How does name analysis work? by Pete Warden.
From the post:
Over the last few months, I’ve been doing a lot more work with name analysis, and I’ve made some of the tools I use available as open-source software. Name analysis takes a list of names, and outputs guesses for the gender, age, and ethnicity of each person. This makes it incredibly useful for answering questions about the demographics of people in public data sets. Fundamentally though, the outputs are still guesses, and end-users need to understand how reliable the results are, so I want to talk about the strengths and weaknesses of this approach.
The short answer is that it can never work any better than a human looking at somebody else’s name and guessing their age, gender, and race. If you saw Mildred Hermann on a list of names, I bet you’d picture an older white woman, whereas Juan Hernandez brings to mind an Hispanic man, with no obvious age. It should be obvious that this is not always reliable for individuals (I bet there are some young Mildreds out there) but as the sample size grows, the errors tend to cancel each other out.
The algorithms themselves work by looking at data that’s been released by the US Census and the Social Security agency. These data sets list the popularity of 90,000 first names by gender and year of birth, and 150,000 family names by ethnicity. I then use these frequencies as the basis for all of the estimates. Crucially, all the guesses depend on how strong a correlation there is between a particular name and a person’s characteristics, which varies for each property. I’ll give some estimates of how strong these relationships are below, and I link to some papers with more rigorous quantitative evaluations below.
Not 100% as Pete points out but an interesting starting point. Plus links to more formal analysis.