We must abandon the p-value

Just a couple of weeks ago I wrote here that we should not abandon statistical significance. No, I did not change my mind. We should not abandon statistical significance but we must abandon the p-values.

Executive summary: p-values make a false impression that statistical significance is a continuous concept, but it is not. It is dichotomous.

You are welcome to read on.

I will not get into a lengthy discussion on the difference between Fisher‘s significance testing and Neyman and Pearson‘s hypothesis testing. In short: Fisher had a null hypothesis, a test statistic and a p-value. The decision whether or not to reject the null hypothesis based on the p-value was flexible. Indeed, in one place Fisher suggested to use the 5% threshold for deciding significance, but in other places he used different thresholds.

Neyman and Pearson’s approach is totally different. In their framework there are two hypotheses: the null hypothesis and the alternative hypothesis. The Neyman-Pearson lemma provides as a tool to construct decision rule. Given data, you evaluate the likelihood of these data in two cases: once assuming that the null hypothesis is true, and then assuming that the alternative hypothesis is true. You reject the null hypothesis in favor of the alternative if the ratio of the two likelihoods exceeds a pre-defined threshold. This threshold is determined by the significance level: an acceptable probability of falsely rejecting the null hypothesis (in a frequentist manner), that meets your scientific standards.

When I was a young under-graduate student, we did not have easy access to computers in my university (yes, I am that old). So we calculated the test statistic (Z, t, F, chi-square) and looked in a table similar to this one to see if our test statistic exceeds the threshold.

Fisher was a genius, and so was Karl Pearson. They developed statistical tests such as F and Chi-square based on geometrical considerations. But when you try to construct a test for equality of means as is in ANOVA or a test for independence of two categorical variables by using the NP lemma, you get the same tests. This opened a computational shortcut: The NP decision rule is equivalent to a p-value decision rule. You can decide to reject the null by comparing your test statistic to the value you look up in a table, or by comparing the p-value to your pre-determined significance level. You will get the same result either way. But still, you either reject the null hypothesis or you don’t. The size of the test statistic or the p-value does not matter. There is are such things as “nearly significant”, “almost significant”, “borderline significant” or “very significant”. IF your pre-determined significance level is 0.05, and your p-value is 0.051, that’s really depressing. But if you declare that your result is “almost significant”, it means that you changed your decision rule after the fact.

Let me illustrate it with a story my teacher and mentor, Professor Zvi Gilula, told me once.

Someone robbed a convenience store in the middle of nowhere. The police caught him and brought him to court. There were three eye witnesses, and a video from the security camera. The jury found him guilty and he was sentenced to serve 6 months in jail. In statistical jargon, the null hypothesis states that the defendant is not guilty, and the jury rejected the null hypothesis.

Let’s look at another robber. This man robbed a convenience store in New York City, next to Yankee Stadium.  The robbery occurred when the game just ended, the Yankees won, 47000 happy fans witnessed the robbery, and they all came to testify in court, because the poor robber was a Red Sox fan. The jury found him guilty.

Is the second robber guiltier than the first robber? Is the first robber “almost not guilty” or “borderline guilty”? Is the second robber “very guilty”? What would be the right punishment to the second robber? 10 years? Life imprisonment? Death sentence? Do you think now that if you were in the jury of the first robber you would acquit him?

Here is what we need to do if we really want to improve science: when publishing results, one should provide all the necessary information, such as data, summary statistics, test statistics and more. And then state: “we reject the null hypothesis at the 5% significance level”, or maybe “we do not reject the null hypothesis at the 5% significance level” (or any other reasonable and justified significance level. No p-values. And if someone wants to know if the null hypothesis will be rejected at the 5.01% significance level, they can go and calculate their own p value. P values should be abandoned.

We should not abandon statistical significance

The call to ban statistical significance is not new. Several researchers, the most notable of them is Andrew Gellman from Columbia University, are promoting this idea in the last few years. What is new is the recent article in the prestigious journal Nature, that brought up the issue again. Actually, it is a petition signed by about 800 researchers, some of them are statisticians (and yes, Gellman signed this petition too). Gellman and Blake McShane from Northwestern University, who is one of the lead authors of the current petition, also expressed this idea  in another provocative article at Nature titled “Five ways to fix statistics” published in November 2017.

