Most interesting statistical paradoxes Because I find them fascinating, I'd like to hear what folks in this community find as the most interesting statistical paradox and why.
 A: Parrondo's Paradox:
From wikipdedia: "Parrondo's paradox, a paradox in game theory, has been described as: A combination of losing strategies becomes a winning strategy. It is named after its creator, Juan Parrondo, who discovered the paradox in 1996. A more explanatory description is:

There exist pairs of games, each with a higher probability of losing
  than winning, for which it is possible to construct a winning strategy
  by playing the games alternately.

Parrondo devised the paradox in connection with his analysis of the Brownian ratchet, a thought experiment about a machine that can purportedly extract energy from random heat motions popularized by physicist Richard Feynman. However, the paradox disappears when rigorously analyzed."
As alluring as the paradox might sound to the financial crowd, it does have requirements that are not readily available in financial time series.  Even though a few of the component strategies can be losing, the offsetting strategies require unequal and stable probabilities of much greater or less than 50% in order for the ratcheting effect to kick in.
It would be difficult to find financial strategies, whereby one has $P_B(W)=3/4+\epsilon$ and the other, $P_A(W)=1/10 + \epsilon$, over long periods.
There's also a more recent related paradox called the "allison mixture," that shows we can take two IID and non-correlated series, and randomly scramble them such that certain mixtures can create a resulting series with non-zero autocorrelation.
A: I like the following: The host is using an unknown distribution on $[0,1]$ to choose, independently, two numbers $x,y\in [0,1]$. The only thing known to the player about the distribution is that $P(x=y)=0$. The player is then shown the number $x$ and is asked to guess whether $y>x$ or $y<x$. Clearly, if player always guesses $y>x$ then player will be correct with probability $0.5$. However, at least surprisingly if not paradoxically, player can improve on that strategy. I'm afraid I don't have a link to the problem (I heard it many years ago during a workshop).
A: It's interesting that the Two Child Problem and the Monty Hall Problem so often get mentioned together in the context of paradox. Both illustrate an apparent paradox first illustrated in 1889, called Bertrand's Box Paradox, which can be generalized to represent either. I find it a most interesting "paradox" because the same very-educated, very-intelligent people answer those two problems in opposite ways with respect to this paradox. It also compares to a principle used in card games like bridge, known as the Principle of Restricted Choice, where it resolution is time-tested.
Say you have a randomly selected item that I'll call a "box." Every possible box has at least one of two symmetric properties, but some have both. I'll call the properties "gold" and "silver." The probability that a box is just gold is P; and since the properties are symmetric, P is also the probability that a box is just silver. That makes the probability that a box has just one property 2P, and the probability that it has both 1-2P.
If you are told a box is gold, but not whether it is silver, you might be tempted to say the chances it is just gold are P/(P+(1-2P))=P/(1-P). But then you would have to state the same probability for a one-color box if you were told it was silver. And if this probability is P/(1-P) whenever you are told just one color, it has to be P/(1-P) even if you aren't told a color. Yet we know it is 2P from the last paragraph.
This apparent paradox is resolved by noting that if a box has only one color, there is no ambiguity about what color you will be told. But if it has two, there is an implied choice. You have to know how that choice was made in order to answer the question, and that is the root of the apparent paradox. If you aren't told, you can only assume a color was chosen at random, making the answer P/(P+(1-2P)/2)=2P. If you insist P/(1-P) is the answer, you are implicitly assuming there was no possibility the other color could have been mentioned unless it was the only color.
