Autocorrelation and other assumptions Diff in Diff model

[Milla Hanzon]

Hello,

I am working on the following Diff in diff model:


ln_sales = b0 + b1*treatment + b2*(treatment*after) + d1*week1 + d2*week2 + d3*week3 +... +dN*week160

This is a DID where I'm looking at the effect on sales for four different grocery store when a dollar store opens nearby. I have the sales for four other grocery stores as a control group. The sales are average weekly sales over a time span of three years.

I have a few questions regarding this model.

I've come to understand that one of the biggest problems in DID models is autocorrelation. For instance, I can imagine that the model's errors will correlate every easter, christmas etc, as sales fluctuate differently in these time periods. However, won't the weekly dummies adjust for this? Or are there other factors I should think of?

Another question: The parallel trend assumption. This far, I have plotted the sales for the different grocery stores and treatment group to see if the assumption about parallel trends hold in the pre-treatment period. The assumption seems to hold, but I find it very hard to asses! The slope might just be a bit more steep for the treatment groups than the control groups. Are there any way to assess this in stata quantitatively?

Hope someone can help!

[Rebecca Water]

I think the headline of this thread is the opposite of what you are asking. Anyways, check out this thread for the dummy variables of your model : https://www.statalist.org/forums/forum/general-stata-discussion/general/1472164-interaction-term-in-a-quadratic-model Clyde Schechter explains how to simplify codes like yours very well.
My answers are not those of an expert and you should not rely on them 100% but here's what I know:
To account for the correlation of observations of each store, you could use

reg whatever, vce(cluster storeid)

The common trend assumption is usually defended by the author itself. Here is an example of a friend of mine who was told by the lecturer to scale the variable of interest on the value of the first observation and check visually if there is a trend. In this example it was approved as a common trend.

Red is control, Blue is treatment.
Another thing is a placebo estimation. You choose a random year as the year the start of the treatment and check whether b2 is still significant.