Treatment Intensity

[vibhuti6]

Hi everyone,

I have been wondering if we should have a subsection on treatment intensity when we present our main results. This will allow us to show that contractors who are more affected by QuickPay have a greater delay on their projects after the reform. I am thinking of a model slightly different from what we did recently. Specifically, a model closer to what Barrot & Nanda have, see excerpt below.

• I specified a similar model by defining treatment intensity as the contractor's budget from small projects relative to total budget before QuickPay. Please see below.

Please let me know what you think. Thanks!

[vob2]

That is OK with me. The only worry is what is that we have two competing specifications and we need to explain why we use contract-level analysis for most of the paper, not the contractor-level analysis. Can we defend/explain our choice?

[vibhuti6]

Thank you, Vlad. All our results hold if we do the analysis at the contractor level (and we do mention it in the paper). But we had decided a couple of months back to change the tables to include all treated projects, and only selected control ones. I think one benefit of the latter approach is that we retain more observations. It also helps study the spillover/sequencing effect.

As for treatment intensity, I think in some sense the analysis can still be at the project level. A more accurate definition may be the following:

• For contractor $k$, define $Relative Budget_k = \text{(Budget from small projects before QP)/(Total budget before QP)}$
• For project $i$ held by contractor $k$, define $TreatIntensity_i = Treat_i \times RelativeBudget_k$

I had run the results with this definition, and there is no qualitative change from previous table. Using relative budget as a measure may help with interpretation as we are able to assign a dollar value to having more small projects.

[vob2]

I am OK either way. My only concern is minimizing the number of specifications. The more different specifications we use, the more we need to defend the choices.

By the way, at the project level TreatIntensity as given above might have a problem.

A project that has a lot of concurrent small projects vs a project that has a few concurrent large projects so that the budgets are the same will have the same TreatIntensity. However, there are reasons to believe that concurrent large and small projects should produce different effects. This nuance is lost in the RelativeBudget variable.

This is not an issue if we do all analyses at the contractor level.

[JNing0]

Thank you, Vibhuti. I also have some concerns about the using the relative budget as treatment intensity.

Would it make sense to directly use project budget, or possibly log(budget), as the treatment intensity? It seems plausible that the more value at stake, the more effect of payment acceleration. This way we also have a project-level intensity. We'll need to restrict our sample to small projects that start before QP and, for ones that start after QP, we can only include non-competitively awarded projects.

[vob2]

That is a good measure. I[small business]*ln(budget). Only one concern. Budget is endogenous. As QP delays the projects it can also change the budget (yes, I remember that we do not see the effect, but this could be econometrics).

We can always take ln(budget) just prior to QP as the measure, this should address concerns about endogenous values.