I think that it is not a good idea. It is not throwing the baby out with the bathwater. It is attempting to take the baby and throw it out of the sixth-floor window.

Let me make it clear: I do not oppose taking into account prior knowledge, plausibility of mechanism, study design and data quality, real-world costs and benefits, and other factors (I actually quoted here what Gellman and McShane wrote in the 2017 article). I think that when editors are considering a paper that was submitted to publication, they must take all of this into account, provided that the results are statistically significant.

“Abandoning statistical significance” are just nice words for abandoning the Neyman-Pearson Lemma. The lemma guarantees that likelihood ratio tests control the rate of the false positive results at the acceptable significance level, which is currently equal to 5%. The rate is guaranteed in a frequentist manner. I do not know how many studies are published every year. Maybe tens of thousands, maybe even more. If all the researchers are applying the Neyman-Pearson theory with a significance level of 5%, then the proportion of the false positive results approaches 0.05. No wonder that Gellman, who is a Bayesian statistician, opposes this.

Once statistical significance is abandoned, bad things will happen. First, the rate of false positive published results will rise. On the other hand, there is no guarantee that the rate of false negative results (that are not published anyway) will fall. The Neyman-Pearson Lemma provides protection against the false negative rates as well. Granted, you need to have an appropriate sample size to control it. But if you don’t care for statistical significance anymore, you will not care for controlling the false negative rate. Researchers will get an incentive for performing many small sample studies. Such studies are less expensive, and have higher variation. The chance of getting a nice and publishable effect is larger. This phenomenon is known as “The law of small numbers“. What about external validity and reproducibility? In a world of “publish or perish”, nobody seems to care.

And this brings us to the question how one would determine if an effect is publishable. I can think of two possible mechanisms. One possibility is to calculate a p-value and apply a “soft” threshold. P is 0.06? that’s fine. What about 0.1? it depends. 0.7? no way (I hope). A second possibility is the creation of arbitrary rules of thumbs that have no theoretical basis. In this paper, for example, the researchers expressed the effects in terms of Hedges’ g, and considered an effect as meaningful if g>1.3. Why 1.3? They did not provide any justification.

I worked many years in the pharmaceutical industry, and now I work in a research institute affiliated with a large healthcare organization. False positive results, as well as false negative results, can lead to serious consequences. Forget about the money involved. It can be a matter of life and death. Therefore, I care for the error rates, and strongly support hypothesis testing and insisting on statistical significance. And you should care too.

 

 

 

 

 

 

 

 

 

 

Cancer clusters and the Poisson distributions

On March 1, 2019, an article was published in Israel’s Ynetnews website, under the title “The curious case of the concentration of cancer”. The story reports on a concentration of cancer cases in the town of Rosh Ha’ayin in central Israel.

In the past few years dozens of cases of cancer have been discovered in the center of Rosh Ha’ayin. About 40 people have already died of the disease. Residents are sure that the cause of the disease is cellular antennas on the roof of a building belonging to the municipality. “Years we cry and no one listens”, They say, “People die one after the other.”

I do not underestimate the pain of the residents of Rosh Ha’ayin. I also do not intend to discuss the numbers mentioned in the article. I accept them as they are. I just want to relate only to the claim that the cause of the disease is cellular antennas. It is easy (at least for me) to explain why this claim is at least questionable: there are many more cellular antennas in many places, and around them there is no high rate of morbidity in cancer. If the antennas are carcinogenic, then they need cancer everywhere, not just in Rosh Ha’ayin. On the other hand, I can’t blame them for blaming the antennas. People try to rationalize what they see and find a cause.

I also must emphasize that the case of Rosh Ha’ayin must not be neglected. If it is not because the cellular antennas, it is possible that there is some other risk factor, and the state authorities must investigate.

