re talking about a random variable or not. If one forbids talk of distributions as first-class objects with their own existence free of random variables, one might argue that concepts (2) and (3) should always be conflated.

### Metonomy and Lexical Semantic Coercion

I think the short story about what’s going on in conflating (2) and (3) is metonymy. For example, I can use “New York” to refer to the New York Yankees or the city government, but no one will understand you if you try to use “New York Yankees” to refer to the city or the government. I’m taking one aspect of the team, namely its location, and using that to refer to the whole.

This can happen implicitly with other kinds of associations. I can talk about the “mean” of a random variable $X$ by implicitly invoking its probability function $p_X(x)$. I can also talk about the expectation of a distribution by implicitly invoking the appropriate random variable. Sometimes authors try to sidestep random variable notation by writing $\mathbb{E}_{\pi(x)}[x]$, which to my type-sensitive mind appears ill-formed; what they really mean to write is $\mathbb{E}[X]$ where $p_X(x) = \pi(x)$.

I found it painfully difficult to learn statistics because of this sloppiness with respect to the types of things, especially among less careful authors. Bayesians, myself now included, often write $x$ for both a random variable and a bound variable; see Gelman et al.’s Bayesian Data Analysis, for a typical example of this style.

Settling down into my linguistic armchair, I’ll conclude by noting that it seems to me more felicitous to say

the expectation of a random variable is the mean of its distribution.

than to say

the mean of a random variable is the expectation of its distribution.