Article in the conversation detailing our research on predicting lethal and near lethal domestic violence. Hopefully some lessons for policymakers.
theconversation.com/can-we-predi...
New open access paper. We look at nationally representative data on the association bw child maltreatment and self reported victimisation after 15 years of age. We find strong associations between exp. multiple forms of maltreatment and later victimisation
www.sciencedirect.com/science/arti...
New paper from us: ijpds.org/article/view... comparing self-reported and police-detected violence over adolescence and early adulthood. tl;dr we undercount violence with both methods and the people missing from the self-report are probably the ones committing the most violence.
I think predicting an offended yes/no would still require you to select a threshold (or cutoff) probability for when a case is considered as predicted to offend in the model? e.g. in cases where the log reg predicts a greater than 50% prob. of offending you are a yes otherwise no.
Agree, but can be massively misleading for log reg in my experience because it depends on where you set the threshold for "correct" predictions.
Yeah, I saw this. The brier score is about the closest to what I was after but not really that informative about the absolute predictive performance of the models (as opposed to performance relative to the other models).
These do not need to be included in the statistical model but can be collected at the same time by officers.
Agree on shorter being better from a model parsimony perspective BUT sometimes the presence of additional questions are viewed positively (provided that they are not asked in a robotic checklist type way) by affected family members and serve purposes other than prediction (e.g. risk management)
Does anybody know of any estimates of the failure rate of experiments (i.e. hypotheses not supported) in physics and chemistry?
Guessed this was based on using the AUC as a measure of predictive performance. The claim of superior performance is silly if you are not directly testing against these tools in the sample.
It would be interesting to see the ppv at different model thresholds.
Excellent. Thanks. Followed.
Thanks Manne. Very helpful. Had no idea about starter packs but I have followed everyone on Matt's.
Hi. I am a researcher interested in quantitative crim. research/methodology/measurement. Can you please recommend me good accounts to follow on these broad topics?
Oh agreed. I am really only familiar with you from your book, so I thought it strange that you would be seen as a puppetmaster type figure based on that. Sorry that you have to put up with such nonsense.
Obviously the 10s of thousands of murders part is insane. But why are they referring to you as the architect behind "Soros DAs"? Is this all just in relation to your book and academic work?
The proportion of "other" is interesting. I am guessing the share of Black recruits is same or higher than the share in the general population. I wonder whether recruits from other backgrounds are underrepresented and whether this is related to how policing is viewed as a career in thesr communities
