[George Ford]

See, e.g., Lechner, about non-linear models and DID. It is generally frowned upon and the DID coefficient is biased.

Lechner offers some alternatives.

https://michael-lechner.eu/ml_pdf/journals/2011_Lechner_DiD_2011_ECO%200403%20Lechner_darf%20aufs%20Netz.pdf

[Sagara Ann]

Thank you! Upon your recommendation, I've created the variable post_q2016 which is defined by the interaction between the dummy for the post period and `q2016', which takes the values 1 to 5 corresponding to each quintile. Using this variable and the code you suggested, I've obtained the following results:




Please confirm whether this is the right way to interpret these estimates:

(For the first row of 0.post_q2016): Relative to the highest quintile, households in the lower 4 quintiles were 1.1 percentage points less likely to report outcome 1 (Better) in the pre-period.
(For the first row of 1.post_q2016): Relative to the highest quintile, households in the lowest quintile were 2.07 percentage points less likely to report outcome 1 (Better) in the post-period.

Also, is this regression equivalent to the following specification: