Conditional probability

18.600 Problem Set 3, due February 28

Welcome to your third problem set! Conditional probability is defined by P (A|B) = P (AB)/P (B)
which implies
P (B)P (A|B) = P (AB) = P (A)P (B|A),
and dividing both sides by P (B) gives Bayes’ rule:
P (B|A)
P (A|B) = P (A) ,
P (B)
which we may view as either a boring tautology or (after spending a few hours online reading about
Bayesian epistemology, Bayesian statistics, etc.) the universal recipe for revising a worldview in response
to new information. Bayes’ rule relates P (A) (our Bayesian prior) to P (A|B) (our Bayesian posterior for
A, once B is given). If we embrace the idea that our brains have subjective probabilities for everything
(existence of aliens, next year’s interest rates, Sunday’s football scores) we can imagine that our minds
continually use Bayes’ rule to update these numbers. Or least that they would if we were clever enough
to process all the data coming our way.
By way of illustration, here’s a fanciful example. Imagine that in a certain world, a normal person
says 105 things per year, each of which has a 10−5 chance (independently of all others) of being truly
horrible. A truly horrible person says 105 things, each of which has a 10−2 chance (independently of all
others) of being truly horrible. Ten percent of the people in this world are truly horrible. Suppose we
meet someone on the bus and the first thing that person says is truly horrible. Using Bayes’ rule, we
conclude that this is probably a truly horrible person.
Then we turn on cable news and see an unfamiliar politician saying something truly horrible. Now
we’re less confident. We don’t know how the quote was selected. Perhaps the politician has made 105
recorded statements and we are seeing the only truly horrible one. So we make the quote selection
mechanism part of our sample space and do a more complex calculation.
The problem of selectively released information appears in many contexts. For example, lawyers
select evidence to influence how judges and jurors calculate conditional probability given that evidence.
If I’m trying to convince you that a number you don’t know (but which I know to be 49) is prime, I
could give you some selective information about the number without telling you exactly what it is (it’s a
positive integer, not a multiple of 2 or 3 or 5, less than 50) and if you don’t consider my motives, you’ll
say “It’s probably prime.”
Note also that legal systems around the world designate various “burdens of proof” including prob-
able cause, reasonable suspicion, reasonable doubt, beyond a shadow of a doubt, clear and convincing
evidence, some credible evidence, and reasonable to believe. Usually, these terms lack clear meaning as
numerical probabilities (does “beyond reasonable doubt” mean with probability at least .95, or at least
.99, or something else?) but there is an exception: preponderance of evidence generally indicates that a
probability is greater than fifty percent, so that something can be said to be “more likely than not.” An
interesting question (which I am not qualified to answer) is whether numerical probabilities should be
assigned to the other terms as well.

1
A. On Interrogation Planet, there are 730 suspects, and it is known that exactly one of them is guilty of
a crime. It is also known that any time you ask a guilty person a question, that person will give a
“suspicious-sounding” answer with probability .9 and a “normal-sounding” answer with probability .1.
Similarly, any time you ask an innocent person a question, that person will give a suspicious-sounding
answer with probability .1 and a normal-sounding answer with probability .9. (And these probabilities
apply regardless of how the suspect has answered questions in the past; in other words, once a person’s
guilt or innocence is fixed, that person’s answers are independent from one question to the next.)
Interrogators pick a suspect at random (all 730 people being equally likely) and ask that person nine
questions. The first three answers sound normal but the next six answers all sound suspicious. The
interrogators say “Wow, six suspicious answers in a row. Only a one in a million chance we’d see that
from an innocent person. This person is obviously guilty.” But you want to do some more thinking.
Given the answers thus far, compute the conditional probability that the suspect is guilty. Give an
exact numerical answer.
B. Every year University X has 30,000 applicants and a target class size of 1,000. Each student
(independently of the others) has a 10 percent chance of being a would-enroll-if-accepted (WEIA)
student — i.e., a student who would enroll at University X if an offer were made. The university cannot
tell who the WEIA students are. So it accepts N = 10, 000 students. Each has a .1 chance of being
WEIA. (This implies that the “expected enrollment” is .1N = 1, 000.)
Then one year, a consultant develops an “80 percent accurate” algorithm to identify WEIA students. If
a student is WEIA, the algorithm identifies the student as such with probability .8. If the student is
not WEIA the algorithm declares the student to be WEIA with probability .2. More formally, if E is
the event that a given student is WEIA and A is the event that the algorithm identifies the student as
WEIA, then P (E) = .1 and P (A|E) = .8 and P (A|E c ) = .2.
1. Compute P (A). That is, find the probability the algorithm will identity the student as WEIA.

2. Compute P (E|A). That is, find the conditional probability that a student is WEIA given that the
algorithm identified the student as WEIA.

3. If the university decides to accept only students who the algorithm identifies as WEIA, and it
accepts N such students, then the “expected enrollment” is N p where p = P (E|A). What value
of N ensures that N p = 1, 000?
Remark: A university that accepts only students the algorithm identifies as WEIA can achieve its
target class size with a much lower acceptance rate and a much higher yield rate. But is that a good
thing? Is it fair to the students the algorithm misidentifies? The above story is not intended to be
realistic, but it hints at real-world debate about the use of enrollment likelihood models. Real-world
algorithms may output a range of probabilities (rather than a binary “WEIA or not”) and also account
for financial aid. Google enrollment management and aid algorithms for some strong opinions.

2
C. In population genetics, one may model the matrilineal descendant P tree in a population beginning
with a single female (often called “Eve”).1 Fix p0 , p1 , p2 , . . . with ∞
j=0 pj = 1. Assume that for each
j ≥ 0, the probability that Eve has exactly j daughters is pj . Write Ek for the event that PEve has a
∞
“kth daughter” (i.e., that Eve has at least k daughters). Then P (Ek ) = ak where ak := j=k pj .
1. Let R = ∞
P P∞
j=0 jpj denote the expected number of children Eve will have. Check that R = 1 ak .
More generally, let Ek1 ,k2 ,...,km denote the event that Eve has a k1 th daughter and that daughter has a
k2 th daughter and so on until the mth generation m (i.e., distance m from Eve in the family tree).
Assume that given that a female is born, her probability of having at least
k children is the same as
Eve’s was—in other words, assume that P Ek1 ,k2 ,...,km−1 ,km |Ek1 ,k2 ,...,km−1 = akm .


2. Use the above equation and induction on m to show that P Ek1 ,k2 ,...,km = ak1 ak2 · · · akm .
3. Show that Eve expects to have R2 granddaughters and R3 greatgranddaughters—and generally
Rm descendants in generation m.
4. Conclude that if R < 1 then the matrilineal descendant tree is almost surely finite, and the
expected number of females in the tree (counting Eve herself) is 1 + R + R2 + . . . = 1/(1 − R).
5. Imagine that after her own birth, Eve lives an integer number of years, and the probability that
she lives at least k years is bk (where bk decreases to zero as k → ∞ since Eve is mortal). Assume
that given that she lives at least k years, she expects to have dk daughters (and tkP
children total)
during her kth year. Argue that her overall expected number of daughters is R = ∞ k=1 bk dk .
Remark:
P Each year, demographers observe age-specific death/fertility rates and work out what R and
T = tk would be if these age-specific rates were the same every year. The resulting R value is called
the net reproduction rate while T is called the total fertility rate (it includes male and female births
and does not account for deaths). Replacement-level fertility is R = 1, which corresponds to T ≈ 2.1 (in
developed countries where early death is less common) or T ≈ 2.3 (globally). In times/places with R
well above 1, leaders may try to reduce fertility. In times/places with R well below 1, leaders may try
to boost fertility. The global R value is currently close to 1 (good news for sustainability) but fertility
inequality between nations is high, with almost a factor of eight separating e.g. DR Congo and South
Korea. Migration from high to low fertility countries is politically limited (net between ±.3%/year for
most countries) and high fertility has boosted Africa recently. UN forecasters expect regional fertility
gaps to narrow tremendously over this century. (Stein’s law?) But forecasters admit great uncertainty,
despite some aspects of demographic forecasting being easier than (say) stock market forecasting—e.g.
to bound the number of Japanese-born 50-year-olds in 2075, just count the Japanese births in 2025.
Remark: If there are K founders and the total future population (including founders) is N ≫ K, then
there must be N − K future births, so the average female has (N − K)/N ≈ 1 daughters. So if the
number of descendants is large but finite, then the R > 1 and R < 1 times/places have to balance out.
1
The same math applies to patrilineal descendant trees. The foundational work by Galton and Watson in this area
considered the spread (and possible extinction) of paternally inherited aristocratic surnames in Great Britain in the 1870’s.

