This notebooks explores several problems related to interpreting the results of medical tests.
Copyright 2016 Allen Downey
MIT License: http://opensource.org/licenses/MIT
In [1]:
from __future__ import print_function, division
from thinkbayes2 import Pmf, Suite
from fractions import Fraction
Suppose we test a patient to see if they have a disease, and the test comes back positive. What is the probability that the patient is actually sick (that is, has the disease)?
To answer this question, we need to know:
The prevalence of the disease in the population the patient is from. Let's assume the patient is identified as a member of a population where the known prevalence is p
.
The sensitivity of the test, s
, which is the probability of a positive test if the patient is sick.
The false positive rate of the test, t
, which is the probability of a positive test if the patient is not sick.
Given these parameters, we can compute the probability that the patient is sick, given a positive test.
To do that, I'll define a Test
class that extends Suite
, so it inherits Update
and provides Likelihood
.
The instance variables of Test
are:
p
, s
, and t
: Copies of the parameters.d
: a dictionary that maps from hypotheses to their probabilities. The hypotheses are the strings sick
and notsick
.likelihood
: a dictionary that encodes the likelihood of the possible data values pos
and neg
under the hypotheses.
In [2]:
class Test(Suite):
"""Represents beliefs about a patient based on a medical test."""
def __init__(self, p, s, t, label='Test'):
# initialize the prior probabilities
d = dict(sick=p, notsick=1-p)
super(Test, self).__init__(d, label)
# store the parameters
self.p = p
self.s = s
self.t = t
# make a nested dictionary to compute likelihoods
self.likelihood = dict(pos=dict(sick=s, notsick=t),
neg=dict(sick=1-s, notsick=1-t))
def Likelihood(self, data, hypo):
"""
data: 'pos' or 'neg'
hypo: 'sick' or 'notsick'
"""
return self.likelihood[data][hypo]
Now we can create a Test
object with parameters chosen for demonstration purposes (most medical tests are better than this!):
In [3]:
p = Fraction(1, 10) # prevalence
s = Fraction(9, 10) # sensitivity
t = Fraction(3, 10) # false positive rate
test = Test(p, s, t)
test.Print()
And here's how we update the Test
object with a positive outcome:
In [4]:
test.Update('pos')
test.Print()
The positive test provides evidence that the patient is sick, increasing the probability from 0.1 to 0.25.
t
So far, this is basic Bayesian inference. Now let's add a wrinkle. Suppose that we don't know the value of t
with certainty, but we have reason to believe that t
is either 0.2 or 0.4 with equal probability.
Again, we would like to know the probability that a patient who tests positive actually has the disease. As we did with the Red Die problem, we will consider several scenarios:
Scenario A: The patients are drawn at random from the relevant population, and the reason we are uncertain about t
is that either (1) there are two versions of the test, with different false positive rates, and we don't know which test was used, or (2) there are two groups of people, the false positive rate is different for different groups, and we don't know which group the patient is in.
Scenario B: As in Scenario A, the patients are drawn at random from the relevant population, but the reason we are uncertain about t
is that previous studies of the test have been contradictory. That is, there is only one version of the test, and we have reason to believe that t
is the same for all groups, but we are not sure what the correct value of t
is.
Scenario C: As in Scenario A, there are two versions of the test or two groups of people. But now the patients are being filtered so we only see the patients who tested positive and we don't know how many patients tested negative. For example, suppose you are a specialist and patients are only referred to you after they test positive.
Scenario D: As in Scenario B, we have reason to think that t
is the same for all patients, and as in Scenario C, we only see patients who test positive and don't know how many tested negative.
We can represent this scenario with a hierarchical model, where the levels of the hierarchy are:
t
and their probabilities.t
.To represent the hierarchy, I'll define a MetaTest
, which is a Suite
that contains Test
objects with different values of t
as hypotheses.
In [5]:
class MetaTest(Suite):
"""Represents a set of tests with different values of `t`."""
def Likelihood(self, data, hypo):
"""
data: 'pos' or 'neg'
hypo: Test object
"""
# the return value from `Update` is the total probability of the
# data for a hypothetical value of `t`
return hypo.Update(data)
To update a MetaTest
, we update each of the hypothetical Test
objects. The return value from Update
is the normalizing constant, which is the total probability of the data under the hypothesis.
We use the normalizing constants from the bottom level of the hierarchy as the likelihoods at the top level.
Here's how we create the MetaTest
for the scenario we described:
In [6]:
q = Fraction(1, 2)
t1 = Fraction(2, 10)
t2 = Fraction(4, 10)
test1 = Test(p, s, t1, 'Test(t=0.2)')
test2 = Test(p, s, t2, 'Test(t=0.4)')
metatest = MetaTest({test1:q, test2:1-q})
metatest.Print()
At the top level, there are two tests, with different values of t
. Initially, they are equally likely.
