I'm graduated now, and swear I remember this exact kind of problem coming up in my Bayesian statistics class, but I can't remember what the answer was.

The Problem

So my wife's brother is red/green colorblind (xY), but her father is not(XY), and her mother is not, so her mother must be a carrier(Xx). Red/green color-blindness is recessive on the X Chromosome(x), which means in order to have the phenotype (actually have it), all of your X chromosomes must have the gene(xx or xY). This means women must get it from both their mother and father, while men must get it from their mother.

As a result, my wife has a 50% chance of being a carrier(Xx or XX), while all of my sons have a 25% chance of being red/green colorblind(XY or xY).

What I'm trying to figure out is if my first son ends up NOT being red-green colorblind (XY), does that change the probability that my wife is a carrier, since I've now observed one datapoint that refutes that possibility? If so, by how much?

My Thoughts

I think that Frequentists would say that it remains a 50% chance due to them being Independent events, but again, I can't recall exactly how it worked. I may be spacing it, but if my memory serves, Bayesians would include the original assumption(50%) as essentially a datapoint, but would subjectively give it more or less weight. I'm wondering if there's an objective way to handle this.

Of course, I can always calculate the odds of her being a carrier and how unlikely it is that she would have X sons without color-blindness (simple Geometric Distribution), but that doesn't tell me anything about her original probability.


Here is my simple attempt.

We have prior of your wife being a carrier $\pi(Y = 1) = 0.5$ and the probability of your first son being colourblind $p(X = 0|Y = 1) = 0.5,\ p(X = 0|Y = 0) = 1$. Now using simple Bayes rule we have:

$$\pi(Y = 1|X = 0) = \frac{p(X = 0| Y = 1)\pi(Y = 1)}{p(X = 0)} = \frac{p(X = 0| Y = 1)\pi(Y = 1)}{p(X = 0| Y = 1)\pi(Y = 1) + p(X = 0| Y = 0)\pi(Y = 0)} = \frac{\frac{1}{2} * \frac{1}{2}}{\frac{1}{2} * \frac{1}{2} + 1 * \frac{1}{2}} = \frac{1}{3}$$

  • $\begingroup$ So for a simplified solution would I be right in saying each additional non-colorblind son results in a 1/3rd decrease in the probability that my wife is a carrier? $\endgroup$ – Ecksters Nov 13 '17 at 16:13
  • 2
    $\begingroup$ Don't think so, but if you assume that $X_i | Y $ are independent (conditioned on $Y$ - which should be a rational assumption) then you can iterate this process, reusing the posterior $\pi(y)$ calculated a step before. From my quick calculations you will obtain: $\pi(Y = 1|X_1 = 0, X_2 = 0) = \frac{1}{5}$, and so on. $\endgroup$ – Łukasz Grad Nov 13 '17 at 16:22

I do not think that Frequentist's would say that (the 50% probability). Imagine the ridiculousness of the statement if you extend it to ten first sons without color blindness.

  • A Bayesian statement would go something like assigning probabilities to the different cases. And then you have P(mother = Xx) and P(mother = XX) changing from estimates of the odds ratio prior to the experiment, $\frac{P(XX)}{P(xX)} = 1$, to odds ratio posterior to the experiment $\frac{P(XX|obs)}{P(xX|obs)} = \frac{P(XX)}{P(xX)} \frac{P(obs|XX)}{P(obs|xX)} = 2$. A frequentist could also make a 95% confidence interval which woudl both include the mother having xX and XX, another not so useful approach in this case.

  • A frequentist would more like create a statement that has a certain success rate. For instance, do not reject the hypothesis that the mother has xX, since the p-value is only 0.5, (which means the approach, if it would have been rejected, fails 50% of the time when the mother has xX after all, this does not mean that the probability for a particular experiment is 50%). In the case of such small numbers, and for a single point estimate, the frequentist approach seems less useful.

    The frequentist might also do something like create a 95% confidence interval, which would in this case include both the xX and XX, another not so useful approach in this case.

Of course this 'a frequentists would...' is a bit ridiculous. No statistician is frequentist or bayesian. Only the methods have this property, and frequentists methods, as this example shows, have a disadvantage for such small experiments, with few observations.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.