30. A Problem that Stumped Milton Friedman#
(and that Abraham Wald solved by inventing sequential analysis)
Contents
Co-authored with Chase Coleman
30.1. Overview#
This lecture describes a statistical decision problem encountered by Milton Friedman and W. Allen Wallis during World War II when they were analysts at the U.S. Government’s Statistical Research Group at Columbia University.
This problem led Abraham Wald [Wal47] to formulate sequential analysis, an approach to statistical decision problems intimately related to dynamic programming.
In this lecture, we apply dynamic programming algorithms to Friedman and Wallis and Wald’s problem.
Key ideas in play will be:
Bayes’ Law
Dynamic programming
Type I and type II statistical errors
a type I error occurs when you reject a null hypothesis that is true
a type II error is when you accept a null hypothesis that is false
Abraham Wald’s sequential probability ratio test
The power of a statistical test
The critical region of a statistical test
A uniformly most powerful test
30.2. Origin of the problem#
On pages 137-139 of his 1998 book Two Lucky People with Rose Friedman [FF98], Milton Friedman described a problem presented to him and Allen Wallis during World War II, when they worked at the US Government’s Statistical Research Group at Columbia University.
Let’s listen to Milton Friedman tell us what happened.
“In order to understand the story, it is necessary to have an idea of a simple statistical problem, and of the standard procedure for dealing with it. The actual problem out of which sequential analysis grew will serve. The Navy has two alternative designs (say A and B) for a projectile. It wants to determine which is superior. To do so it undertakes a series of paired firings. On each round it assigns the value 1 or 0 to A accordingly as its performance is superior or inferior to that of B and conversely 0 or 1 to B. The Navy asks the statistician how to conduct the test and how to analyze the results.
“The standard statistical answer was to specify a number of firings (say 1,000) and a pair of percentages (e.g., 53% and 47%) and tell the client that if A receives a 1 in more than 53% of the firings, it can be regarded as superior; if it receives a 1 in fewer than 47%, B can be regarded as superior; if the percentage is between 47% and 53%, neither can be so regarded.
“When Allen Wallis was discussing such a problem with (Navy) Captain Garret L. Schyler, the captain objected that such a test, to quote from Allen’s account, may prove wasteful. If a wise and seasoned ordnance officer like Schyler were on the premises, he would see after the first few thousand or even few hundred [rounds] that the experiment need not be completed either because the new method is obviously inferior or because it is obviously superior beyond what was hoped for \(\ldots\) ‘’
Friedman and Wallis struggled with the problem but, after realizing that they were not able to solve it, described the problem to Abraham Wald.
That started Wald on the path that led him to Sequential Analysis [Wal47].
We’ll formulate the problem using dynamic programming.
30.3. A dynamic programming approach#
The following presentation of the problem closely follows Dmitri Berskekas’s treatment in Dynamic Programming and Stochastic Control [Ber75].
A decision maker observes iid draws of a random variable \(z\).
He (or she) wants to know which of two probability distributions \(f_0\) or \(f_1\) governs \(z\).
After a number of draws, also to be determined, he makes a decision as to which of the distributions is generating the draws he observers.
To help formalize the problem, let \(x \in \{x_0, x_1\}\) be a hidden state that indexes the two distributions:
Before observing any outcomes, the decision maker believes that the probability that \(x = x_0\) is
After observing \(k+1\) observations \(z_k, z_{k-1}, \ldots, z_0\), he updates this value to
which is calculated recursively by applying Bayes’ law:
After observing \(z_k, z_{k-1}, \ldots, z_0\), the decision maker believes that \(z_{k+1}\) has probability distribution
This is a mixture of distributions \(f_0\) and \(f_1\), with the weight on \(f_0\) being the posterior probability that \(x = x_0\) 1.
To help illustrate this kind of distribution, let’s inspect some mixtures of beta distributions.
The density of a beta probability distribution with parameters \(a\) and \(b\) is
We’ll discretize this distribution to make it more straightforward to work with.
The next figure shows two discretized beta distributions in the top panel.
