Contrary to your currently accepted answer, although obviously there is an artistic element to the design of quantitative graphics (as emphasized by The National Geographic), there do exist objective principles and good advice based on them. This is not wholly a "matter of personal preference."
Fine map-making is an art; it is also a science, and the one should complement the other on equal terms.
(Charles Riddiford, quoted in the Smithsonian article, ibid.) But what exactly is the "science?"
It is helpful to look into the literature on cartography (map making): people have been making maps much longer than they have been making statistical graphics, yet the same principles of construction and interpretation apply.
For a good list of references to the literature, consult the Wikipedia article on cartographic labeling.
Practical advice given by a major vendor of cartographic software for making Web graphics can be found at http://mappingcenter.esri.com/index.cfm?fa=ask.answers&q=1469.
Good web fonts have a generous amount of space between each character, as well as the amount of whitespace within the characters (glyphs) themselves. A tall x-height also opens up the space within a character. These properties are what make these fonts so legible on screen.
The accepted answer to a parallel question on our sister site, GIS, recommends TypeBrewer, software to help
... explore typographic alternatives and see the impact that various elements of type have on the overall look and feel of a [graphic].
Standard textbooks on cartographic design have extensive analyses and discussions of typographics. Borden Dent, author of the popular Cartography: Thematic Map Design (Fifth Edition, McGraw-Hill, 1999) writes
[Graphic] lettering should be viewed first as a functional symbol on the map, and only secondarily as an aesthetic object.
By "map lettering" Dent means the "selection of lettering type and its placement" on the graphic. Selection elements include style and size; placement elements include position and spacing.
Dent discusses several kinds of research, including psychological experiments intended to identify typographic principles that communicate information quickly, clearly and accurately; as well as empirical experiments that characterize how maps are actually made. As one might imagine, much of this is contradictory, but some general tendencies emerge. They lead to advice that ranges from considerations about labeling points (e.g., not all positions of labels relative to point symbols are equally effective; don't use point symbols that can be confused with letters; avoid other "distracting" elements nearby) to labeling linear and polygonal features. Much attention is also paid to titles, axes, legends, and other auxiliary matter that is essential for communicating quantitative information.
One particularly intriguing experimental finding is that
Choice of typeface has little effect on legibility in the [graphic] context.
(Quoting Richard J. Phillips, Elizabeth Noyes, and R. J. Audley, "Searching for Names on Maps," Cartographic Journal 15 (1978): 72-76.)
Instead, other aspects are more important, especially consistency of the type in the graphic and its presence in an associated legend: when the reader knows what kind of typeface to expect, they can search more quickly and accurately within the graphic itself. Surely this has implications for certain kinds of statistical graphics, such as those used to display and summarize raw data.
This leads to a final caution: it is evident that one's design choices for making statistical graphics will depend partly on the intended purpose for the graphic, on the medium (such as a Web page versus print) and on how the audience is intended to interact with and use the graphic. Although such considerations would seem to make typographical choices almost arbitrary and uniquely personal, we should nevertheless seek to understand the visual and psychological principles that underlie how people read and make sense of maps and statistical graphics so that we can make well reasoned choices that are likely to work well in each application.