In the Monty Hall Problem, the analogy for the colors is not very intuitive, but P=1/3. Answers based on the two unopened doors originally being equally likely to have the prize are assuming Monty Hall was required to open the door he did, even if he had a choice. That answer is P/(1-P)=1/2. The answer allowing him to choose at random is 2P=2/3 for the probability that switching will win.
In the Two Child Problem, the colors in my analogy compare quite nicely to genders. With four cases, P=1/4. To answer the question, we need to know how it was determined that there was a girl in the family. If it was possible to learn about a boy in the family by that method, then the answer is 2P=1/2, not P/(1-P)=1/3. It's a little more complicated if you consider the name Florida, or "born on Tuesday," but the results are the same. The answer is exactly 1/2 if there was a choice, and most statements of the problem imply such a choice. And the reason "changing" from 1/3 to 13/27, or from 1/3 to "nearly 1/2," seems paradoxical and unintuitive, is because the assumption of no choice is unintuitive.
In the Principle of Restricted Choice, say you are missing some set of equivalent cards - like the Jack, Queen, and King of the same suit. The chances start out even that any particular card belongs to a specific opponent. But after an opponent plays one, his chances of having any one of the others are decreased because he could have played that card if he had it.
A: Another example is the ecological fallacy. 
Example 
Suppose that we look for a relationship between voting and income by regressing the vote share for then-Senator Obama on the median income of a state (in thousands). We get an intercept of approximately 20 and a slope coefficient of 0.61.
Many would interpret this result as saying that higher income people are more likely to vote for Democrats; indeed, popular press books have made this argument. 
But wait, I thought that rich people were more likely to be Republicans? They are.
What this regression is really telling us is that rich states are more likely to vote for a Democrat and poor states are more likely to vote for a Republican. Within a given state, rich people are more likely to vote Republican and poor people are more likely to vote Democrat. See the work of Andrew Gelman and his coauthors.
Without further assumptions, we cannot use group-level (aggregate) data to make inferences about individual-level behavior. This is the ecological fallacy. Group-level data can only tell us about group-level behavior.
To make the leap to individual-level inferences, we need the constancy assumption. Here, the voting choice of individuals most not vary systematically with the median income of a state; a person who earns \$X in a rich state must be just as likely to vote for a Democrat as someone who earns \$X in a poor state. But people in Connecticut, at all income levels, are more likely to vote for a Democrat than people in Mississippi at those same income levels. Hence, the consistency assumption is violated and we are led to the wrong conclusion (fooled by aggregation bias).
This topic was a frequent hobbyhorse of the late David Freedman; see this paper, for example. In that paper, Freedman provides a means for bounding individual-level probabilities using group data.
Comparison to Simpson's paradox
Elsewhere in this CW, @Michelle proposes Simpson's paradox as a good example, as it indeed is. Simpson's paradox and the ecological fallacy are closely related, yet distinct. The two examples differ in the natures of the data given and analysis used. 
The standard formulation of Simpson's paradox is a two-way table. In our example here, suppose that we have individual data and we classify each individual as high or low income. We would get an income-by-vote 2x2 contingency table of the totals. We'd see that a higher share of high income people voted for the Democrat relative to the share of low income people. Were we to create a contingency table for each state, however, we'd see the opposite pattern.
In the ecological fallacy, we don't collapse income into a dichotomous (or perhaps multichotomous) variable. To get state-level, we get the mean (or median) state income and state vote share and run a regression and find that higher income states are more likely to vote for the Democrat. If we kept the individual-level data and ran the regression separately by state, we'd find the opposite effect.
In summary, the differences are:


