Diff in Diff: pre-post and treated variables (repeated cross-sectional data)

[Passant Aboubakr]

Dear all,

I am posting here as a last resort. I am very doubtful about the viability of diff in diff analysis for the dataset I am working on, and any insights would be greatly appreciated.

I am trying to assess the impact of a global event that occurred in 2015 on a country's domestic environmental policies in terms of environmental policy stringency and policy type. The dataset's type is repeated cross-sectional, and the time frame of the data is 2009 - 2019. In total, I have 1339 observations.

Obviously, the pre-post variable would be before and after 2015. The treated variable is coded the same way as the pre-post variable, since the treatment effect depends only on the time factor. Unsurprisingly, this causes collinearity problems when I create the interaction variable.

Here is how I coded the variables and an example of the dataset:

gen pre-post = (Year >= 2015) if !missing(Year)

[gen treated = ( Group == "B")

Note that Group B are the set of policies introduced from 2015 to 2019.


input int(Number Year) str8 PolicyNumber str1 Group double Policy Stringency str4 PolicyType
1 2009 "2009-001" "A" 48.326923 "CCEP"
2 2009 "2009-002" "A" 153.933747 "CCEP"
3 2009 "2009-003" "A" 33.77584783 "CCEP"
4 2009 "2009-004" "A" 60.460317 "CCEP"
5 2009 "2009-005" "A" 50.3101741 "MBEP"
6 2009 "2009-006" "A" 63.12719533 "CCEP"
7 2009 "2009-007" "A" 78.29167967 "MBEP"
8 2009 "2009-008" "A" 38.09122081 "MBEP"
9 2009 "2009-009" "A" 31.11311381 "PPEP"
10 2009 "2009-010" "A" 27.268017 "CCEP"
11 2009 "2009-011" "A" 61.28461689 "MBEP"
12 2009 "2009-012" "A" 63.63442233 "PPEP"
13 2009 "2009-013" "A" 35.255703 "CCEP"
14 2009 "2009-014" "A" 44.22906543 "MBEP"
15 2009 "2009-015" "A" 30.41389233 "CCEP"
16 2009 "2009-016" "A" 52.2948099 "CCEP"
17 2009 "2009-017" "A" 70.53453633 "MBEP"
18 2009 "2009-018" "A" 52.586475 "MBEP"
19 2009 "2009-019" "A" 46.08626126 "MBEP"
20 2009 "2009-020" "A" 84.01568167 "CCEP"

Forgive me if this seems like a basic question. Is there anyway where the treatment effect could be driven only by the time factor in the diff in diff analysis?

[Clyde Schechter]

This data is not suitable for diff-in-diff analysis. The data cannot possibly identify a distinct treatment and time effect since the two are exactly the same thing. You either need a control group that is never treated in the data set (classic DID), or you need different groups adopting the policy at different times (generalized DID).

What you have is suitable only for a pre-post comparison, which is a weak design because it is confounds the treatment effect with other events that occur at the same time as the onset of treatment.

[Passant Aboubakr]

Many thanks for you reply.