Difference in Differences with cross sectional data - continuous treatment (didregress)

[Shiva Jayanth]

Hello Members,

I have a pressing doubt in implementing the diff in diff in Stata 17. - didregress command.

My research question explores the impact of mobile phone usage on the crop yield of farmers. In this model, I have a continuous outcome (crop yield of farmers), treatment variable (usage of phones), few covariates (farm level variables), Time variable (crop season A and crop season B) and group variable - ??

The didregress requires the following command for cross sectional - repeated measures data -
"didregress (y1 x1 x2) (treat1), group(grpvar1) time(tvar)"

In the current case, I have crop yield data (outcome variable) of 1000 farmers in two crop seasons (time-variable), whereas the treatment is the usage of phones (measured in number of minutes over the study period). Given the above command, my research question, and the data available, I have two doubts -

1. What is the group variable for the didregress model in this case? As you see, I am not giving intervention based on any groups. I thought each farmer was his/her own group. But I got errors while executing this
2. How do I implement continuous treatment variable in didregress? I know that we need to specify "(treatment_variable, continuous)" but that doesn't seem to work and throws errors.

Can someone please help me in this?
Regards

[Enrique Pinzon (StataCorp)]

Dear Shiva,

In the traditional setup (2 by 2) people would think about a treatment group, a period before the intervention and a period after the intervention, and an interaction of the two. You would fit something like:

regress outcome i.treatment_group i.post_period i.postperiod#i.treatment_group

Here some individuals belong to the treatment group and some do not. A typical example would be a repeated cross section of states where individuals in some states receive the treatment and some do not. Individuals in your sample should belong to a treatment or a control group.

-didregress- needs a time variable, which you have in your crop season variable, a group variable, perhaps farmers in certain regions are treated, and an indicator for observations that are treated. This indicator usually correspond to the interaction of post_period and belonging to the treatment_group, something like treated = 1.postperiod*1.treatment_group. You would estimate something like:

didregress (outcome x1 x2 x3) (treated), group(region) time(crop_period)

If treatment is continuous, you would type

didregress (outcome x1 x2 x3) (treated, continuous), group(region) time(crop_period)

[Mat Jock]

Dear Shiva,

In the traditional setup (2 by 2) people would think about a treatment group, a period before the intervention and a period after the intervention, and an interaction of the two. You would fit something like:

regress outcome i.treatment_group i.post_period i.postperiod#i.treatment_group

Here some individuals belong to the treatment group and some do not. A typical example would be a repeated cross section of states where individuals in some states receive the treatment and some do not. Individuals in your sample should belong to a treatment or a control group.

-didregress- needs a time variable, which you have in your crop season variable, a group variable, perhaps farmers in certain regions are treated, and an indicator for observations that are treated. This indicator usually correspond to the interaction of post_period and belonging to the treatment_group, something like treated = 1.postperiod*1.treatment_group. You would estimate something like:

didregress (outcome x1 x2 x3) (treated), group(region) time(crop_period)

If treatment is continuous, you would type

didregress (outcome x1 x2 x3) (treated, continuous), group(region) time(crop_period)

Dear Enrique,

Can you please give the underlying principle of what basically does

didregress

do? Like writing the maths formula or providing a traditional

regress

code line that's equivalent to the new

didregress

command.

Thanks.

[Enrique Pinzon (StataCorp)]

Hi Mat,

Please take a look at the first section of the Methods and Formulas in https://www.stata.com/manuals/tedidregress.pdf . We provide a discussion of how -areg- maps to -didregress-

[Alex Weng]

If treatment is continuous, you would type

didregress (outcome x1 x2 x3) (treated, continuous), group(region) time(crop_period)

@ Enrique Pinzon, I looked the didregress help manual, the description is not very detailed about how to understand the treatment effect of a continuous treatment variable.
I was hoping to have a paper to read to understand it more, and it looks like the papers that the manual cited are predominately binary treatment.

So my question is: what does the treatment coefficient measure? if the continuous treatment is a percentage, does it measure the effect of a one percent increase in treatment intensity on the outcome variables, compared to never treated individual or (never treated individuals + individual with a lower treated intensity)?
Can I find a paper that this the continuous treatment method this manual is based on?

Thank you so much for your help,
Alex

[Jared Greathouse]

Continous DD is a rapidly advancing topic. Whatever paper you find about it will likely be written by and for econometricians. If that's cool with you, read this. Alex Wang