It seems like your approach would suffer a lot of selection bias and would need to collect a lot of data before it would be useful, hence it'll be hard to boots

It seems like your approach would suffer a lot of selection bias and would need to collect a lot of data before it would be useful, hence it'll be hard to bootstrap contributors. I'm not an academic, but wouldn't a better strategy be to do co-authorship analysis on papers to generate similar information. That is you could say researchers who co-wrote their initial papers with more senior researcher X (normally indicating a supervisor or mentor relationship) at lab Y tend to end up at Z and have an average research impact score of R. That way you could get all the data from analyzing public citation databases and avoid the chicken-and-egg problem with trying to crowdsource the data, and go directly to solving the main problem of comparing potential PhD supervisors based on their trackrecord.