*

*Mode of analysis: We could say, following our SAT prep skills, that Simpson's paradox is to contingency tables as the ecological fallacy is to correlation coefficients and regression. 

*Degree of aggregation/nature of data: Whereas the Simpson's paradox example compares two numbers (Democrat vote share among high income individuals versus the same for low income individuals), ecological fallacy uses 50 data points (i.e., each state) to calculate a correlation coefficient. To get the full story from in the Simpson's paradox example, we'd just need the two numbers from each of the fifty states (100 numbers), while in the ecological fallacy case, we need the individual-level data (or else be given state-level correlations/regression slopes).


General observation
@NeilG comments that this just seems to be saying that you can't have any selection on unobservables/omitted variables bias issues in your regression. That's right! At least in the regression context, I think that nearly any "paradox" is just a special case of omitted variables bias. 
Selection bias (see my other response on this CW) can be controlled for by including the variables that drive the selection. Of course, these variables are typically unobserved, driving the problem/paradox. Spurious regression (my other other response) can be overcome by adding a time trend. These cases say, essentially, that you have enough data, but need more predictors.
In the case of the ecological fallacy, it's true, you need more predictors (here, state-specific slopes and intercepts). But you need more observations, individual-, rather than group-level, observations as well to estimate these relationships.
(Incidentally, if you have extreme selection where the selection variable perfectly divides treatment and control, as in the WWII example that I give, you may need more data to estimate the regression as well; there, the downed planes.)
A: My contribution is Simpson's paradox because:

*

*the reasons for the paradox are not intuitive to many people, so


*it can be really hard to explain why the findings are the way they are to lay people in plain English.
tl;dr version of the paradox: the statistical significance of a result appears to differ depending on how the data are partitioned. The cause appears often to be due to a confounding variable.
Another good outline of the paradox is here.
A: There are no paradoxes in statistics, only puzzles waiting to be solved.
Nevertheless, my favourite is the two envelope "paradox". Suppose I put two envelopes in front of you and tell you that one contains twice as much money as the other (but not which is which). You reason as follows. Suppose the left envelope contains $x$, then with 50% probability the right envelope contains $2x$ and with 50% probability it contains $0.5x$, for an expected value of $1.25x$. But of course you can simply reverse the envelopes and conclude instead the left envelope contains $1.25$ times the value of the right envelope. What happened?
A: The Sleeping Beauty Problem.
This is a recent invention; it was heavily discussed within a small set of philosophy journals over the last decade.  There are staunch advocates for two very different answers (the "Halfers" and "Thirders").  It raises questions about the nature of belief, probability, and conditioning, and has caused people to invoke a quantum-mechanical "many worlds" interpretation (among other bizarre things).
Here is the statement from Wikipedia:

Sleeping Beauty volunteers to undergo the following experiment and is
told all of the following details. On Sunday she is put to sleep. A
fair coin is then tossed to determine which experimental procedure is
undertaken. If the coin comes up heads, Beauty is awakened and
interviewed on Monday, and then the experiment ends. If the coin comes
up tails, she is awakened and interviewed on Monday and Tuesday. But
when she is put to sleep again on Monday, she is given a dose of an
amnesia-inducing drug that ensures she cannot remember her previous
awakening. In this case, the experiment ends after she is interviewed
on Tuesday.
Any time Sleeping beauty is awakened and interviewed, she is asked,
"What is your credence now for the proposition that the coin landed
heads?"

The Thirder position is that S.B. should respond "1/3" (this is a simple Bayes' Theorem calculation) and the Halfer position is that she should say "1/2" (because that's the correct probability for a fair coin, obviously!).  IMHO, the entire debate rests on a limited understanding of probability, but isn't that the whole point of exploring apparent paradoxes?

(Illustration from Project Gutenberg.)

Although this is not the place to try to resolve paradoxes--only to state them--I don't want to leave people hanging and I'm sure most readers of this page don't want to wade through the philosophical explanations.  We can take a tip from E. T. Jaynes, who replaces the question “how can we build a mathematical model of human common sense”—which is something we need in order to think through the Sleeping Beauty problem—by “How could we build a machine which would carry out useful plausible reasoning, following clearly defined principles expressing an idealized common sense?”  Thus, if you like, replace S. B. by Jaynes' thinking robot.  You can clone this robot (instead of administering a fanciful amnesiac drug) for the Tuesday portion of the experiment, thereby creating a clear model of the S. B. setup that can be unambiguously analyzed.  Modeling this in a standard way using statistical decision theory then reveals there are really two questions being asked here (what is the chance a fair coin lands heads? and what is the chance the coin has landed heads, conditional on the fact that you were the clone who was awakened?).  The answer is either 1/2 (in the first case) or 1/3 (in the second, using Bayes' Theorem).  No quantum mechanical principles were involved in this solution :-).