So why does Rosh Ha’ayin have such a large cluster of cancer morbidity? One possible answer is that there is an environmental factor (other than the antennas) that does not exist elsewhere. Another possible answer is that there may be a non-environmental factor that does not exist elsewhere, maybe a genetic factor (most of the town’s residents are immigrants from Yemen and their descendants). A third and particularly sad possibility is that the local residents suffer from really bad luck. Statistics rules can be cruel.

Clusters happen: if there are no local (environmental or other) risk factors that cause cancer (or any other disease), and the disease spreads randomly across the whole country, then clusters are formed. If the country is divided into units of equal size, then the number of cases in a given area unit follows a Poisson distribution. Then there is a small, but not negligible possibility that one of these units will contain a large number of cases. The problem is that there is no way of knowing in advance where it will happen.

The opposite is also true: If the distribution of the number of cases in a given area unit is a Poisson distribution, then it can be concluded that the dispersion on the surface is random.

I will demonstrate the phenomenon using simulation.

Consider a hypothetical country that has a perfect square shape, and its size is 100 x 100 kilometers. I randomized 400 cases of morbidity across the country:

# generate n random points on 0:100 x 0:100
> set.seed(21534)
> n=400
> x=runif(n)*100
> y=runif(n)*100
> dat=data.frame(x,y)
> head(dat)
x y
1 15.73088 8.480265
2 12.77018 78.652808
3 45.50406 31.316797
4 86.46181 6.669138
5 27.25488 48.164316
6 17.42388 98.429575

I plotted the 400 cases.

> # plot the points
> plot(dat$x, dat$y, ylim=c(0,100), xlim=c(0,100), 
+ asp=1, frame.plot=FALSE, axes=FALSE,
+ xlab=' ', ylab=' ', col='aquamarine', pch=16,
+ las=1, xaxs="i", yaxs="i",
+ )
> axis(1, at=0:10*10)
> axis(2, at=0:10*10, las=1, pos=0)
> axis(3, at=0:10*10, labels=FALSE)
> axis(4, at=0:10*10, labels=FALSE, pos=100)

Here’s the map I got.

Next, I divided the map into 100 squares, each 10 x 10 kilometers:

> #draw gridlines
> for (j in 1:9){
+ lines(c(0,100), j*c(10,10))
+ lines(j*c(10,10), c(0,100))
+ }
>

In order to count the cases in each square, I assigned row and column numbers for each of the squares, and recorded the position (row and column) for every case/dot:

> # row and column numbers and cases positions
> dat$row=0
> dat$col=0
> dat$pos=0
> for (j in 1:nrow(dat)){
+ dat$row=ceiling(dat$y/10)
+ dat$col=ceiling(dat$x/10)
+ dat$pos=10*(dat$row-1)+dat$col
+ }
>

Now I can count the number of points/cases in each square:

> # calculate number of points for each position
> # ppp=points per position
> dat$count=1
> ppp=aggregate(count~pos, dat, sum)
> dat=dat[,-6]

But of course, it is possible that there are squares with zero cases; (actually the data frame ppp has only 97 rows). Let’s identify them:

> # add positions with zero counts, if any
> npp=nrow(ppp)
> if(npp<100){
+ w=which(!(1:100 %in% ppp$pos))
+ addrows=(npp+1):(npp+length(w))
+ ppp[addrows,1]=w
+ ppp[addrows,2]=0
+ ppp=ppp[order(ppp$pos),]
+ }
>

And now we can get the distribution of number of cases in each of the 100 squares:

> # distribution of number of points/cases in each position
> tb=table(ppp$count)
> print(tb)
0  1  2  3  4  5  6  7  8  9 11 
3  9 12 21 15 17 13  5  1  3  1 
>

We see that there is one very unlucky cluster with 11 cases, and there also 3 squares with 9 cases each. Let’s paint them on the map:

> # identify largest cluster
> mx=max(ppp$count)
> loc=which(ppp$count==11)
> clusters=dat[dat$pos %in% loc,]
> points(clusters$x, clusters$y, col='red', pch=16)
> 
> # identify second lasrgest cluster/s
> loc=which(ppp$count==9)
> clusters=dat[dat$pos %in% loc,]
> points(clusters$x, clusters$y, col='blue', pch=16)
>