3
D. Suppose that a fair coin is tossed infinitely many times, independently. Let Xi denote the outcome
of the ith coin toss (an element of {H, T }). Compute:
1. the conditional probability the first toss is H given that exactly 7 of the first 10 tosses are H.

2. the probability that the pattern THTH appears at least once in the sequence X1 , X2 , X3 , . . .

3. the probability that TTTT appears in the sequence X1 , X2 , X3 , . . . before HH appears.

4. the conditional probability that the first two tosses are both heads given that exactly 6 of the first
12 tosses are heads. Is this number greater or smaller than 1/4?

5. the probability that every finite-length pattern appears infinitely many times in the sequence
X1 , X2 , X3 , . . .

E. Suppose that the quantities P [A|X1 ], P [A|X2 ], . . . , P [A|Xk ] are all equal. Check that P [Xi |A] is
proportional to P [Xi ]. In other words, check that the ratio P [Xi |A]/P [Xi ] does not depend on i. (This
requires no assumptions about whether the Xi are mutually exclusive.)
Remark: This can be viewed as a mathematical version of Occam’s razor. We view A as an
“observed” event and each Xi as an event that might “explain” A. What we showed is that if each Xi
“explains” A equally well (i.e., P (A|Xi ) doesn’t depend on i) then the conditional probability of Xi
given A is proportional to how likely Xi was a a priori. For example, suppose A is the event that there
are certain noises in my attice, X1 is the event that there are squirrels there, and X2 is the event that
there are noisy ghosts. I might say that P (X1 |A) >> P (X2 |A) because P (X1 ) >> P (X2 ). Note that
after looking up online definitions of “Occam’s razor” you might conclude that it refers to the above
tautology plus the common sense rule of thumb that P (X1 ) > P (X2 ) when X1 is “simpler” than X2 or
“requires fewer assumptions.”
F. On Cautious Science Planet, science is done as follows. First, a team of wise and well informed
experts concocts a hypothesis. Experience suggests the hypotheses produced this way are correct ninety
percent of the time, so we write P (H) = .9 where H is the event that the hypothesis is true. Before
releasing these hypotheses to the public, scientists do an additional experimental test (such as a clinical
trial or a lab study). They decide in advance what constitutes a “positive” outcome to the experiment.
Let T be the event that the positive outcome occurs. The test is constructed so that P (T |H) = .95 but
P (T |H c ) = .05. The result is only announced to the public if the test is positive. (Sometimes the test
involves checking whether an empirically observed quantity is “statistically significant.” The quantity
P (T |H) is sometimes called the power of the test.)
(a) Compute P (H|T ). This tells us what fraction of published findings we expect to be correct.

(b) On Cautious Science Planet, results have to be replicated before they are used in practice. If the
first test is positive, a second test is done. Write T̃ for the event that the second test is positive,

4
 and assume the second test is like the first test, so that P (T̃ |HT ) = .95 but P (T̃ |H c T ) = .05.
Compute the reproducibility rate P (T̃ |T ).

(c) Compute P (H|T T̃ ). This tells us how reliable the replicated results are. (Pretty reliable, it turns
out—your answer should be close to 1.)

On Speculative Science Planet, science is done as follows. First creative experts think of a hypothesis
that would be rather surprising and interesting if true. These hypotheses are correct only five percent
of the time, so we write P (H) = .05. Then they conduct a test. This time P (T |H) = .8 (lower power)
but again P (T |H c ) = .05. Using these new parameters:

(d) Compute P (H|T ).

