Randomization inference has been increasingly recommended as a way of analyzing data from randomized experiments, especially in samples with a small number of observations, with clustered randomization, or with high leverage (see for example Alwyn Young’s paper, and the books by Imbens and Rubin, and Gerber and Green). However, one of the barriers to widespread ...
Topics:
David McKenzie considers the following as important:
This could be interesting, too:
Bank of Japan writes Speech by Deputy Governor Nakaso at the Federal Reserve Bank of New York on Oct. 18, 2017 (Evolving Monetary Policy: The Bank of Japan’s Experience)
David McKenzie writes Six Questions with Rohini Pande
Jp Koning writes An example of tax-driven money during the greenback era
Swiss National Bank writes 10 Franc Note: Banknote App Updated
Randomization inference has been increasingly recommended as a way of analyzing data from randomized experiments, especially in samples with a small number of observations, with clustered randomization, or with high leverage (see for example Alwyn Young’s paper, and the books by Imbens and Rubin, and Gerber and Green). However, one of the barriers to widespread usage in development economics has been that, to date, no simple commands for implementing this in Stata have been available, requiring authors to program from scratch.
This has now changed with a new command ritest written by Simon Hess, a PhD student who I met just over a week ago at Goethe University in Frankfurt. This command is extremely simple to use, so I thought I would introduce it and share some tips after playing around with it a little. The Stata journal article is also now out.
How do I get this command?
Simply type findit ritest in Stata.
[edit: that will get the version from the Stata journal. However, to get the most recent version with a couple of bug fixes noted below, type
net describe ritest, from(https://raw.githubusercontent.com/simonheb/ritest/master/)
How does the command work?
- Using it when you have randomized at the individual level within strata
One of the outcomes I consider is truncated profits in the second round follow-up, for firms that were existing at the time . My stata command for doing this in the paper was:
areg s_prof_trunc assigntreat, a(strata) robust
where s_prof_trunc was the profit outcome, and assigntreat my treatment assignment variable, and I am controlling for randomization strata. The p-value from this regression is 0.051.
The code for using this new command is then:
ritest assigntreat _b[assigntreat], reps(1000) seed(125) strata(strata): areg s_prof_trunc assigntreat, a(strata) robust
where here I am telling it what the treatment variable is (assigntreat), and what coefficient I want the p-value for, telling it to do randomization inference within the randomization strata, and do use 1000 replications or draws. The output looks like this:
You can see the p-value is 0.047 here, very similar to what I got from the regression. This is individual-level randomization, with a sample size of 497, so we might expect randomization inference to give similar results to regression.
2. Individual-level randomization, but with multiple time periods and clustered standard errors
Here I have re-shaped the Nigeria data so that I am estimating a pooled effect over the second and third rounds of follow-up data together. The p-value when I use regression to do this is 0.096. The command is then very similar to that above, except now telling it also to take into account the clustered nature when doing the permutations, by using the cluster option (with uid being my individual identifier):
ritest assigntreat _b[assigntreat], reps(1000) strata(strata) cluster(uid) seed(124): areg prof_trunc assigntreat time3 if group<=2 & existing==1, a(strata) robust cluster(uid)
The p-value I get is 0.100, which is again very similar to the regression value I estimate. In contrast, if I had not specified the cluster() option, it would give me the artificially low p-value of 0.06.
3. Clustered randomization with stratification
A third example I tried comes from ongoing work in Colombia on shortening supply chains. Here we have firms in 63 market blocks, which were formed into 31 strata (pairs and one triplet), with clustered randomization occurring at this market block level. An outcome of interest here is how many days a week firms shop at the central market. The p-value I get in the regression with clustered standard errors is 0.024. Randomization inference is meant to make more of a difference with clustered randomizations with relatively few clusters, so I was curious to see what difference it makes here:
ritest b_treat _b[b_treat], reps(5000) cluster(b_block) strata(b_pair) seed(546): areg dayscorab b_treat b_dayscorab miss_b_dayscorab round2 round3, cluster(b_block) a(b_pair)
So here randomization inference does make a big difference – increasing my p-value from 0.024 to 0.106. So if you are doing clustered randomization, randomization inference will be particularly important to use.
Some observations and caveats
This is fantastically easy to use for the simple randomization cases, and should lead to many more people using randomization inference, at least as a check. Here are some observations from my initial uses to keep in mind:
- Despite the ability to set the seed for randomization, it currently does NOT seem to give replicable responses. For example, in case 1 above, I ran the code with 1000 replications and the same seed 5 times in a row, and got different p-values each time: 0.047, 0.054, 0.039, 0.053, and 0.057. These are all close, but obviously it is not ideal that the same code will produce different p-values each time it is run.[edit: Simon has now fixed this in the latest version available on his github https://github.com/simonheb/ritest/blob/master/README.md]
- This raises the second issue of how many replications you should use. The default is 100 replications. This is likely to be way too low in many settings. Alwyn Young uses 10,000 draws, but says he finds little difference once he goes beyond 2000 draws. Once you are doing enough draws, point 1 will become less of an issue.
- Time taken: in my simple case above, it takes my laptop about 35 seconds to run this with 1000 draws, and 3 minutes to do 5000 draws. However, this is after I have dropped all other variables from my dataset. When I had my full dataset, it took more than 2 minutes to do 1000 draws. So if you want this to run relatively fast, drop all the other data from your dataset apart from what is needed for your regression.
- The command doesn’t allow for xi to be used in the regression you are running – so if your regression had been using this, you should remove this from your command. Jason Kerwin notes that you can still just put e.g. i.groupvariable to add group variable fixed effects, without using xi.
- Jason Kerwin notes to me that if your strata variable is a string rather than numeric, then the code ignores your strata and will give incorrect standard errors. So make sure any strata variables are numeric. [edit: Simon has now fixed this in his latest version, available on his github https://github.com/simonheb/ritest/blob/master/README.md]
- More complicated randomizations such as two-level clustered randomization
7. Multiple treatments:
The standard textbook explanation of randomization inference always seems to involve just a single binary treatment, so that a unit is either treatment or control. Then under the sharp null that all treatment effects are zero, the outcomes for each unit should be the same whether or not they are assigned to treatment or control, and so we can just estimate the treatment effect under all the possible permutations and see in how many of these permutations do we get a coefficient at least as large as the one in the actual data.
However, suppose now that we assign firms to control (T=0), traditional business training (T=1), or personal initiative training (T=2) as in this recent paper. Then we have at least three hypotheses we might want to test: no effect of traditional training, no effect of personal initiative training, and no effect of either training. It seems to me that we then need to use three separate sets of permutations to test these, and that you will need to program up each of these as a file of permutations for the program to then call on. For example, if I had 10 observations and my treatment assignment variable has values (0, 1, 0, 2, 1, 2, 0, 1, 2, 0) then to test no effect of traditional training, I would want to only permute the 0s and 1s with one another, without changing the assignments of those getting personal initiative training. I have yet to program this.
If anyone has examples they want to share of using this command to do randomization inference with multiple treatments or multi-level randomizations, please link to them in the comments.
For further reading:
Jason Kerwin also blogged last week about the difference between randomization inference and bootstrapping, and this command.
Alwyn Young has also developed a Stata command randcmd.ado available on his webpage, which I have yet to play around with.