The new ABA employment outcomes data for law schools have been released, and that means it's time for a new WITNESSETH law school employment rankings. As in the past, I've ranked law schools using all of the data in the ABA employment reports. The technique used incorporates each category of employment outcome with a unique weight, rather than simply counting "good" and "bad" outcomes. As a result, the ranking uses every piece of data released by schools, and assigns a (positive or negative weight) to that item.
Unlike other ranking systems, this technique does not rely at all on anyone's subjective determination of what outcomes are good or bad. It's an unsupervised machine learning technique that determines the weight of each outcome based on its relationship to other outcomes among the 195 schools. Therefore, this is a truly objective ranking measure.
Unlike in past years, however, this year I decided to present the data in an "unranked" alphabetical format, and introduced uncertainty estimates (the horizontal bars). This avoids artificial distinctions among schools that are not substantively very different, and shows how much uncertainty there is in ranking schools even with a rich dataset like the ABA employment outcomes data. A note about objectivity and methodology follows the rankings.
Here are the 2020 law school unranked "rankings"!
* Methodological Note. The method used here is principal component analysis run over all of the items of employment data as defined in the ABA report, presented in proportion form and transformed with an arcsin transformation. The resulting score is the first principal component of the schools based on the employment criteria that best explain the variance in the data among the schools. There is no weighting on my part or other assessment of "good" or "bad," other than to say which end of the ranking (#1 or #195) is "better" or "worse." Given that one end of the ranking is dominated by biglaw jobs and federal clerkships, and the other end is dominated largely by unemployment, I feel this is a safe judgment. I created the uncertainty estimates using the non-parametric bootrstrap with 10,000 replicates.
Comments