How do I need to scale variables before I check the DiD parallel trends assumption? I have a panel data set over several years (2010-2017). In that figure I have countries that I grouped as control (not receiving the treatment) and others as treatment (receiving the treatment).
The variable of interest is net margin.
In the control group and the treatment group there is large heterogeneity between the size of assets of banks in each country. E.g relative share of assets to total assets of German banks is 20% whereas Estonia is 0.04%.
If I would just plot this in R ggplot it would look like this:
margins_eurozone_avg <- assets_year %>% 
  group_by(fiscalyear)%>% 
  summarize(net_margin=mean(net_margin))

ggplot(margins_eurozone_avg, aes(x=fiscalyear, y=net_margin))+
  geom_line()


I could do the same for my control group and plot them side by side.
But this chart is deceiving to me since it totally misinterprets the data in the sample.
I've nowhere found any explanation online how to deal with a problem like this. Should I just scale each net income figure by asset value, would that make sense?
 A: It's not totally clear to me what the problem is, but I'm assuming there is heterogeneity in net margin between observations.
In your DiD model, the between-observation heterogeneity is generally (or can be) removed either by demeaning or differencing. It is still possible that heterogeneity between observations can pose an issue, as within-observation, annual changes could have a larger variance for observations with greater panel-average levels of net margin.
For example, if Germany's bank's net margin is consistently much greater than Estonia's over the panel, and banks in Germany and Estonia both increased their net margin by 10% (of net margin) on average during each of the pre-treatment years, the raw values would appear as a non-parallel trend, as the change in net margin would have a much steeper slope for Germany than Estonia (as raw net margin). However, if you converted to annual percent change in net margin, this would appear as a parallel trend. Change in logged values is often a good proxy for percent change.
Net margin is already a ratio. Adding an additional scaling component would mean you are no longer working with net margin, right? If converting to percent change introduces bias or still does not satisfy parallel trends, perhaps your treatment and control groups do not have pre-treatment parallel trends. The did package let's you test whether a parallel trend exists conditional on covariates if that is the case.
