search

xuyiqing / gsynth

Generalized Synthetic Control Method
Other
134 stars 41 forks source link

Adding autocovariance to gsynth to fix autocorrelated errors #46

Open feliuserra opened 4 years ago

feliuserra commented 4 years ago

Hi,

I've been using this package for a while and I have to say I am very happy with it. However, I've got a question. I've seen that the paper on NNMCM from Athey et. al, and it says that one drawback of MC-NNM is that it does not take into account the time series nature of the observations. Thus, it is likely that the errors (e_{it}) are correlated over time. However, one can modify the objective function to include an autoregressive coefficient to solve this issue. (Section 8.3 from the paper).

I've been looking at your repository, and I haven't found this correction. I've seen some autocorrelation terms but not used on the objective function (correct me if I'm wrong). My question is then, has anyone looked at this? Has anyone tried to implement it? I would like to help if possible, or at least know if anyone is working on this issue, since it seems quite relevant, especially for long time series with high autocorrelation.

Thanks in advance,

Feliu

xuyiqing commented 4 years ago

Thanks for reaching out! I think we do have an AR(1) option, but we have shut it off because it involves calculating the cumulative effects.

Licheng, do we still support (automatically) Lagged Dependent Variables?

On Fri, Jul 10, 2020 at 6:31 AM Feliu Serra Burriel < notifications@github.com> wrote:

Hi,

I've been using this package for a while and I have to say I am very happy with it. However, I've got a question. I've seen that the paper on NNMCM from Athey et. al, and it says that one drawback of MC-NNM is that it does not take into account the time series nature of the observations. Thus, it is likely that the errors (e_{it}) are correlated over time. However, one can modify the objective function to include an autoregressive coefficient to solve this issue. (Section 8.3 from the paper).

I've been looking at your repository, and I haven't found this correction. I've seen some autocorrelation terms but not used on the objective function (correct me if I'm wrong). My question is then, has anyone looked at this? Has anyone tried to implement it? I would like to help if possible, or at least know if anyone is working on this issue, since it seems quite relevant, especially for long time series with high autocorrelation.

Thanks in advance,

Feliu

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/xuyiqing/gsynth/issues/46, or unsubscribe https://github.com/notifications/unsubscribe-auth/AB2PKGFNMYMFBGOHGX2JAD3R24JZRANCNFSM4OWTB6HA .

-- Yiqing Xu

Assistant Professor Department of Political Science Stanford University http://yiqingxu.org/

feliuserra commented 4 years ago

Hi, thanks for the quick response. Let's see what Licheng has to say.

Also, perhaps I understood it wrong, but you mean that I could introduce the lagged dependent variable manually? Would that work?

Thanks again for everything

xuyiqing commented 4 years ago

It can work, but what you get is the immediate effect, not the cumulative effect.

On Thu, Jul 16, 2020 at 7:13 AM Feliu Serra Burriel < notifications@github.com> wrote:

Hi, thanks for the quick response. Let's see what Licheng has to say.

Also, perhaps I understood it wrong, but you mean that I could introduce the lagged dependent variable manually? Would that work?

Thanks again for everything

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/xuyiqing/gsynth/issues/46#issuecomment-659439664, or unsubscribe https://github.com/notifications/unsubscribe-auth/AB2PKGCYFLLZYOJOLPQSPALR34DHJANCNFSM4OWTB6HA .

-- Yiqing Xu

Assistant Professor Department of Political Science Stanford University http://yiqingxu.org/