The bottom panel presents mixtures of these distributions, with various mixing probabilities \(p_k\).
using LinearAlgebra, Statistics
using Distributions, LaTeXStrings, Printf, Random, Roots, Plots
using StatsPlots
begin
base_dist = [Beta(1, 1), Beta(3, 3)]
mixed_dist = MixtureModel.(Ref(base_dist),
(p -> [p, one(p) - p]).(0.25:0.25:0.75))
plot(plot(base_dist, labels = [L"f_0" L"f_1"],
title = "Original Distributions"),
plot(mixed_dist, labels = [L"1/4-3/4" L"1/2-1/2" L"3/4-1/4"],
title = "Distribution Mixtures"),
# Global settings across both plots
ylab = "Density", ylim = (0, 2), layout = (2, 1))
end
30.3.1. Losses and costs#
After observing \(z_k, z_{k-1}, \ldots, z_0\), the decision maker chooses among three distinct actions:
He decides that \(x = x_0\) and draws no more \(z\)’s.
He decides that \(x = x_1\) and draws no more \(z\)’s.
He postpones deciding now and instead chooses to draw a \(z_{k+1}\).
Associated with these three actions, the decision maker can suffer three kinds of losses:
A loss \(L_0\) if he decides \(x = x_0\) when actually \(x=x_1\).
A loss \(L_1\) if he decides \(x = x_1\) when actually \(x=x_0\).
A cost \(c\) if he postpones deciding and chooses instead to draw another \(z\).
30.3.2. Digression on type I and type II errors#
If we regard \(x=x_0\) as a null hypothesis and \(x=x_1\) as an alternative hypothesis, then \(L_1\) and \(L_0\) are losses associated with two types of statistical errors.
a type I error is an incorrect rejection of a true null hypothesis (a “false positive”)
a type II error is a failure to reject a false null hypothesis (a “false negative”)
So when we treat \(x=x_0\) as the null hypothesis
We can think of \(L_1\) as the loss associated with a type I error.
We can think of \(L_0\) as the loss associated with a type II error.
30.3.3. Intuition#
Let’s try to guess what an optimal decision rule might look like before we go further.
Suppose at some given point in time that \(p\) is close to 1.
Then our prior beliefs and the evidence so far point strongly to \(x = x_0\).
If, on the other hand, \(p\) is close to 0, then \(x = x_1\) is strongly favored.
Finally, if \(p\) is in the middle of the interval \([0, 1]\), then we have little information in either direction.
This reasoning suggests a decision rule such as the one shown in the figure
As we’ll see, this is indeed the correct form of the decision rule.
The key problem is to determine the threshold values \(\alpha, \beta\), which will depend on the parameters listed above.
You might like to pause at this point and try to predict the impact of a parameter such as \(c\) or \(L_0\) on \(\alpha\) or \(\beta\).
30.3.4. A Bellman equation#
Let \(J(p)\) be the total loss for a decision maker with current belief \(p\) who chooses optimally.
With some thought, you will agree that \(J\) should satisfy the Bellman equation
where \(p'\) is the random variable defined by
when \(p\) is fixed and \(z\) is drawn from the current best guess, which is the distribution \(f\) defined by
In the Bellman equation, minimization is over three actions:
accept \(x_0\)
accept \(x_1\)
postpone deciding and draw again
Let
Then we can represent the Bellman equation as
where \(p \in [0,1]\).
Here
\((1-p) L_0\) is the expected loss associated with accepting \(x_0\) (i.e., the cost of making a type II error).
\(p L_1\) is the expected loss associated with accepting \(x_1\) (i.e., the cost of making a type I error).
\(c + A(p)\) is the expected cost associated with drawing one more \(z\).
The optimal decision rule is characterized by two numbers \(\alpha, \beta \in (0,1) \times (0,1)\) that satisfy
and
The optimal decision rule is then
Our aim is to compute the value function \(J\), and from it the associated cutoffs \(\alpha\) and \(\beta\).
One sensible approach is to write the three components of \(J\) that appear on the right side of the Bellman equation as separate functions.
Later, doing this will help us obey the don’t repeat yourself (DRY) golden rule of coding.
30.4. Implementation#
Let’s code this problem up and solve it.
We implement the cost functions for each choice considered in the Bellman equation (63.7).
First, consider the cost associated to accepting either distribution and compare the minimum of the two to the expected benefit of drawing again.
Drawing again will only be worthwhile if the expected marginal benefit of learning from an additional draw is greater than the explicit cost.