References
Arntzenius, Frank (2002).  Reflections on Sleeping Beauty.  Analysis 62.1 pp 53-62.
Elga, Adam (2000).  Self-locating belief and the Sleeping Beauty Problem.  Analysis 60 pp 143-7.
Franceschi, Paul (2005).  Sleeping Beauty and the Problem of World Reduction.  Preprint.
Groisman, Berry (2007).  The end of Sleeping Beauty’s nightmare.
Lewis, D (2001).  Sleeping Beauty: reply to Elga.  Analysis 61.3 pp 171-6.
Papineau, David and Victor Dura-Vila (2008).  A thirder and an Everettian: a reply to Lewis’s ‘Quantum Sleeping Beauty’.
Pust, Joel (2008).  Horgan on Sleeping Beauty.  Synthese 160 pp 97-101.
Vineberg, Susan (undated, perhaps 2003).  Beauty’s Cautionary Tale.
All can be found (or at least were found several years ago) on the Web.
A: This is Simpson's Paradox again but 'backwards' as well as forwards, comes from Judea Pearl's new book Causal Inference in Statistics: A primer[^1]
The classic Simpon's Paradox works as follows: consider trying to choose between two doctors. You automatically choose the one with the best outcomes. But suppose the one with the best outcomes chooses the easiest cases. The other's poorer record is a consequence of trickier work. 
Now who do you choose? Better to look at the results stratified by difficulty and then decide.
There is another side to the coin (another paradox) which says that the stratified outcomes can also lead you to the wrong choice. 
This time consider choosing to use a drug or not. The drug has a toxic side effect, but its therapeutic mechanism of action is through lowering blood pressure. Overall, the drug improves outcomes in the population, but when stratifying on post-treatment blood pressure the outcomes are worse in both the low and the high blood pressure groups. How can this be true? Because we have unintentionally stratified on the outcome, and within each outcome all that remains to observe is the toxic side effect.
To clarify, imagine the drug is designed to fix broken hearts, and it does this by lowering the blood pressure, and instead of stratifying on blood pressure we stratify on fixed hearts. When the drug works, the heart is fixed (and the blood pressure will be lower), but some of the patients will also get the toxic side effect. Because the drug works, the 'fixed heart' group will have more patients who have taken the drug, than there are patients taking the drug in the 'broken' heart group. More patients taking the drug means more patients getting side effects, and apparently (but falsely) better outcomes for patients who didn't take the drug.
The patients who get better without taking the drug are just lucky. The patients who took the drug and got better are a mixture of those who needed the drug to get better, and those who would have been lucky anyway. Examining only patients with 'fixed hearts' means excluding patients who would have been fixed had they taken the drug. Excluding such patients means excluding the harm from not taking the drug which in turn means we only see the harm from taking the drug. 
Simpson's paradox arises when there is a cause  for the outcome other than the treatment such as the fact that your doctor only does tricky cases. Controlling for the common cause (tricky versus easy cases) allows us to see the true effect. In the latter example, we have unintentionally stratified on an outcome not on a cause which means the true answer is in the aggregate not the stratified data.
[^1]: Pearl J. Causal Inference in Statistics. John Wiley & Sons; 2016
A: The St.Petersburg paradox, which makes you think differently on the concept and meaning of Expected Value. The intuition (mainly for people with background in statistics) and the calculations are giving different results.
A: The Jeffreys-Lindley paradox, which shows that under some circumstances default frequentist and Bayesian methods of hypothesis testing can give completely contradictory answers. It really forces users to think about exactly what these forms of testing mean, and to consider whether that's what the really want. For a recent example see this discussion. 
A: I find a simplified graphical illustration of the ecological fallacy (here the rich State/poor State voting paradox) helps me to understand on an intuitive level why we see a reversal of voting patterns when we aggregate State populations:

A: 
Suppose you obtained a data on births in royal family of some kingdom.
  In the family tree each birth was noted. What is peculiar about this
  family was that parents were trying to have a baby only as soon first
  boy was born and then did not have any more children.

So your data potentially looks similar to this:
G G B
B
G G B
G B
G G G G G G G G G B
etc.