When we update the MetaTest
, it updates the embedded Test
objects and then the MetaTest
itself.
In [7]:
metatest.Update('pos')
Out[7]:
Here are the results.
In [8]:
metatest.Print()
Because a positive test is more likely if t=0.4
, the positive test is evidence in favor of the hypothesis that t=0.4
.
This MetaTest
object represents what we should believe about t
after seeing the test, as well as what we should believe about the probability that the patient is sick.
In [9]:
def MakeMixture(metapmf, label='mix'):
"""Make a mixture distribution.
Args:
metapmf: Pmf that maps from Pmfs to probs.
label: string label for the new Pmf.
Returns: Pmf object.
"""
mix = Pmf(label=label)
for pmf, p1 in metapmf.Items():
for x, p2 in pmf.Items():
mix.Incr(x, p1 * p2)
return mix
Here's the posterior predictive distribution:
In [10]:
predictive = MakeMixture(metatest)
predictive.Print()
After seeing the test, the probability that the patient is sick is 0.25, which is the same result we got with t=0.3
.
In [11]:
def MakeMetaTest(p, s, pmf_t):
"""Makes a MetaTest object with the given parameters.
p: prevalence
s: sensitivity
pmf_t: Pmf of possible values for `t`
"""
tests = {}
for t, q in pmf_t.Items():
label = 'Test(t=%s)' % str(t)
tests[Test(p, s, t, label)] = q
return MetaTest(tests)
def Marginal(metatest):
"""Extracts the marginal distribution of t.
"""
marginal = Pmf()
for test, prob in metatest.Items():
marginal[test.t] = prob
return marginal
def Conditional(metatest, t):
"""Extracts the distribution of sick/notsick conditioned on t."""
for test, prob in metatest.Items():
if test.t == t:
return test
MakeMetaTest
makes a MetaTest
object starting with a given PMF of t
.
Marginal
extracts the PMF of t
from a MetaTest
.
Conditional
takes a specified value for t
and returns the PMF of sick
and notsick
conditioned on t
.
I'll test these functions using the same parameters from above:
In [12]:
pmf_t = Pmf({t1:q, t2:1-q})
metatest = MakeMetaTest(p, s, pmf_t)
metatest.Print()
Here are the results
In [13]:
metatest = MakeMetaTest(p, s, pmf_t)
metatest.Update('pos')
metatest.Print()
Same as before. Now we can extract the posterior distribution of t
.
In [14]:
Marginal(metatest).Print()
Having seen one positive test, we are a little more inclined to believe that t=0.4
; that is, that the false positive rate for this patient/test is high.
And we can extract the conditional distributions for the patient:
In [15]:
cond1 = Conditional(metatest, t1)
cond1.Print()
In [16]:
cond2 = Conditional(metatest, t2)
cond2.Print()
Finally, we can make the posterior marginal distribution of sick/notsick, which is a weighted mixture of the conditional distributions:
In [17]:
MakeMixture(metatest).Print()
At this point we have a MetaTest
that contains our updated information about the test (the distribution of t
) and about the patient that tested positive.
Now, to compute the probability that both patients are sick, we have to know the distribution of t
for both patients. And that depends on details of the scenario.
In Scenario A, the reason we are uncertain about t
is either (1) there are two versions of the test, with different false positive rates, and we don't know which test was used, or (2) there are two groups of people, the false positive rate is different for different groups, and we don't know which group the patient is in.
So the value of t
for each patient is an independent choice from pmf_t
; that is, if we learn something about t
for one patient, that tells us nothing about t
for other patients.
So if we consider two patients who have tested positive, the MetaTest we just computed represents our belief about each of the two patients independently.
To compute the probability that both patients are sick, we can convolve the two distributions.
In [18]:
convolution = metatest + metatest
convolution.Print()
Then we can compute the posterior marginal distribution of sick/notsick for the two patients:
In [19]:
marginal = MakeMixture(metatest+metatest)
marginal.Print()
So in Scenario A the probability that both patients are sick is 1/16.
