Testval in distribution

What is testval in :

mu = pm.Normal('mu', mu=178, sd=20, testval=df.height.mean()) ?

Using it gives me a very different posterior of mu.

Usually, we use the testval to

  1. infer/validate the shape of the random variable,
  2. set the starting value of the sampler

This is an indication of problem of your model, that it is sensitive to initial condition of the chain.

Can you give me more details to exactly how we do 1 or 2?

If I dont have a very large dataset, wont the model be sensitive to initial conditions anyways?