For every belief \(p\), we can compute the difference between accepting a distribution and choosing to draw again.
The solution \(\alpha\), \(\beta\) occurs at indifference points.
Define the cost function be the minimum of the pairwise differences in cost among the choices.
Then we can find the indifference points when the cost function is zero.
We can use any roots finding algorithm to solve for the solutions in the interval [0, 1].
Lastly, verify which indifference points correspond to the definition of a permanent transition between the accept and reject space for each choice.
Here’s the code
accept_x0(p, L0) = (one(p) - p) * L0
accept_x1(p, L1) = p * L1
function bayes_update(p, d0, d1)
p * pdf(d0, p) / pdf(MixtureModel([d0, d1], [p, one(p) - p]), p)
end
function draw_again(p, d0, d1, L0, L1, c, target)
candidate = 0.0
cost = 0.0
while candidate < target
p = bayes_update(p, d0, d1)
cost += c
candidate = min(accept_x0(p, L0), accept_x1(p, L1)) + cost
if candidate >= target
break
end
target = candidate
end
return candidate
end
function choice(p, d0, d1, L0, L1, c)
if isone(p)
output = (1, 0)
elseif iszero(p)
output = (2, 0)
elseif zero(p) < p < one(p)
target, option = findmin([accept_x0(p, L0), accept_x1(p, L1)])
candidate = draw_again(p, d0, d1, L0, L1, c, target)
if candidate < target
target, option = (candidate, 3)
end
output = (option, target)
else
throw(ArgumentError("p must be ∈ [0, 1]"))
end
return output
end
choice (generic function with 1 method)
Next we solve a problem by finding the \(\alpha\), \(\beta\) values for the decision rule
function decision_rule(d0, d1, L0, L1, c)
function cost(p, d0, d1, L0, L1, c)
if c < zero(c)
throw(ArgumentError("Cost must be non-negative"))
end
x0 = accept_x0(p, L0)
x1 = accept_x1(p, L1)
draw = draw_again(p, d0, d1, L0, L1, c, min(x0, x1))
output = min(abs(draw - x0), abs(draw - x1), abs(x1 - x0))
return output
end
# Find the indifference points
roots = find_zeros(p -> cost(p, d0, d1, L0, L1, c), 0 + eps(), 1 - eps())
# Compute the choice at both sides
left = first.(choice.(roots .- eps(), d0, d1, L0, L1, c))
right = first.(choice.(roots .+ eps(), d0, d1, L0, L1, c))
# Find beta by checking for a permanent transition from the area accepting to
# x₁ to never again accepting x₁ at the various indifference points
# Find alpha by checking for a permanent transition from the area accepting of
# x₀ to never again accepting x₀ at the various indifference points
beta = findlast((left .== 2) .& (right .≠ 2)) |>
(x -> isa(x, Int) ? roots[x] : 0)
alpha = findfirst((left .≠ 1) .& (right .== 1)) |>
(x -> isa(x, Int) ? roots[x] : 1)
if beta < alpha
@printf("Accept x1 if p <= %.2f\nContinue to draw if %.2f <= p <= %.2f
\nAccept x0 if p >= %.2f", beta, beta, alpha, alpha)
else
x0 = accept_x0(beta, L0)
x1 = accept_x1(beta, L1)
draw = draw_again(beta, d0, d1, L0, L1, c, min(x0, x1))
if draw == min(x0, x1, draw)
@printf("Accept x1 if p <= %.2f\nContinue to draw if %.2f <= p <= %.2f
\nAccept x0 if p >= %.2f", beta, beta, alpha, alpha)
else
@printf("Accept x1 if p <= %.2f\nAccept x0 if p >= %.2f", beta, alpha)
end
end
return (alpha, beta)
end
decision_rule (generic function with 1 method)
We can simulate an agent facing a problem and the outcome with the following function
function simulation(problem)
(; d0, d1, L0, L1, c, p, n, return_output) = problem
alpha, beta = decision_rule(d0, d1, L0, L1, c)
outcomes = fill(false, n)
costs = fill(0.0, n)
trials = fill(0, n)
for trial in 1:n
# Nature chooses
truth = rand(1:2)
# The true distribution and loss are defined based on the truth
d = (d0, d1)[truth]
l = (L0, L1)[truth]
t = 0
choice = 0
while iszero(choice)
t += 1
outcome = rand(d)
