Minimizing expected loss for hypothesis testing

Loading...

En provenance du cours de Duke University

Bayesian Statistics

405 notes

Duke University

Bayesian Statistics

405 notes

Cours 4 sur 5 dans Specialization Statistics with R

This course describes Bayesian statistics, in which one's inferences about parameters or hypotheses are updated as evidence accumulates. You will learn to use Bayes’ rule to transform prior probabilities into posterior probabilities, and be introduced to the underlying theory and perspective of the Bayesian paradigm. The course will apply Bayesian methods to several practical problems, to show end-to-end Bayesian analyses that move from framing the question to building models to eliciting prior probabilities to implementing in R (free statistical software) the final posterior distribution. Additionally, the course will introduce credible regions, Bayesian comparisons of means and proportions, Bayesian regression and inference using multiple models, and discussion of Bayesian prediction. We assume learners in this course have background knowledge equivalent to what is covered in the earlier three courses in this specialization: "Introduction to Probability and Data," "Inferential Statistics," and "Linear Regression and Modeling."

À partir de la leçon

Decision Making

In this module, we will discuss Bayesian decision making, hypothesis testing, and Bayesian testing. By the end of this week, you will be able to make optimal decisions based on Bayesian statistics and compare multiple hypotheses using Bayes Factors.

Losses and decision making3:17

Working with loss functions6:34

Minimizing expected loss for hypothesis testing5:24

Posterior probabilities of hypotheses and Bayes factors6:19

Rencontrer les enseignants

Mine Çetinkaya-Rundel
Associate Professor of the Practice
Department of Statistical Science
David Banks
Professor of the Practice
Statistical Science
Colin Rundel
Assistant Professor of the Practice
Statistical Science
Merlise A Clyde
Professor
Department of Statistical Science

In Bayesian statistics,

all inference about a parameter is made base on the posterior distribution.

Including in the setting of a hypothesis test.

Suppose we have two competing hypothesis, H1 and H2.

Then the probability that H1 is true given the data

is the posterior probability of H1.

And the probability that H2 is true

given the data is the posterior probability of H2.

One straightforward way of choosing between H1 and

H2 would be to choose the one with the higher posterior probability.

In other words, you reject H1 if the posterior probability

of H1 is smaller than the posterior probability of H2.

However, since hypothesis testing is a decision problem,

we should also consider a loss function.

Revisiting the HIV testing example from earlier in the course.

Suppose we want to test H1 which is patient does not have HIV,

versus H2, patient has HIV.

Note that these are the only two possibilities, hence these are mutually

exclusive hypotheses that cover the entire decision space.

We can define the loss function as L(d)

as the loss that occurs when decision d is made.

The Bayesian testing procedure then minimizes the posterior expected loss.

The possible decisions, in other words actions,

in our two competing hypotheses example are d1.

Which is choosing H1,

in other words deciding that the patient does not have HIV.

And d2, which is choosing H2, in other words deciding that the patient has HIV.

Suppose our decision is d1.

We might be right, or we might be wrong.

If we decide that the patient does not have HIV, and indeed they don't.

Then the decision is right and the loss associated with d1 is zero,

in other words, no loss.

If we decide that the patient does not have HIV, but they actually do.

Then the decision is wrong, and the loss associated with d1 is some value w1.

Now suppose our decision is d2.

Again, we might be right or we might be wrong.

If we decide that the patient has HIV, and indeed they do,

then the decision is right.

And the loss associated with d2 is zero, in other words, no loss.

If we decide the patient has HIV but they actually do not,

then the decision is wrong and the loss associated with d2 is some value w2.

The consequences of making a wrong decision d1 or d2 are different.

If the decision is d1 and it is wrong,

then you decided that the patient does not have HIV when in reality they do.

This is a false negative, potential consequences are no treatment and

premature death.

These are grave consequences.

If the decision is d2 and it is wrong then you decided that the patient

has HIV when in reality they don’t.

This is a false positive.

Potential consequences are distrust and unnecessary further investigation.

These consequences are certainly not ideal but

they are much less grave than the consequences of a false negative decision.

Let's put these definitions in the context of the HIV testing example with ELISA.

The loss function for d1 can take on the values zero or w1 say 1000 and

the loss function for d2 can take on the values zero or ten.

The values of w1 or w2 are arbitrarily chosen.

But the important thing to realize is that w1,

the loss associated with a false negative determination.

Is much higher than w2,

the loss associated with a false positive determination.

Remember from the earlier video that our patient had tested positive on the ELISA.

In that video, we also calculated the posterior probability of the patient not

having HIV given positive ELISA result.

As approximately 0.88 and

the posterior probability of having HIV given the positive ELISA result.

As the complement of this value, approximately .12.

Then the expected loss for d1 can be calculated as the sum of the two

posterior probabilities weighted by their associated losses.

That is, the posterior possibility of the patient not having HIV.

Given positive ELISA times zero, since the loss for

deciding the patient does not have HIV would be the right decision in this case.

Times the posterior probability of having HIV given positive ELISA times 1,000,

the loss associated with a false negative.

We can similarly calculate the expected loss for d2.

Since the expected loss for d2 is lower, we should make that decision.

That is, we should decide that the patient has HIV.

Note that our decision is highly influenced by the losses we assigned to

d1 and d2.

If the losses were symmetric, say both w1 and w2 were ten.

Then the expected loss for d1 would be 1.2 and the expected loss for

d2 wouldn't change and we would choose d1 instead.

That is, we would decide that the patient does not have HIV.

To recap, Bayesian methodologies allow for

the integration of losses into the decision making framework easily.

And in Bayesian testing, we minimize the posterior expected loss.

Explorer notre catalogue

Rejoignez-nous gratuitement et obtenez des recommendations, des mises à jour et des offres personnalisées.

Coursera propose un accès universel à la meilleure formation au monde, en partenariat avec des universités et des organisations du plus haut niveau, pour proposer des cours en ligne.

© 2018 Coursera Inc. Tous droits réservés.

Télécharger dans l'App Store

Disponible sur Google Play