Prior

The prior describes before we look at the data, what are our beliefs of the parameters θ? This part can feel a bit strange, because it can feel like you're injecting some subjectivity into your analysis. You can get different results if you make different choices for the priors. How do you choose good values for your priors?

Informative priors

You might think to take a look at the distribution of any historic data, and use that distribution to give you a prior distribution for your parameters. Say for example you calculate the first statistical moment (the mean) and the second statistical moment (the variance). Why not use that for your prior? This is where the terminology moment matching priors comes from.

Informative priors seem to be recommended against. There is a risk that a tight prior distribution will dominate the posterior inferences that you make.

Some people talk about eliciting priors, where you talk to a lot of independent experts, or doing a literature review to get a range of values you can use for the prior. Here the priors are informative because they provide numeric information to the results, but these seem to be more acceptable because there is less room for bias. If you are using informative priors, then you have to be very explicit about having done it.

Weakly informative priors

The STAN wiki and Andrew Gelman's blog talk about a concept called weakly informative priors. The idea is that you flatten the informative prior distribution and make it less constraining. You still get some regularisation because you tell the model where the less likely values are, but you lower the risk of ruling out parts of the parameter space which should be explored.

Uninformative priors

There are techniques that you can use to create prior distributions that contain no information in them. For example the Jefferys prior. These seem to be tricky to use in real world analysis, and don't give any real benefit other than making you feel better.

Don't use constrained uniform priors unless it is a true constraint

One bit of clear advice I've picked up is that you should be very careful when picking a probability distribution that has zero at any point. What you are saying with a prior that has zero values is that those values are impossible, and will never come through in the posterior distribution.

The example where you see this most often is when using the constrained uniform prior. Say you want to describe the average lifespan of a human, it is very tempting to use a uniform prior constrained between 0-100 for example. This makes you feel better because you think that it is less informative than for example a normal distribution, but what it does is guarantee that you'll never explore the space above 100. This is bad and shouldn't be done.