p = bayes_update(p, d0, d1)
if p <= beta
choice = 1
elseif p >= alpha
choice = 2
end
end
correct = choice == truth
cost = t * c + (correct ? 0 : l)
outcomes[trial] = correct
costs[trial] = cost
trials[trial] = t
end
@printf("\nCorrect: %.2f\nAverage Cost: %.2f\nAverage number of trials: %.2f",
mean(outcomes), mean(costs), mean(trials))
return return_output ? (alpha, beta, outcomes, costs, trials) : nothing
end
function Problem(;d0 = Beta(1, 1), d1 = Beta(9, 9),
L0 = 2, L1 = 2,
c = 0.2, p = 0.5,
n = 100, return_output = false)
return (; d0, d1, L0, L1, c, p, n, return_output)
end;
Random.seed!(0);
simulation(Problem());
Accept x1 if p <= 0.35
Continue to draw if 0.35 <= p <= 0.57
Accept x0 if p >= 0.57
Correct: 0.53
Average Cost: 1.22
Average number of trials: 1.40
30.4.1. Comparative statics#
Now let’s consider the following exercise.
We double the cost of drawing an additional observation.
Before you look, think about what will happen:
Will the decision maker be correct more or less often?
Will he make decisions sooner or later?
Random.seed!(0);
simulation(Problem(;c = 0.4));
Accept x1 if p <= 0.41
Continue to draw if 0.41 <= p <= 0.54
Accept x0 if p >= 0.54
Correct: 0.55
Average Cost: 1.38
Average number of trials: 1.20
Notice what happens?
The average number of trials decreased.
Increased cost per draw has induced the decision maker to decide in 0.18 less trials on average.
Because he decides with less experience, the percentage of time he is correct drops.
This leads to him having a higher expected loss when he puts equal weight on both models.
30.5. Comparison with Neyman-Pearson formulation#
For several reasons, it is useful to describe the theory underlying the test that Navy Captain G. S. Schuyler had been told to use and that led him to approach Milton Friedman and Allan Wallis to convey his conjecture that superior practical procedures existed.
Evidently, the Navy had told Captail Schuyler to use what it knew to be a state-of-the-art Neyman-Pearson test.
We’ll rely on Abraham Wald’s [Wal47] elegant summary of Neyman-Pearson theory.
For our purposes, watch for there features of the setup:
the assumption of a fixed sample size \(n\)
the application of laws of large numbers, conditioned on alternative probability models, to interpret the probabilities \(\alpha\) and \(\beta\) defined in the Neyman-Pearson theory
Recall that in the sequential analytic formulation above, that
The sample size \(n\) is not fixed but rather an object to be chosen; technically \(n\) is a random variable.
The parameters \(\beta\) and \(\alpha\) characterize cut-off rules used to determine \(n\) as a random variable.
Laws of large numbers make no appearances in the sequential construction.
In chapter 1 of Sequential Analysis [Wal47] Abraham Wald summarizes the Neyman-Pearson approach to hypothesis testing.
Wald frames the problem as making a decision about a probability distribution that is partially known.
(You have to assume that something is already known in order to state a well posed problem. Usually, something means a lot.)
By limiting what is unknown, Wald uses the following simple structure to illustrate the main ideas.
A decision maker wants to decide which of two distributions \(f_0\), \(f_1\) govern an i.i.d. random variable \(z\).
The null hypothesis \(H_0\) is the statement that \(f_0\) governs the data.
The alternative hypothesis \(H_1\) is the statement that \(f_1\) governs the data.
The problem is to devise and analyze a test of hypothesis \(H_0\) against the alternative hypothesis \(H_1\) on the basis of a sample of a fixed number \(n\) independent observations \(z_1, z_2, \ldots, z_n\) of the random variable \(z\).