Let’s also mark the squares with zero points/cases. In order to do this, we first need to identify the row and column locations of these squares:

> # identify sqaures without cases
> # find row and column locations
> loc=which(ppp$count==0)
> zeroes=data.frame(loc)
> zeroes$row=ceiling(zeroes$loc/10)
> zeroes$col=zeroes$loc %% 10
> w=which(zeroes$col==0)
> if(length(w)>0){
+ zeroes$col[w]=10
+ }
> print(zeroes)
loc row col
1 8 1 8
2 31 4 1
3 99 10 9
>

So there is one empty square in the 8th column of the first row, one in the first column of the 4th row, and one in the 9th column of the 10th row. Let’s mark them. To do that, we need to know the coordinates of each of the four vertices of these squares:

# mark squares with zero cases
> for (j in 1:nrow(zeroes)){
+ h1=(zeroes$col[j]-1)*10
+ h2=h1+10
+ v1=(zeroes$row[j]-1)*10
+ v2=v1+10
+ lines(c(h1,h2), c(v1,v1), lwd=3, col='purple')
+ lines(c(h1,h2), c(v2,v2), lwd=3, col='purple')
+ lines(c(h1,h1), c(v1,v2), lwd=3, col='purple')
+ lines(c(h2,h2), c(v1,v2), lwd=3, col='purple')
+ }

Do you see any pattern?

How well does the data fit the Poisson distribution? We can perform a goodness of fit test.
Let’s do the log-likelihood chi-square test (also known as the G-test):

> # log likelihood chi square to test the goodness of fit 
> # of the poisson distribution to the data
> 
> # the obserevd data
> observed=as.numeric(tb)
> values=as.numeric(names(tb))
> 
> # estimate the poisson distribution parameter lambda
> # it is the mean number of cases per square
> lambda=nrow(dat)/100
> print(lambda)
[1] 4
> 
> # calculate the expected values according to 
> # a poisson distribution with mean lambda
> expected=100*dpois(values, lambda)
> 
> # view the data for the chi-square test
> poisson_data=data.frame(values, observed, expected)
> print(poisson_data)
values observed expected
1 0 3 1.8315639
2 1 9 7.3262556
3 2 12 14.6525111
4 3 21 19.5366815
5 4 15 19.5366815
6 5 17 15.6293452
7 6 13 10.4195635
8 7 5 5.9540363
9 8 1 2.9770181
10 9 3 1.3231192
11 11 1 0.1924537
> 
> # calculate the degrees of freedom
> df=max(values)
> print(df)
[1] 11
> 
> # calculate the test statistic and p-value
> g2=sum(observed*log(observed/expected))
> pvalue=1-pchisq(g2,df)
> log_likelihood_chi_squrae_test=data.frame(g2, df, pvalue)
> print(log_likelihood_chi_squrae_test)
g2 df pvalue
1 4.934042 11 0.9343187
>

We cannot reject the hypothesis that the data follows the Poison distribution. This does not imply, of course, that the data follows the Poisson distribution, but we can say that the Poisson model fits the data well.

The delta method and its implementation in R

Suppose that you have a sample of a variable of interest, e.g. the heights of men in certain population, and for some obscured reason you are interest not in the mean height μ but in its square μ². How would you inference on μ², e.g. test a hypothesis or calculate a confidnce interval? The delta method is the trick you need.

The delta method is mathematical assertion that can yield estimates for the varinance of functons of statistics under mild condition. Loosley speaking, let bₙ is an estimate of β, where n is the sample size, and the distribution of bₙ is approximately normal with mean β and variance σ²/n.

Then, if g is a function, then g(bₙ) is approximately normal with mean g(β) and and variance [g’(β)² ]⋅σ²/n, provided that the sample size is large. Don’t let the calculus scare you. R’s deltmethod function will handle the calculus for you.

If bₙ is a Maximum Likelihood Estimate (MLE) of β then under mild conditions it is approximately normal for a large sample size, and g(bₙ) and g’(bₙ) are MLEs for g(β) and g’(β) respectively. Add on top of this a MLE for σ, and you can implement statistical inference.

