The value of the areas you mentioned is problematic.  Although its value 
is technically correct in accordance with our distribution algorithm, there 
is still reason to be cautious.

Here is why

1) GRUMP proportionally allocates population based on land area
2) Land area is determined through the application of a water mask such
that each pixel value represents the total land area of that pixel minus
the amount that is covered with water.
3) The administrative units and water mask come from different sources
and might not be registered completely accurately (with a pixel size of
1 sq km, and positional accuracy within ~500 m)
4) As a result, certain coastal pixels might be assigned quite low area
values, but have quite large population counts in accordance with census
block group date in the US.
5) Low area values and high population counts = high population density
values.

This really highlights the fact that GRUMP is not appropriate for pixel
level analyses, or even block or tract level.  The raster resolution is
too low for results at these scales to be meaningful and sensitivity to
outliers is increased.  It is also difficult to identify outliers
because at this scale the sample size of pixels is too small for
meaningful statistical analysis.

More appropriate would be an analysis at the county scale, where the
sample of pixels is large enough to control for potential outliers and
the influence of positional inaccuracy is diminished.