Will the proportion of boys and girls in this sample reflect the general probability of giving a birth to a boy (say 0.5)? The answer and explanation can be found in this thread.
A: One of my "favorites", meaning that it's what drives me crazy about the interpretation of many studies (and often by the authors themselves, not just the media) is that of Survivorship Bias.
One way to imagine it is suppose there's some effect that is very detrimental to the subjects, so much so that it has a very good chance of killing them.  If subjects are exposed to this effect before the study, then by the time study begins, the exposed subjects that are still alive have a very high probability of having being unusually resilient. Literally natural selection at work. When this happens, the study will observe that exposed subjects are unusually healthy (since all the unhealthy ones already died or made sure to stop being exposed to the effect).This is often misinterpreted as implying that exposure is actually good for the subjects. This is a result of ignoring truncation (i.e. ignoring the subjects who died and did not make it to the study). 
Similarly, subjects who stop being exposed to the effect during the study are often incredibly unhealthy: this is because they have realized that continued exposure will probably kill them. But the study merely observes that those who quit are very unhealthy!
@Charlie's answer about the WWII bombers can be thought of as an example of this, but there's plenty of modern examples too. A recent example are the studies reporting that drinking 8+ cups of coffee a day (!!) is linked to much higher heart health in subjects over 55 years of age. Plenty of people with PhD's interpreted this as "drinking coffee is good for your heart!", including the authors of the study. I read this as you have to have an incredibly healthy heart to be still drinking 8 cups of coffee a day after 55 years of age and not have a heart attack. Even if it doesn't kill you, the moment something looks worrisome about your health, everyone that loves you (plus your doctor) will immediately encourage you to stop drinking coffee. Further studies found that drinking so much coffee had no beneficial effects in younger groups, which I believe is more evidence that we are seeing a survivorship effect, rather than a positive causal effect. Yet there's plenty of PhD's running around saying "Science says drinking 8+ cups of coffee is good for seniors!"
A: Try the Borel–Kolmogorov paradox, where conditional probabilities behave badly.  One example had the question

Let $X_1, X_2$ be independent exponential random variables with
parameter $1$.

*

*Find the conditional PDF of $X_1+X_2$ given that $\frac{X_1}{X_2}=1.$

*Find the conditional PDF of $X_1+X_2$ given that $X_1-X_2=0.$

*The events $\frac{X_1}{X_2}=1$ and  $X_1-X_2=0$ are the same. Does this mean that conditioning on either of these two events should give the same answer?


to which the answer appears to be

*

*$f_{X_1+X_2 \mid \frac{X_1}{X_2}=1}(x) = x e^{-x}$, a $\text{Gamma}(2,1)$ distribution  with mean $2$