I will demonstrate the use of the delta method using the Titanic survival data. For the sake of the example I will use only 3 variables: Survived is a 0/1 variable indicating whther a passenger survived the sinking of the Titanic, Age is obviously the passenger’s age, and Parch is the passenger’s number of parents/children aboard. Lets take a look at some of the data:

> # install.packages("titanic")
> library(titanic)
> titanic=titanic_train[, c(1, 2, 6,8)]
> head(titanic)
  PassengerId Survived Age Parch
1           1        0  22     0
2           2        1  38     0
3           3        1  26     0
4           4        1  35     0
5           5        0  35     0
6           6        0  NA     0
>

Let π be the probability of surviving. Then the odds of survival is π/(1-π). The sample size is n=891 is considered large, so we can apply the Central Limit Theorem to conclude that p is approximately normal with mean π and variance π/(1-π)/n. Then π can be estimated by p=0.384, the odds are estimated by p/(1-p)=0.623, and the variance of p is estimated by 0.000265:

> p_survival=mean(titanic$Survived)
> print(p_survival)
[1] 0.3838384
> survival_odds=p_survival/(1-p_survival)
> print(survival_odds)
[1] 0.6229508
> n=nrow(titanic)
> var_p_survival=(p_survival*(1-p_survival))/n
> print(var_p_survival)
[1] 0.0002654394
>

Let g(x)=p/(1-x). Then g`(x)=1/(1-x)², (you can check this at https://www.wolframalpha.com) so the variane of the odds can be estimated by

Var(p/(1-p))=[g’(p)² ]⋅σ²/n=[1/(1-p)²]² ⋅ [p(1-p)/n].

This can be implemented in the following R code:

> dev_g_odds=1/(1-survival_odds)^2
> var_odds=(dev_g_odds^2)*var_p_survival
> print(var_odds)
[1] 0.01313328
> se_odds=sqrt(var_odds)
> print(se_odds)
[1] 0.1146005
>

But of course, instead of doing all the calculus, you can use the deltamethod function of R’s msm package.

The function has three parameters:

• g is a formula object representating of the transformation g(x). The formula variables must be labeled x1, x2 and so on.
• mean is the estimate of g(β)
• cov is the estimate of var(β) which is σ²/n

The use of the delta function provides the same result:

> # install.packages("msm")
> library(msm)
> se_odds_delta=deltamethod(g=~x1/(1-x1), mean=survival_odds, cov=var_p_survival)
> print(se_odds_delta)
[1] 0.1146005
>

The second example considers logistic regression. We will model the (logit of the) probability of survival using Age nd Parch. Using R’s glm function we can obtain estimates to the logistic regression coefficients band their standard errors se_b:

> model=glm(Survived ~ Age + Parch, family=binomial(link="logit")
> model_beta=data.frame(summary(model)$coefficients[2:3, 1:2])
> names(model_beta)=c("b", "se_b")
> print(model_beta)
                 b        se_b
Age   -0.008792681 0.005421838
Parch  0.191531303 0.090643635
>

Since exp(β) is usually interpreted as the odd ratio (OR) of the response variable with regard to the regression independent variable, reseachers are interested in inference on exp(β). In order to perform such inference one nees to estimate the standard error of exp(β). Since b is a Maximum Likelihood Estimate for β, it is approximately normal and its variance is estimated by se_b², and the delta method can be applied.

The calculus in this case is easy: g(x)=exp(x), so g’(x)=exp(x). Therefore, the standard error of exp(β) can be estimated by exp(b)⋅ se_b:

> model_beta$OR=exp(model_beta$b)
> model_beta$se_OR=exp(model_beta$b)*model_beta$se_b
> print(model_beta)
                 b        se_b        OR       se_OR
Age   -0.008792681 0.005421838 0.9912459 0.005374374
Parch  0.191531303 0.090643635 1.2111027 0.109778755
>

To use the deltamethod function, we will first use the vcov function to obtain the variance-covariance matrix of the resression coefficient estimates, and the variances will be the inputs of the deltamethod function:

