I was recently admitted at a university which disseminates a list of the admitted students along with their emails in order to facilitate communication and increase camaraderie before the semester begins. I was wondering if people with last names in certain ranges (e.g. one of A, B, C... or one of Aa-Hq, Hr-Qh, Qi-Zz) are more likely to make up the graduating class.

I initially considered just taking the proportion of people with names in each range, but then I realized that I would have the same bias present in maps that measure phenomena by state--the size of the name pool would make it seem like names in those pool are more likely to be in the graduating class. For instance, if 99% of last names in general begin with A, then of course it will seem like people with A-names are more likely to be in the class.

Since the solution to this bias is usually per capita measurements, my next thought was to take per nomine measurements by dividing each number of students with some last name type by the number of names of that type in existence. So if there were 200 students with last names starting with A, and 10000 last names in existence starting with A, then the statistic for the As would be $\frac{200}{10000}$. But then again, I'm not sure if this might skew the data because of name types like J which may not have a lot of names, but include names like Johnson which capture a lot of people.

Is my thinking correct, or will my proposed "per nomine" measure do the trick? Is there a better way to do this? Maybe dividing the number of people with a certain name type by the proportion of children born in 2000 who have that name type?

I'm interested in both the most precise process from a theoretical standpoint and the most practical approach from a data-acquisition standpoint (assuming said approaches differ).


1 Answer 1


I am not sure that you can derive any valuable knowledge concerning being accepted/rejected based on the described dataset, since you only have the "accepted" class, and not the "rejected" ones. (Although, I believe such bias in the acceptance taste of many institutes/corporations exists and any study in this direction would be very valuable and much appreciated.)

To elaborate more on this; there is an important piece in your research that is missing, and that is the list of the names of the people who applied for that program in the first place. You cannot assume that it was entirely random, and we are not being very picky here. There is indeed some significant bias in the form of the names of the applicants (specially for graduate level) , which should be considered. Applicants with Asian/Indian/Middle-Easter names. Those names have their own unique structures. For instance, a western-name rarely starts with the letter 'X', which happens very often in Asian names. You would need this piece of information to be able to make any judgment regarding the interesting patterns of the accepted names. Otherwise, everything would be relative results. Similar to the location of an object without having set any origin point for our space.

Apart from this missing piece, I agree that your weighting model could reduce the bias. The other method would be to down-sample the list of names to get a balanced data but I do not think you have enough names to be able to remove a good chunk of data yet have a large-enough data to mine from.

Btw, I am not sure if it was entirely a right thing to do by that university to publish people's name and email addresses! But, that is for a different debate I guess.

  • $\begingroup$ Thanks for your answer! Regarding the list, it was an opt-in service, so the publication was consensual. $\endgroup$
    – actinidia
    Commented Apr 17, 2018 at 1:54

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.