We re-analyse recent Cepheid data to estimate the Hubble parameter $H_0$ by
using Bayesian hyper-parameters (HPs). We consider the two data sets from Riess
et al 2011 and 2016 (labelled R11 and R16, with R11 containing less than half
the data of R16) and include the available anchor distances (megamaser system
NGC4258, detached eclipsing binary distances to LMC and M31, and MW Cepheids
with parallaxes), use a weak metallicity prior and no period cut for Cepheids.
We find that part of the R11 data is down-weighted by the HPs but that R16 is
mostly consistent with expectations for a Gaussian distribution, meaning that
there is no need to down-weight the R16 data set. For R16, we find a value of
$H_0 = 73.75 \pm 2.11 \, \mathrm{km} \, \mathrm{s}^{-1} \, \mathrm{Mpc}^{-1}$
if we use HPs for all data points, which is about 2.6 $\sigma$ larger than the
Planck 2015 value of $H_0 = 67.81 \pm 0.92 \,\mathrm{km}\, \mathrm{s}^{-1} \,
\mathrm{Mpc}^{-1}$ and about 3.1 $\sigma$ larger than the updated Planck 2016
value $66.93 \pm 0.62 \,\mathrm{km}\, \mathrm{s}^{-1} \, \mathrm{Mpc}^{-1}$. We
test the effect of different assumptions, and find that the choice of anchor
distances affects the final value significantly. If we exclude the Milky Way
from the anchors, then the value of $H_0$ decreases. We find however no evident
reason to exclude the MW data. The HP method used here avoids subjective
rejection criteria for outliers and offers a way to test datasets for unknown
systematics.