🚨Replication alert🚨 I reanalyze Moretti (2021), which tests for the...

@michael_wiebe
Michael Wiebe@michael_wiebe
75 views Oct 08, 2025 ~2 min read
Advertisement
1
🚨Replication alert🚨

I reanalyze Moretti (2021), which tests for the effect of tech cluster size on innovation. I find that the event study and instrumental variable estimates—necessary for a causal interpretation—are caused by coding errors.

#EconTwitter
1/
Media image
2
The main results are based on OLS regressions of patents on cluster size. The event study rules out sorting (promising inventors selecting into large clusters) and omitted variable bias (eg. local subsidies increasing both patents and size)

aeaweb.org/articles?id=10…

2/
3
Both supporting results are caused by coding errors.

The event study should interact the treatment variable with an indicator for t=0. But Moretti doesn't, so the estimate for β0 is based on all event years, instead of only t=0. Correcting makes the effect disappear.

3/
Media image
4
The IV should be calculated after sorting the data by firm, field, and city. But Moretti doesn't sort by city, so the first-difference is mistakenly calculated across cities. Correcting makes the first stage and 2SLS estimates disappear.

4/
Media image
5
Not sorting by city also leads to unreproducible results. Stata randomly orders the data before sorting, creating a random order by city. So running the code gives a different result each time. Here's the distribution:

5/
Media image
6
Here's a video where I walk through the code and show how the errors overturn the results.

youtube.com/watch?v=mXpGyd…

6/
7
After I made that video, I discovered another source of unreproducibility: Moretti uses several many-to-many merges with a nonunique sort order, which creates a slightly different sample each time you run the cleaning code.

7/
Media image
8
Takeaway: the OLS results remain valid, but now there's no evidence against confounding from sorting or omitted variable bias.

See here for the writeup:
michaelwiebe.com/assets/moretti…

I've submitted this to AER as a comment. Moretti didn't reply to emails.

8/
9
Should the paper be retracted?

I don't think this would have been published in AER without the event study and IV results. But I'm not sure retraction is the right call.

Thoughts?

9/
10
I also have an extended commentary that digs into:
- other event studies
- defining the IV in levels
- imputing missing observations
- changing the time unit
- heterogeneity by movers vs stayers
- what is the mechanism?

michaelwiebe.com/assets/moretti…

10/
11
If you want to see more replications, support me on Patreon:
patreon.com/michael_wiebe

If you have research you want to see scrutinized, please get in touch:
michaelwiebe.com/hire-me/

11/11
12
@RoyalTS Try this:
ko-fi.com/michaelwiebe
Actions
What You Can Do
  • Export as PDF or Markdown
  • Batch Export to Notion
  • Bookmark & Highlight
  • Screenshot Tweet
Create Free Account

Includes 7-day Premium trial

Advertisement