AKM SD = 0

[qgallea]

Good morning,

I managed to solve my first problem. However, now that the command is running I get a AKM SD of 0.

Here is the command I use: ivreg_ss conflict_num, endogenous_var(lnvalue) shiftshare_iv(IV_global_bin) control_varlist(polity2 lngdp lnexp internal_conflictpast15 yc id_) sharevarlist(share*) akmtype(1) firststage(1)

And here is the command without "shift-share" errors: ivreghdfe conflict_num polity2 lngdp lnexp internal_conflict_past15 (lnvalue= IV_global_bin) if continent=="Africa" , absorb(yc ctry_code) cluster(ctry_code)

Could you please help me?

Have a pleasant day, Quentin

[yijuhung]

I have run into a similar issue in that the AKM standard error is close to 0, which is way smaller than the usual or EHW one. Have you figured out what is going on?

[clowenstein]

I am also finding this to be the case (either AKM = 0 or is very small) in my panel data setting, especially when including high-dimensional fixed effects as in the example provided by @qgallea. Reducing the number of FEs helps (by de-meaning), but the corrected SEs are still very small relative to EHW and AKM is never > EHW except when removing location FEs. I've tried on both Stata and R, and both yield similarly tiny corrected SEs. Any assistance would be greatly appreciated, and thanks for making this code available!

[zhangxiang0822]

@yijuhung @clowenstein Hi guys, thanks so much for pointing it out. It would be a bit hard for us to figure out what's happening based on the command you use and the Stata output.

If you can, could you please prepare (1) a dataset that contains your outcome, control variables, FEs, and shift shares (2) a short readme file telling us what the variables mean (3) the command that you use? Ideally, we hope that we can use the dataset you provided to replicate your results, and we can help figure out what's happening. You could share it via google drive or Dropbox, or other methods that you prefer. Just shoot me an email at xiangzhang@princeton.edu.

Thanks so much!

[yijuhung]

@zhangxiang0822, thanks for your reply. I have sent you an email with the data, do-file, and the readme. Look forward to hearing back from you. I appreciate it!

[lampeggiante]

@zhangxiang0822, thanks for your reply. I have sent you an email with the data, do-file, and the readme. Look forward to hearing back from you. I appreciate it!

@yijuhung Hey, I have the same problem. Have you solved it?

[clowenstein]

@yijuhung I never solved this problem, although @qgallea directed me to an innovative solution in his recent paper in J. Dev Econ. I ended up using the Borusyak, Hull and Jaravel approach in which recasting the location-level regression as a shock-level regression yields corrected SEs. Perhaps this approach will work for you as well?