Tuesday, May 28, 2019

Triple Differences Models

We all remember the first time we were introduced to differences in differences models. Was it the Card and Krueger minimum wage paper? Was it Card's Mariel boatlift paper?  You surely saw the nice tables with before and after in the treatment and control groups. Pretty intuitive, right? A bit more complicated to link the tables to regression estimates but doable. Since then, you've added to your difs in difs repertoire and maybe you've even done a few triple differences analyses. Everything is OK until...you go to write up your results and realize that the estimate on that triple interaction is not so easy to explain! Don't worry, we've all been there. 

My suggestion: Have another look at my favorite explanation of differences in differences (in differences) models. This video starring my colleague, Nishith Prakash, and his coauthor, Karthik Muralidharan. Then make similar diagrams for your own paper. Tell your story in the same way that they tell their story about bicycles. Maybe make a video? I think this will help you to tell your story in your paper. 

I also recommend reading the section on triple differences in Scott Cunningham's online book, Causal Inference: The Mixtape. I really like the entire chapter on differences in differences (starts on page 263), but I especially like the discussion of differences in differences in differences (starting on page 273). He provides lots of examples of papers that use triple differences techniques. You can refer to them to help you write up your results. He also provides sample Stata code for a DDD model! 

Thank you, Shiyi, for inspiring this post! I hope it's helpful.

Wednesday, May 22, 2019

It's May and You're Going on the Job Market Next Year?

Random trivia about me: May is my favorite month of the year! Sure, the better weather. Flowers are blooming. I survived the end of semester craziness. But the biggest things: It's the beginning of the summer break, and I have such high expectations for all I'll get done by the time classes start again. May is the month of promise and hopefulness. 

Ok, but for those of you planning on going on the market this coming year: Don't get too relaxed. There is lots to do this summer. See this very helpful twitter thread written just for you. The big things: Start writing! Spend time polishing your papers (e.g., make sure you put your footnote after the period). And also, don't ignore your advisors over the summer. Yes, they may be basking in the glory of May (see above paragraph), but they also want you to do well on the market next year and may not have as much time to help you once the semester starts. 

From our friends at PhDComics:
Image result for phdcomics, summer

Sunday, May 19, 2019

Read this When Your Paper Gets Rejected

It turns out that rejection improves the eventual impact of the paper. See here

This past Monday, one of my papers was rejected at a good journal. Of course it wasn't a happy day, but the referees' and editor's comments were very good. Since then, my coauthors and I have been brainstorming, and I really believe the new draft of the paper will be significantly better than the draft we submitted.

Also, listen to this by Adam Grant. 

But editors, take note...this does not mean I want you to reject more of my papers in the future. :)

  

Tuesday, May 7, 2019

Should We Use Figures for Statistical Inference? Evidence from RD

Alternative titles of this blog entry: 

Is the .05 P-Value Threshold too Lenient? 
or
When In Doubt, Simulate! 

But here goes: 

Whenever I teach RD design, I always urge students to make the pictures. In fact, I think the reason I love RD so much (and dislike IV so much) is precisely that we can see the magic right in front of our eyes. But every now and then, I see an RD figure that is .. less than impressive. And I find myself doubting the results. But should I?  

Kirabo Jackson wondered the same thing, and so he made a gif to show us simulated RD plots associated with various t-statistics. Brilliant! Yes, when the t-statistics are really big, the plots are beautiful, but even with a t-statistic of 3.8, the plot is "less than compelling". Clearly, our standards for beautiful RD plots are much more demanding than our standards for beautiful tables. So what should we do? Be more lenient with the RDers? More demanding of our tables? I'm not sure---read the comments for more insight on that. 

In the meantime, another big lesson: If you're ever not sure of something, simulate it! Read the entire thread. Kirabo includes the very simple simulation code he used to make his gif.