Post estimation tests in Difference in Difference analysis of unbalanced panel data

[Pranshu Tripathi]

Dear Members

I am working on the impact of a law enacted in 2016 on firms' leverage with unbalanced panel data.

The treatment group variable Dividend_dummy (i.e., Dividend-paying firms vs. non-dividend-paying firms) is available for all the firms (This variable is balanced).

Also, I have 2 years of data for both pre and post-treatment periods (which means that other than Dividend_dummy rest of the variables are unbalanced).


I am performing DiD and PSM DiD for the analysis purpose



1. For Did

Why is Stata not running post-estimation tests like estat ptrends and estat granger?
It shows error

treatment assignments times vary; not allowed with estat granger

2. For PSM DiD

When I apply psmatch2 with the logit model with any of propensity matching, I get very few observations. (Which was around 26,000 for plain DiD whereas only 2852 for PSM DiD)

Thanks

[George Ford]

what's your didregress command?

how did you define the Dividend_dummy variable?

[Pranshu Tripathi]

Hey George,
Sorry for being late to reply

what's your didregress command?

It is :

xtdidregress (sd1 laggedgrowth laggedsize laggedtang laggedroa laggedliquid laggedage dividend_dummy ibc i.t##i.iid) (did), group(dividend_dummy) time(t) vce(robust)

&
xtreg sd1 did ibc laggedgrowth laggedsize laggedtang laggedroa laggedage dividend_dummy i.t##i.iid,fe vce(robust)

How did you define the Dividend_dummy variable?

The dividend dummy is defined (for 2016) : if the firm had not paid any dividend in the previous four years. It makes this variable time-invarying.

Also, I found no difference when I checked for no. of observations for this variable for all the years.







Also, I have got the answer for PSM DiD. So, the question is for only DiD.

Thanks and Regards

[George Ford]

What is the treatment? How is the treatment date chosen?

[Pranshu Tripathi]

2016 is the treatment year. This year, the law was enacted.

2011-2015 is assigned a value of 0, and 2016-2023 is assigned a value of 1.

So, I made the variable did (treatment effect)= ibc (law)*dividend_dummy (group dummy)


and sd1 is the dependent variable
Thanks

[George Ford]

what's your thinking on " i.t##i.iid". Is that a full interaction in N and T?

[Pranshu Tripathi]

It is interaction between time and industry-fixed effect.

It gives me i.t (time effect) , and i.t*i.iid (time varying industry effect) and drops i.iid (being fixed effect).





Is unbalanced panel data an issue or there is problem i n defining the group dummy, treatment dummy and treatment effect variable?

[George Ford]

How did you xtset your data for xtreg?

[George Ford]

Did the 2016 law apply only to firms that pay dividends?

[George Ford]

What do you mean by balanced and unbalanced?

Is this a collection of firms within industries, where the firms in the industries are not the same in each year?

[Pranshu Tripathi]

How did you xtset your data for xtreg?

The panel is xtset on firm id and time.

xtset firmid t

Did the 2016 law apply only to firms that pay dividends?

Theoretically no. However, my research design (quasi-experimental design) is such that it is more important for non-dividend paying firms. So, I have created the dummy where a firm is considered a non-dividend-paying firm in 2016 if it has not paid any dividend in the previous four years; otherwise, it is a dividend-paying firm. if a firm is a non-dividend paying firm in 2016, then it is considered as a non-dividend firm for the whole study, and similarly, we construct the dividend-paying firm. It makes this variable time invarying.

What do you mean by balanced and unbalanced?

By balanced, I mean that the firms should have non-missing data for all the years, whereas by unbalanced data, I mean that the firms may have some years missing in their data, but I have only considered those firms in unbalanced which should have at least three years before and after the Law.

Is this a collection of firms within industries, where the firms in the industries are not the same in each year?

I have used both industry and firm fixed effect.

A firm remains in the same industry across all the years.

[Pranshu Tripathi]

I have also used industry as a control in the psmatch2 command.

[George Ford]

In xtdidregress, group should be id not dummy_variable. cluster on id. I don't think dummy_variable should be a covariate in that model.

start simple.

ssc install reghdfe

reghdfe sd1 c.law16#c.dividend_dummy , absorb(id year) cluster(id)

reghdfe sd1 c.law16#c.dividend_dummy Xvars , absorb(id year) cluster(id)

[Pranshu Tripathi]

Thanks, George.

I don't think dummy_variable should be a covariate in that model.

I agree with you about no use of the dummy variable as a covariate.

In xtdidregress, group should be id not dummy_variable. cluster on id

When I use group (firmid), it gives me a result similar to group(dividend_dummy). It's just the Number of groups and treatment time table that differs.



My question is, if I operate on unbalanced data with the xtdidregress command, can I run the estate trends and estat granger command commands?
When I try these post-estimation test commands, I get an error.

treatment assignment times vary; not allowed with estat granger

Regards

[George Ford]

could you run them when used used group(firmid)? group(dividend_dummy) is not correct.