Interpretation did.

[amalie larsen]

Hello,

I have made a difference in differences estimation. I have 3 graphs. They are supposed to help me determine parallel assumptions. But I think they look a bit odd. I have risk weighted assets, loan loss provisions and log-score. Please help me if you have suggestions, or anything to help me interpret these correct.

[Maxence Morlet]

I don't know which trends you ran, but your trends do not look parallel to me in all three graphs... There may be substantial non-random selection into treatment based on pre-existing characteristics. You may want to try community contributed commands sdid and scul.

[amalie larsen]

Hey how do I run sdid and scul? I ran linear parallel trends and prob > F = 0,89 what does this mean then? Does it mean that the parallel don't hold?

[amalie larsen]

What does this mean ''There may be substantial non-random selection into treatment based on pre-existing characteristics'' in other words?

[Jeff Wooldridge]

Based on those graphs I can believe that the null of parallel trends wasn't rejected. I don't see systematic differences, either. Trends estimated trends won't ever be exactly the same. The "selection into treatment" based on differences in levels (which there clearly are) is handled by DiD. The issue is if there's violation of PT -- and selection is based on the difference in trends -- then maybe that can be eliminated by conditioning on covariates, X. Amalie, do you have some pre-treatment control variables?

I assume your F statistic is for one of those graphs?

[Maxence Morlet]

Dear Professor Wooldridge,

Just out of curiosity (very far from the idea of questioning what you've said) I would be very interested in having your opinion on the two following issues, because I have probably misunderstood something:

- Are parallel trends tests generally credible, in the sense that they do not under-reject the null of parallel trends due to low power?

- Are there any potential drawbacks to including time-varying covariates / confounders in DiD estimation if they are also affected by the treatment? In the sense that that may thwart the uncovering of the unbiased ATT and shut down potential channels through which treatment affects the outcome?

[amalie larsen]

Hey Jeff, Thank so much for response. Im so happy to hear from you. Just to add some info, I am trying find out whether negative interest rate policy have an impact on banks risk-taking. To do this I have the above mentioned dependent variables. Treated group (countries with and without nirp) and time (before and after 2015) which was the year nirp was first implemented.

My control variables are: inflation, HHI, logGDP, capitalization, size, deposits, liquidity.

So as far as my graphs go, my PT is not violated ?

[Jeff Wooldridge]

Just as with any test, PT tests can suffer from low power -- so we might commit a Type II error. But what is the alternative? We can always just include heterogenous trends (in a very simply way, such as linear). Bu then the collinearity will often blow up the standard errors. So should we ignore the results of the statistical test if it means producing large standard errors? It's a tough question.

Those pictures are informative, but they do not provide us with any measure of uncertainty. I don't even know how many observations are used to generate those graphs, and so they may be very noisy. That's why we have tests -- to try to help sort it out.

There's no good answer. We'd like the trends to look perfectly parallel before the intervention, but reality is it's not always true. The Loan Loss Provision graph looks sketchy to me, but not the other two.

I'm not a big fan of controlling for time-varying covariates for the reason Maxence says. Many of the control variables can be influenced by the intervention. I'd rather see flexibly controlling for pre-treatment variables.

[amalie larsen]

Thank you so much Jeff.

There is 2300 observations each.
So from what you say here, my DID is valid enough? What does the F test provide me with then? as prob>F =0,89 does this mean RWA is not parallel?

Thank you again, im just learning this.