To quote Abraham Wald,
A test procedure leading to the acceptance or rejection of the hypothesis in question is simply a rule specifying, for each possible sample of size \(n\), whether the hypothesis should be accepted or rejected on the basis of the sample. This may also be expressed as follows: A test procedure is simply a subdivision of the totality of all possible samples of size \(n\) into two mutually exclusive parts, say part 1 and part 2, together with the application of the rule that the hypothesis be accepted if the observed sample is contained in part 2. Part 1 is also called the critical region. Since part 2 is the totality of all samples of size 2 which are not included in part 1, part 2 is uniquely determined by part 1. Thus, choosing a test procedure is equivalent to determining a critical region.
Let’s listen to Wald longer:
As a basis for choosing among critical regions the following considerations have been advanced by Neyman and Pearson: In accepting or rejecting \(H_0\) we may commit errors of two kinds. We commit an error of the first kind if we reject \(H_0\) when it is true; we commit an error of the second kind if we accept \(H_0\) when \(H_1\) is true. After a particular critical region \(W\) has been chosen, the probability of committing an error of the first kind, as well as the probability of committing an error of the second kind is uniquely determined. The probability of committing an error of the first kind is equal to the probability, determined by the assumption that \(H_0\) is true, that the observed sample will be included in the critical region \(W\). The probability of committing an error of the second kind is equal to the probability, determined on the assumption that \(H_1\) is true, that the probability will fall outside the critical region \(W\). For any given critical region \(W\) we shall denote the probability of an error of the first kind by \(\alpha\) and the probability of an error of the second kind by \(\beta\).
Let’s listen carefully to how Wald applies a law of large numbers to interpret \(\alpha\) and \(\beta\):
The probabilities \(\alpha\) and \(\beta\) have the following important practical interpretation: Suppose that we draw a large number of samples of size \(n\). Let \(M\) be the number of such samples drawn. Suppose that for each of these \(M\) samples we reject \(H_0\) if the sample is included in \(W\) and accept \(H_0\) if the sample lies outside \(W\). In this way we make \(M\) statements of rejection or acceptance. Some of these statements will in general be wrong. If \(H_0\) is true and if \(M\) is large, the probability is nearly \(1\) (i.e., it is practically certain) that the proportion of wrong statements (i.e., the number of wrong statements divided by \(M\)) will be approximately \(\alpha\). If \(H_1\) is true, the probability is nearly \(1\) that the proportion of wrong statements will be approximately \(\beta\). Thus, we can say that in the long run [ here Wald applies a law of large numbers by driving \(M \rightarrow \infty\) (our comment, not Wald’s) ] the proportion of wrong statements will be \(\alpha\) if \(H_0\)is true and \(\beta\) if \(H_1\) is true.
The quantity \(\alpha\) is called the size of the critical region, and the quantity \(1-\beta\) is called the power of the critical region.
Wald notes that
one critical region \(W\) is more desirable than another if it has smaller values of \(\alpha\) and \(\beta\). Although either \(\alpha\) or \(\beta\) can be made arbitrarily small by a proper choice of the critical region \(W\), it is possible to make both \(\alpha\) and \(\beta\) arbitrarily small for a fixed value of \(n\), i.e., a fixed sample size.
Wald summarizes Neyman and Pearson’s setup as follows:
Neyman and Pearson show that a region consisting of all samples \((z_1, z_2, \ldots, z_n)\) which satisfy the inequality
\[ \frac{ f_1(z_1) \cdots f_1(z_n)}{f_0(z_1) \cdots f_1(z_n)} \geq k \]is a most powerful critical region for testing the hypothesis \(H_0\) against the alternative hypothesis \(H_1\). The term \(k\) on the right side is a constant chosen so that the region will have the required size \(\alpha\).
Wald goes on to discuss Neyman and Pearson’s concept of uniformly most powerful test.
Here is how Wald introduces the notion of a sequential test
A rule is given for making one of the following three decisions at any stage of the experiment (at the m th trial for each integral value of m ): (1) to accept the hypothesis H , (2) to reject the hypothesis H , (3) to continue the experiment by making an additional observation. Thus, such a test procedure is carried out sequentially. On the basis of the first observation one of the aforementioned decisions is made. If the first or second decision is made, the process is terminated. If the third decision is made, a second trial is performed. Again, on the basis of the first two observations one of the three decisions is made. If the third decision is made, a third trial is performed, and so on. The process is continued until either the first or the second decisions is made. The number n of observations required by such a test procedure is a random variable, since the value of n depends on the outcome of the observations.