Page 150-167 of the Publication Manual of the American Psychological Association, 6th Edition has a detailed description of how figures should be displayed, and what they should convey. Here are two quotes from the book that should help you:
pg 152-153:
A good figure
- augments rather than duplicates the text,
- is easy to read-its elements (type, lines, labels, symbols, etc.) are large enough to read with ease,
- is consistent with and in the same style as similar figures in the same article ...
Be certain in figures of all types that
- lines are smooth and sharp,
- has typeface that is simple (sans serif) and legible
pg 161:
... Avoid the use of three-dimensional and other effects (including
color), except in rare instances in which they demonstrably enhance
the presentation of your data. Individual publishers have stated
policies with regard to color printing.
Size and proportion of elements. Each element must be large enough and sharp enough to be legible. Use a simple typeface (such as Arial,
Futura, or Helvetica) with enough space between letters to avoid
crowding. Letters should be clear, sharp, and uniformly dark and
should be sized consistently throughout the figure. Type style affects
legibility. For example, boldface type tends to thicken and become
less readable. The size of lettering should be no smaller than 8
points and no larger than 14 points. As a general guideline, plot
symbols should be about the size of a lowercase letter of an average
label within the figure. Also consider the weight (i.e., size,
density) of each element in a figure in relation to that of every
other element, making the most important elements the most prominent.
For example, curves on line graphs and outlines of bars on bar graphs
should be bolder than axis labels, which should be bolder than the
axes and tick marks. Shading. Limit the number of different shadings
used in a single graphic. If different shadings are used to
distinguish bars or segments of a graph, choose shadings that are
distinct (e.g., the best option to distinguish two sets of bars is no
shading [open] and black [solid]). If more than three shadings are
required, a table may be a better presentation of the data. Use
computer-generated art in such a way as to maximize the clarity of the
resulting graphic. And as always, keep it simple and clean looking.
So it's best practice to use a sans serif font, since, while serif helps with the reading-flow in body text, it is not as legible as sans serif in figures. I use Calibri for figures and Times Roman for body. You might choose to use CMU Sans Serif in your figures since you use CMU Serif in your body. Font size should be between 8-14 points (as long as it is easy to read). The text mentions nothing about the font size needing to be smaller or equal to the body text.
In the example you provided, a good font size would be between (a) and (b). In (a), the font size makes it unclear which tick each number is associated with on the graph. In (b), the text seems to not be optimally sized for legibility. Assuming your printing this, be sure to check the hardcopy for figure legibility.
The quote above says "consistent with and in the same style as similar figures in the same article". This doesn't mean that the font sizes in all figures should be the same. If a large font is used in a large graph, the same font size would look too large in a small graph (as in (a) from your example).
Final note. You can use
\the\textwidth
to print the width of the body, and
\the\textheight
to print the height of the body
This can help in determining the dimensions of a figure you're producing.
Update: according to a recent conversation in the Tex stackexchange chat, it is more common to use sans serif in figures in the fields of biology and chemistry papers where as the fields of mathematics, physics, and engineering tend to keep the font in figures consistent with the serif font used in the body of the article.
pgfplots
. That way you have all the flexibility to control the fonts and can scale the graphic without it affecting the font size.