Limit sample to firms that exist pre and post treatment time point

[vob2]

This is a robustness check for now (will need to discuss if we want to use this reduced sample as our main analysis)

Quoting Harish:

Two other thoughts:

Could quick payment affect the composition of firms (small business primes)? In the original (non-quickpay) world, "weak" small businesses may not bid for government contracts because of the late payment. (Weak firms could be poorly managed firms that do not do a good job of managing their working capital; and this may be correlated with weak management overall.) Under quickpay, some weak firms may bid and win some contracts. The increase in delays may then not be a direct (moral hazard) effect on the behavior of firms, but may be the result of more weak firms entering the government contracting world.

The same argument can be made as a survival effect; more weak firms may survive under quickpay.

The way to address this may be to only consider firms that exist both pre- and post-quickpay.

[JNing0]

I have been thinking along the lines of Harish's comments. We have been treating this as a repeated cross-section data. From my superficial glimpse of the data set, I think we may have enough data points to treat it as panel data. We can just consider the small prime firms whose projects last over the implementation of quickpay. For a project of a small business, we would have multiple observations. If we aggregate data on a finer time scale, e.g., monthly, so that there are several observations before and after treatment (implementation of quickpay), then the interaction term between time and the implementation of quickpay is the effect we look for. Does this make sense?

This way we don't need to bother with the parallel trend assumption. This also makes it possible to expand the time horizon till 2017. We may find some interesting things due to the addition and removal of large firms in the quickpay reform.

[JNing0]

Actually, as a first step, maybe we can simply do repeated cross section with just the small businesses? We can aggregate the data to a time scale so that there are several observations before and after treatment. This makes "time" a continuous co-variate in the regression. We can run regression and simply look at the interaction term between time and the dummy variable of treatment? This way we tease out the trend in the regression automatically so we no longer need the large businesses to estimate the trend and use the parallel trend assumption?

[vibhuti6]

Hi Jie, I think we still need to have large businesses and parallel trends -- details in my response to your comment here.

But I think it is a good idea to look only at the contracts that were active both before and after quickpay. I included the analysis for this subsample with quarterly delays here -- please see the last regression table. Thanks!

[vob2]

Completed

[vob2]

I reopened this issue because I have a related question: Did we end up running regressions with the sample restricted only to projects that existed prior to QP law? If so, in DiD regression (eq (18) in the paper), are you using both X and X*Post? Values of X do not change over time. Does this create collinearity?

[vibhuti6]

Hi Vlad, we had checked that the results hold for this subsample but it's not currently included in the paper.