search

jeremysze / LPIS

Repository of code used to analyze LPIs in NYC
0 stars 0 forks source link

Non-spatial Panel regression #8

Open jeremysze opened 5 years ago

jeremysze commented 5 years ago

https://github.com/jeremysze/LPIS/blob/d91897f809ccf55b52f6a089c639c79bf316abcd/analysis_qt_panel_3b.ipynb

Under Time Trends, I ran the interaction of flag_school (intersections within 200ft of school), flag_seniors (intersections within safe senior zone) and flag_priorityinters (intersections within 10 ft of signal intersection).

I think that was what we had spoken about.

jeremysze commented 5 years ago

The coefficient on flag_LPIS is smaller compared to the results from excluding the time trends of school, seniors and priority intersections (in Outcome of collision counts).

jhconning commented 5 years ago

This is great. Smaller but still clearly significant for almost all specifications which suggests your results a very robust to the inclusion of many controls and practically anything you throw at it. So the LPIS interevention seems effective..

Can you translate this into human-understandable marginal effects? Adding LPIS reduces reported pedestrian injuries by how much in a year?

jeremysze commented 5 years ago

Margins is taking a surprisingly long time to finish running.

jeremysze commented 5 years ago

https://github.com/jeremysze/LPIS/blob/master/analysis_qt_panel_3b.ipynb

It ran! With the Poisson fixed effect panel regression, 5.45% decrease in collision counts when LPIS is implemented. I ran margins, dydx(flag_LPIS). The marginal effects are identical to the coefficients in the regressions.

With the linear fixed effects panel regression, 0.165 decrease in collision counts when LPIS is implemented.