search

QuickPay-Operational-Performance / Data-and-code

Data and code for econometric analysis
0 stars 0 forks source link

Portfolio effects model #60

Open JNing0 opened 3 years ago

JNing0 commented 3 years ago

I have posted the model for the portfolio effects in the "notes" folder. See here for the pdf and here for the markdown file. Please review and give comments. Thank you!

vibhuti6 commented 3 years ago

Thank you, Jie. This is helpful! I will get back to you if I have any questions or comments. Also, just for reference, here's the paper on war and wages (see equation 8) that I mentioned yesterday: https://economics.mit.edu/files/11620.

vibhuti6 commented 3 years ago

Hi Jie, thanks again for sharing these notes. They were very helpful. I have a few comments/questions. I think one issue with this specification is that the control group is the same for all terciles. The control group is all large projects. And we difference this out from small projects of firms in different terciles. The parallel trends assumption is likely to be violated if we model it this way, and our estimates could be misleading.

My understanding is that we should pick the control group from firms with a similar exposure to Quickpay. I think a small tweak to the specification you proposed would work:

Screen Shot 2021-03-11 at 6 46 51 PM

The treatment effect coefficients will be the same, but I think this specification allows us to ensure that control group comes from firms in the same tercile. Please let me know what you think, or if you have any comments or suggestions. Thank you!

JNing0 commented 3 years ago

Hi Vibhuti,

Good question about the parallel trend. I thought about that and I think we are OK. The reason is that the parallel trend is about the treatment and control groups WITHOUT the treatment. That is, absent of treatment, the treated and control projects will have the same trend. What we are capturing here is the different effect WITH treatment. So we are saying all treated projects, regardless of the QP exposure, have the same trend as the control group without QP. Then once QP occurs, they "split" because they are affected differently. This model is in essence the same as the intensity model, only that we are discretizing it and, instead of having an intensity metric, we have the portfolio weight. If you think about the intensity model, there is one control group and all treated observations have different intensity.

Jie Ning //////////////////////// Associate Professor Department of Operations Weatherhead School of Management Case Western Reserve University Cleveland, OH 44106 e-mail: @.*** tel: 216-368-3841 ////////////////////////

On Thu, Mar 11, 2021 at 9:52 PM Vibhuti Dhingra @.***> wrote:

Hi Jie, thanks again for sharing these notes. They were very helpful. I have a few comments/questions. I think one issue with this specification is that the control group is the same for all terciles. The control group is all large projects. And we difference this out from small projects of firms in different terciles. The parallel trends assumption is likely to be violated if we model it this way, and our estimates could be misleading.

My understanding is that we should pick the control group from firms with a similar exposure to Quickpay. I think a small tweak to the specification you proposed would work:

[image: Screen Shot 2021-03-11 at 6 46 51 PM] https://user-images.githubusercontent.com/47764977/110884611-358dec00-829a-11eb-8b0d-2d4c34d388ba.png

The treatment effect coefficients will be the same, but I think this specification allows us to ensure that control group comes from firms in the same tercile. Please let me know what you think, or if you have any comments or suggestions. Thank you!

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/QuickPay-Operational-Performance/Data-and-code/issues/60#issuecomment-797197759, or unsubscribe https://github.com/notifications/unsubscribe-auth/AOBDD5Q77KTTFCXBD3I2VA3TDFXVFANCNFSM4Y2W3GWA .

vibhuti6 commented 3 years ago

Thanks, Jie. I don't think that would suffice. The reason is that the pretends on small projects are likely to be different for firms in different categories. I tried to illustrate through an example here: https://www.dropbox.com/s/88x9fa028o2bczf/example.xlsx?dl=0. In this simulation, the parallel trends holds (up to time 4) when we compare the average of all small projects to all large projects. But it does not hold when we compare a subcategory of small projects to all large projects.

JNing0 commented 3 years ago

How does a treatment intensity model check the parallel trend assumption? The usual way by pooling all the treated objects together. Our formulation is not and should not be different from a treatment intensity model. The tercile model is the discretized version of the continuous model and is used only because it is easier to see the nonlinear effects. We can go with the continuous version if you are more comfortable with it. Let's discuss more on Monday,

Nonetheless, could you please run the models I formulated, both the tercile and the continuous versions? Thanks!

vibhuti6 commented 3 years ago

Hi Jie, I am of course more than happy to run the models you proposed and I will certainly post those results before our next meeting.

However, to the best of my understanding, the current discrete model is likely to give us biased estimates. I think the model should have a different baseline for each firm category.

I need to think more about the continuous model but I believe it should have a firm-specific baseline as well. That is, we should have firm fixed effect or include the firm’s "rank order" as a variable. Some references:

Let's discuss more on Monday, and I will post the existing model's results in the meantime. Thanks.

vibhuti6 commented 3 years ago

Hi Jie, I have posted the results here -- please see Section 8 in the document. Thanks.