As an aside, we could have computed the marginal distributions first and then convolved them, which is computationally more efficient:
In [20]:
marginal = MakeMixture(metatest) + MakeMixture(metatest)
marginal.Print()
We can confirm that this result is correct by simulation. Here's a generator that generates random pairs of patients:
In [21]:
from random import random
def flip(p):
return random() < p
def generate_pair_A(p, s, pmf_t):
while True:
sick1, sick2 = flip(p), flip(p)
t = pmf_t.Random()
test1 = flip(s) if sick1 else flip(t)
t = pmf_t.Random()
test2 = flip(s) if sick2 else flip(t)
yield test1, test2, sick1, sick2
And here's a function that runs the simulation for a given number of iterations:
In [22]:
def run_simulation(generator, iters=100000):
pmf_t = Pmf([0.2, 0.4])
pair_iterator = generator(0.1, 0.9, pmf_t)
outcomes = Pmf()
for i in range(iters):
test1, test2, sick1, sick2 = next(pair_iterator)
if test1 and test2:
outcomes[sick1, sick2] += 1
outcomes.Normalize()
return outcomes
In [23]:
outcomes = run_simulation(generate_pair_A)
outcomes.Print()
As we increase iters
, the probablity of (True, True) converges on 1/16, which is what we got from the analysis.
Good so far!
In [24]:
metatest1 = MakeMetaTest(p, s, pmf_t)
metatest1.Update('pos')
metatest1.Print()
And the marginal distribution of sick/notsick is the same:
In [25]:
marginal = MakeMixture(metatest1)
marginal.Print()
Now suppose the second patient arrives. We need a new MetaTest
that contains the updated information about the test, but no information about the patient other than the prior probability of being sick, p
:
In [26]:
metatest2 = MakeMetaTest(p, s, Marginal(metatest1))
metatest2.Print()
Now we can update this MetaTest
with the result from the second test:
In [27]:
metatest2.Update('pos')
metatest2.Print()
This distribution contains updated information about the test, based on two positive outcomes, and updated information about a patient who has tested positive (once).
After seeing two patients with positive tests, the probability that t=0.4
has increased to 25/34, around 74%.
For either patient, the probability of being sick is given by the marginal distribution from metatest2
:
In [28]:
predictive = MakeMixture(metatest2)
predictive.Print()
After two tests, the probability that the patient is sick is slightly lower than after one (4/17 is about 23.5%, compared to 25%). That's because the second positive test increases our belief that the false positive rate is high (t=0.4), which decreases our belief that either patient is sick.
Now, to compute the probability that both are sick, we can't just convolve the posterior marginal distribution with itself, as we did in Scenario A, because the selection of t
is not independent for the two patients. Instead, we have to make a weighted mixture of conditional distributions.
If we know t=t1
, we can compute the joint distribution for the two patients:
In [29]:
cond_t1 = Conditional(metatest2, t1)
conjunction_t1 = cond_t1 + cond_t1
conjunction_t1.Print()
If we know that t=t1
, the probability of sicksick
is 0.111. And for t=t2
:
In [30]:
cond_t2 = Conditional(metatest2, t2)
conjunction_t2 = cond_t2 + cond_t2
conjunction_t2.Print()
If we know that t=t2
, the probability of sicksick
is 0.04
.
The overall probability of sicksick
is the weighted average of these probabilities:
In [31]:
posterior_t = Marginal(metatest2)
posterior_t[t1] * conjunction_t1['sicksick'] + posterior_t[t2] * conjunction_t2['sicksick']
Out[31]:
1/17
is about 5.88%, somewhat smaller than in Scenario A (1/16
, which is 6.25%).
To compute the probabilities for all four outcomes, I'll make a Metapmf
that contains the two conditional distributions.
In [32]:
metapmf = Pmf()
for t, prob in Marginal(metatest2).Items():
cond = Conditional(metatest2, t)
conjunction = cond + cond
metapmf[conjunction] = prob
metapmf.Print()
And finally we can use MakeMixture
to compute the weighted averages of the posterior probabilities:
In [33]:
predictive = MakeMixture(metapmf)
predictive.Print()
To confirm that this result is correct, I'll use the simuation again with a different generator:
In [34]:
def generate_pair_B(p, s, pmf_t):
while True:
sick1, sick2 = flip(p), flip(p)
t = pmf_t.Random()
test1 = flip(s) if sick1 else flip(t)
# Here's the difference
# t = pmf_t.Random()
test2 = flip(s) if sick2 else flip(t)
yield test1, test2, sick1, sick2
The difference between Scenario A and Scenario B is the line I commented out. In Scenario B, we generate t
once and it applies to both patients.
In [35]:
outcomes = run_simulation(generate_pair_B)
outcomes.Print()
As iters
increases, the results from the simulation converge on 1/17
.
In summary:
P(sick|pos) P(sicksick|pospos)
Scenario A 1/4 = 25% 1/16 = 6.25%
Scenario B 1/4 = 25% 1/17 ~= 5.88%
If we are only interested in one patient at a time, Scenarios A and B are the same. But for collections of patients, they yield different probabilities.
A real scenario might combine elements of A and B; that is, the false positive rate might be different for different people, and we might have some uncertainty about what it is. In that case, the most accurate probability for two patients might be anywhere between 1/16
and 1/17
.
In [ ]: