bar graph diff-in-diff between subjects

[Frieder Neunhoeffer]

Dear Statalists,

Given the following dataset

ID    var1    treatment    subtreatment
1    3    1    1
2    5    1    1
3    4    1    2
4    6    1    2
5    3    2    1
6    5    2    1
7    6    2    2
8    4    2    2
9    1    2    2

I would like to create a bar graph that compares the treatment effect as the difference in var1 between subtreatments. I know the following command doesn't conform with the row-wise data structure in Stata, but I would like to have something like:

cibar (mean(var1) if treatment==1) - (mean(var1) if treatment==2), over(subtreatment

I'm grateful for any ideas,
Frieder

[Nick Cox]

Where does that command come from (FAQ Advice #12) and what does the result look like?

stripplot from SSC can show the data and means and confidence intervals too.

clear 
input ID    var1    treatment    subtreatment
1    3    1    1
2    5    1    1
3    4    1    2
4    6    1    2
5    3    2    1
6    5    2    1
7    6    2    2
8    4    2    2
9    1    2    2
end 

set scheme s1color 
stripplot var1, over(subtreatment) by(treatment, note("")) bar cumul cumprob vertical height(0.2) xline(1.5, lw(thin) lc(gs8)) yla(, ang(h)) subtitle(Treatment)

If the real dataset is much larger, height(0.2) should be deleted.

[Frieder Neunhoeffer]

I appreciate the suggestion, Nick. But it is not exactly what I'm looking for.

cibar is from SSC. I'm using it because it comes with confidence intervals and suitable coloring.

My question was based on the following command:

cibar var1, over(treatment subtreatment)

which creates this graph

What I'm after is this graph structure with just one bar for each subtreatment that represents the corresponding difference between treatment 1 and treatment 2. Ideally, I would also get the confidence intervals for the computed difference in means.

[Nick Cox]

That's the much detested detonator plot or dynamite plot or plunger plot.

See commentary at e.g.

http://emdbolker.wikidot.com/blog:dynamite

https://simplystatistics.org/posts/2...lots-must-die/

You want something different -- and certainly stripplot doesn't support it and I don't know what does.

I fear that you need to write your own code.

[Nick Cox]

If I understand this correctly, your differences should come out of an appropriate model fit, so the question hinges on what that would be. It could be a plain ANOVA or something else, depending on your design. I think you'll get better advice on that from other people but you'd need to give more information first.

[Frieder Neunhoeffer]

Thank you for making me aware of the general dislike for these types of graphs. I will gladly adjust. Essentially, I would like to show the difference between the mean or median statistic of treatment 1 and treatment 2 using one bar/strip for each of the two subtreatments. I want to reduce the effort for the human eye to extracting the treatment effect across subtreatments by comparing two bars/strips to each other.

In other words, I would like to visualize the "difference-in-difference" effect of two between-subject variables from the following regression model:

reg var1 i.treatment##i.subtreatment

[Nick Cox]

OK, but medians are not estimated by that model. Again, this is territory more frequently traversed by other people here.