*$f_{X_1+X_2 \mid X_1 - X_2=0}(x) = e^{-x}$, and $\text{Exp}(1)$ distribution with mean $1$
and this can be confirmed by simulation.
Whether conditioning on $X_1=X_2$ is really consistent with the assumption that $X_1$ and $X_2$ are independent is a deeper question.
A: Misspecification paradox
If $T$ is a method of statistical inference with a certain model assumption, say the true $P$ is assumed to be in some set ${\cal P}$ (e.g., $P$ may be assumed to be an i.i.d. normal distribution model for data $X_1,\ldots,X_n$), it is standard practice (in some quarters) to run a model misspecification test $M$, i.e., a test, that tests the null hypothesis $P\in {\cal P}$.
Assuming that $P(M$ rejects$)>0$ for $P\in {\cal P}$ (which is pretty much always fulfilled), it follows that the conditional distribution upon non-rejection $P(\bullet|M$ does not reject$)$ cannot be in ${\cal P}$. This is because
$$
P(M\mbox{ rejects}|M\mbox{ does not reject})=0
$$
in contradiction to $P(M$ rejects$)>0 \forall P\in {\cal P}$.
If $T$ is only applied in case that $M$ does not reject, it means that the distribution that generates the data that go into $T$ is the conditional distribution that is not in ${\cal P}$.
In other words, testing the model assumption and passing it (i.e., not rejecting it and taking the model as valid for the data) actively violates the model assumption, even if it was fulfilled before!
See
https://academic.oup.com/philmat/article-abstract/15/2/166/1572953?redirectedFrom=fulltext, https://arxiv.org/abs/1908.02218.
As I am (co-) author of these papers, I should acknowledge that in principle this is known already at least since Bancroft (1944), see reference in the arxiv paper, although I believe I was the first to call it a paradox and to present it in a way that its paradoxical "nature" comes out.
A: Sorry, but I can't help myself (I, too, love statistical paradoxes!).
Again, perhaps not a paradox per se and another example of omitted variables bias. 
Spurious causation/regression
Any variable with a time trend is going to be correlated with another variable that also has a time trend. For example, my weight from birth to age 27 is going to be highly correlated with your weight from birth to age 27. Obviously, my weight isn't caused by your weight. If it was, I'd ask that you go to the gym more frequently, please.
Here's an omitted variables explanation. Let my weight be $x_t$ and your weight be $y_t$, where
$$\begin{align*}x_t &= \alpha_0 + \alpha_1 t + \epsilon_t \text{ and} \\ y_t &= \beta_0 + \beta_1 t + \eta_t.\end{align*}$$
Then the regression
$$\begin{equation*}y_t = \gamma_0 + \gamma_1 x_t + \nu_t\end{equation*}$$
has an omitted variable---the time trend---that is correlated with the included variable, $x_t$. Hence, the coefficient $\gamma_1$ will be biased (in this case, it will be positive, as our weights grow over time).
When you are performing time series analysis, you need to be sure that your variables are stationary or you'll get these spurious causation results.
(I fully admit that I plagiarized my own answer given here.)
A: One of my favorites is the Monty Hall problem.  I remember learning about it in an elementary stats class, telling my dad, as both of us were in disbelief I simulated random numbers and we tried the problem.  To our amazement it was true.  
Basically the problem states that if you had three doors on a game show, behind which one is a prize and the other two nothing, if you chose a door and then were told of the remaining two doors one of the two was not a prize door and allowed to switch your choice if you so chose you should switch you current door to the remaining door.
Here's the link to an R simulation as well: LINK
A: It's not a paradox per se, but it is a puzzling comment, at least at first.
During World War II, Abraham Wald was a statistician for the U.S. government. He looked at the bombers that returned from missions and analyzed the pattern of the bullet "wounds" on the planes. He recommended that the Navy reinforce areas where the planes had no damage. 
Why? We have selection effects at work. This sample suggests that damage inflicted in the observed areas could be withstood. Either planes were never hit in the untouched areas, an unlikely proposition, or strikes to those parts were lethal. We care about the planes that went down, not just those that returned. Those that fell likely suffered an attack in a place that was untouched on those that survived.
For copies of his original memoranda, see here. For a more modern application, see this Scientific American blog post.
Expanding upon a theme, according to this blog post, during World War I, the introduction of a tin helmet led to more head wounds than a standard cloth hat. Was the new helmet worse for soldiers? No; though injuries were higher, fatalities were lower.
A: I'm surprised no one has mentioned Newcombe's Paradox yet, although it is more heavily discussed in decision theory. It's definitely one of my favorites.
A: There are lists of paradoxes on Wikipedia!
https://en.wikipedia.org/wiki/Category:Statistical_paradoxes
https://en.wikipedia.org/wiki/Category:Probability_theory_paradoxes
A: The hot hand paradox.
Quoting Miller and Sanjurjo's paper:

Jack takes a coin from his pocket and decides that he will flip it 4 times in a row, writing down the outcome of each flip on a scrap of paper. After he is done flipping, he will look at the flips that immediately followed an outcome of heads, and compute the relative frequency of heads on those flips. Because the coin is fair, Jack of course expects this conditional relative frequency to be equal to the probability of flipping a heads: 0.5. Shockingly, Jack is wrong. If he were to sample 1 million fair coins and flip each coin 4 times, observing the conditional relative frequency for each coin, on average the relative frequency would be approximately 0.4.

Intuitively, the problem is that the amount of available data is positively correlated with the number of heads - so a failure to follow heads with tails is more likely to occur in a sample where there are fewer instances of heads to analyse, giving it a larger impact on the calculated mean, introducing bias.
Sampling bias caused by this paradox went undetected in a notable study on the hot hand phenomenon in basketball for over thirty years (Wikipedia).
A: Let x, y, and z be uncorrelated vectors.  Yet x/z and y/z will be correlated.