(e) Compute the reproducibility rate P (T̃ |T ). Assume the second test is like the first test, so that
P (T̃ |HT ) = .8 but P (T̃ |H c T ) = .05.

Remark: If you google Nosek reproducibility you can learn about one attempt to systematically
reproduce 100 psychology studies, which succeeded a bit less than 40 perent of the time. Note that
P (T̃ |T ) ≈ .4 is (for better or worse) closer to Speculative Science Planet than Cautious Science Planet.
The possibility that P (H|T ) < 1/2 for real world science was famously discussed in a paper called Why
Most Published Research Findings Are False by Ioannidis in 2005. A more recent mass replication
attempt (involving just Science and Nature) allowed scientists to bet on whether a study would be
replicated and found that to some extent scientists were good at predicting such things. See
[Link] Another study found that even when
using the same data set and asking the same question, researchers may arrive at very different
conclusions because they formalize the question and analyze the data in different ways.
Questions for thought: What are the pros and cons of the two planets? Is it necessarily bad for
P (T̃ |T ) and P (H|T ) to be low in some contexts (assuming that people know this and don’t put too
much trust in single studies)? Do we need to do larger and more careful studies? What improvements
can be made in fields like medicine, where controlled clinical data is sparse and expensive but life and
death decisions have to be made nonetheless? And I do mean expensive. The cost of recruiting and
pre-screening a single Alzheimer’s patient for trial is $100,000, per this article
[Link] These
questions go well beyond the scope of this course, but we will say a bit more about the tradeoffs
involved when we study the central limit theorem.
G. Doomsday: Many people think it is likely that intelligent alien civilizations exist somewhere
(though perhaps so far separated from us in space in time that we will never encounter them). When a
species becomes roughly as advanced and intelligent as our own, how long does it typically survive
before extinction? A few thousand years? A few millions years? A few billion years? Closely related
question: how many members of such a species typically get to exist before it goes extinct?

5
Let’s consider a related problem. Suppose that one factory has produced 10 million baseball cards in
100,000 batches of 100. Each batch is numbered from 1 to 100. Another factory has produced 10
million baseball card in 10,000 batches of 1,000, each batch numbered from 1 to 1,000. A third factory
produced a 10 million baseball card in 1000 batches of 10,000, with each batch numbered from one to
10,000. You chance upon a baseball card from one of these three factories, and a priori you think it is
equally likely to come from each of the three factories. Then you notice that the number on it is 87.

(a) Given the number you have seen, what is the conditional probability that the card comes from
the first factory? The second? The third?

Now consider the following as a variant of the card problem. Suppose that one universe contains 1050
intelligent beings, grouped into civilizations of size 1012 each. Another universe contains 1050 intelligent
beings, grouped into civiliations of size 1016 each. A final universe contains 1050 intelligent beings,
grouped into civilizations of size 1020 each. You pick a random one of these 3 × 1050 beings and learn
that before this being was born, exactly 141, 452, 234, 521 other beings were born in its civilization.

(b) What is the conditional probability that the being comes from the first universe?

Remark: The doomsday argument is that it is relatively likely that the number of future humans will
not be too large, for the following reason: if advanced civilizations typically lasted for tens of millions of
years (with perhaps 108 beings born per year—and over 1015 beings total), then it would seem very
coincidental for us to find ourselves among the first 1011 or so. People disagree on what to make of this
argument (what the Bayesian prior on civilization length should be, what to do with all the other
information we have about our world, what measure to put on the set of alternative universes, etc.) By
some accounts, typical mammalian species tend to last for millions of years. Should one a priori expect
the duration of a species like ours to be longer (we are uniquely resourceful) or shorter (we have the
potential to be uniquely self-destructive)? In any case, the doomsday argument may at least help us
consider the possibility of near-term human extinction, and whether preparing for apocalyptic scenarios
(giant asteroids, incurable plagues, nuclear war, climate disaster, supervolcanos, resource depletion,
fertility collapse, famine, the next ice age, etc.) might improve our chance of surviving long enough for
another ten billion (hundred billion? trillion? quadrillion?) more humans to enjoy this world.