The Alt-Ac Job Beat Newsletter Post 12

Hi Everyone,

My advice this week is to apply now, don't wait. I often talk to individuals who are waiting for the perfect opportunity (which is a mistake, since the interviewing cycle can take months in the best of times), or waiting to get skilled up (which most of you already have the skills, and the job adverts you are often silly).

Even if you get an interview and decide it is not a good fit, there was not much harm in the process (besides some of your time). Things can take a significant amount of time though (I would guess typical range in our field is 3-9 months looking). And if you get an offer, it is often OK to say "can I wait a month until I have finished the semester" or "can I wait two months until I have finished my dissertation". Those often are not deal breakers.

JOBS

Job board link

For some of the recent gigs:

EXAMPLE SCIENTIST

Renee Mitchell, has PhD from Cambridge in Criminology. Renee started the American Society of Evidence Based Policing, and has had research positions at RTI and Axon. To me Renee is a great example of "just do it yourself" -- no one made ASEBP, Renee did it because she saw a need for a group like that.

TECH ADVICE

I do not have hard data on this, but in my experience, for tree based machine learning models, you only start to get improvements over more typical linear models at around 20k observations in the data. (By "linear" here I mean linear in the parameters, so that includes logistic or other generalized linear models.)

It of course matters the nature of the underlying functions, but my personal theory is in noisy data settings (as we are dealing with in social science data) linear and additive is the best you can uncover without very large sample sizes. (And with small samples, hold out evaluations will also be more noisy.) So many examples in criminal justice simpler models often perform quite well, and even with large samples more complicated machine learning models only offer slight improvement.

Best, Andy Wheeler