> vc=vcov(model)[2:3, 2:3]
> print(vc)
               Age        Parch
Age   2.939632e-05 9.236876e-05
Parch 9.236876e-05 8.216269e-03

> model_beta$se_OR_delta[1]=deltamethod(~exp(x1), mean=model_beta$b[1], cov=vc[1,1])

> model_beta$se_OR_delta[2]=deltamethod(~exp(x1), mean=model_beta$b[2], cov=vc[2,2])

> print(model_beta)
                 b        se_b        OR       se_OR se_OR_delta
Age   -0.008792681 0.005421838 0.9912459 0.005374374 0.005374374
Parch  0.191531303 0.090643635 1.2111027 0.109778755 0.109778755
>

Of course, we get the same results.

 

What is logistic in the logistic regression?

Suppose that you are interviewed for a data scientist role. You are asked about logistic regression, and you answer all sorts of questions: How to run it in Python, how would you perform feature selection, and how would you use it for prediction. For the last question you answer that if you have the estimated of the regression coefficients and the data of the features, then you perform the necessary multiplications and additions, and the result will be L=log(p/(1-p)) where p is the probability of the event to be predicted. This transformation is known as the logit transformation. From this you can calculate p as exp(L)/(1+exp(L)). Then comes the critical question: why is that?

One possible answer is that since p is between 0 and 1, then L is between -∞ to ∞, that is, it can be any real number, and therefore the logistic regression is transformed into “regular” linear regression. However, this answer is wrong. In the train data, the values of the event/label to be predicted or classified are either 0 or 1. You cannot apply the logit transformations to zeros and ones. And even if you could, the linear regression assumptions do not hold.

A more sophisticated answer is to say that the logit transformation is the link function of choice. This choice has a nice property: if β is the coefficient of a feature X, then exp(β) is a (biased) estimate for the corresponding odds ratio. This is useful if you want to identify risk factors , e.g. for a disease like cancer. But if you are interested in predictions, you may not care about the odds ratio (although you should).

Let’s assume that we know to explain what a link function is. The question still remains: Why not choose another link function? Almost any inverse distribution function will do this trick. Why not choose the inverse of the normal distribution function as the link function?

Moreover, the history does not support this answer. The logit transformation and the logistic regression model came first. This model was developed by Sir David Cox in the 1960’s — the same David Cox who later introduced the proportional hazards model (see https://papers.tinbergen.nl/02119.pdf) . The extension of the model to general linear models using various link functions came later.

So the question remains: what is logistic in the logistic regression?

The key is in the statistical model of the logistic regression, or any other binary regression. Let’s review the model.

Suppose you have data of a response/label/outcome Y that takes values of zeroes and ones. Let assume for the sake of simplicity that you have only one feature/predictor X, which can be any type of variable.

The key assumption of the model is that there exists a continuous/latent/unobservable Y* that relates somehow to the observed values of Y. Note that Y* is not a part of your data. It is a part of your model.

The next assumption is about the relationship between Y and Y*. You assume that Y equals to 1 if the signal of Y* is above some threshold, and otherwise Y is equal to zero. Furthermore, assume, without loss of generality, that this threshold is zero.

This model assumption is not new, and most of the readers are familiar with this approach. This is how the perceptron, the building block of neural networks, works. The idea itself is much older. Karl Pearson used similar modelling when he attempted to develop a correlation coefficient for categorical data, back in the 1910’s.

Assuming you know the values of Y*, you can model the relationship between Y* and X using simple linear regression:

The third and last assumption is about the distribution of the errorepsilon. As I said before, you can choose any distribution you like. If you assume, for example, that epsilon is normally distributed, you will get something that is called probit regression. But if assume that epsilon follows the logistic distribution:

then these three assumptions and some basic probability and algebra you get a logistic regression — a regression with a logit link function.

For simplicity I will assume that X is discrete variable. One can do the whole trick for any X by using density functions for a continuous X.

Let

be the conditional distribution of Y given that X is equal to some value x.

Since Y=1 if and only if Y*>0, we get that

By the second assumption of the model we get that

Using the third assumption that states that the distribution of Y* is logistic we get that

and therefore

Good luck in your interview!