Should I use district fixed effects

[Shane Bond]

Hi, I have individual level sample data and I know whether each individual worked in manufacturing or non-manufacturing (mf =1 if manufacturing, 0 otherwise). I know the district of the person. I am looking at the impact of trade liberalization that affected the entire country. So all the districts face the effects of trade liberalization. I want to see if districts closer to seaports (say within 50 miles) had a strong effect of people shifting from non-manufacturing to manufacturing. I have data for 3 years 2000 and 2005 and 2010. Trade liberalization happened in 2002. I have calculated distance of each district from the nearest seaport and made a variable (port=1, if distance < 50 miles, 0 otherwise).

I am running the following regression:

reg mf i.port i.year i.port#i.year i.state

My primary question is if I should include district fixed effect instead of the port fixed effect?

[Maxence Morlet]

I see a few problems with your regression:

reg mf i.port i.year i.port#i.year i.state

i.year by itself is redundant. Also, I would advise you to cluster standard errors by district.

You should also give this paper a read: "A simple, robust test for choosing the level of fixed effects in linear panel data models" (Papke and Wooldridge, 2022)

Why did you dichtomise distance? I would use it as a continuous treatment, in which case I would also recommend that you read "Difference-in-Difference with Continuous Treatment", NBER working paper.

I would add district fixed effects, in which case i.state becomes presumably redundant, and then see how it affects results. How many districts do you have?

[Shane Bond]

Thanks Maxence. I am clustering standard errors by district. Can I ask why the year fixed effect is redundant? I thought I had 3 years and it will absorb some year specific shocks that is common across districts. I am looking at using distance as both continuos and with some categories (2 as it is now or more). The district fixed effect will absorb both the state and port fixed effect. Do you recommend doing this? I have 350 districts. My initial thought in the regression was that I am comparing two indviduals, in the same state, one who is in a district close to ports and one who is away, and seeing if trade liberalization had a differential impact in the probability of employment in manufacturing sector.

I have a question about clustering too. My data comes from a sample as follows. Every district in the country is sampled. Now, some villages are chosen from a district and they are sampled. Within a village, some households are chosen and all individuals are surveyed in the household. Does this mean that I can also think of clustering at the village level as it is the first sampling unit? Thanks

[Carlo Lazzaro]

Shane:
if (as it would seem fron the last line of your post #3) you have a nested study design (village nested within districts) you may want to consider -mixed-.

[Shane Bond]

Thank you Carlo. Can you please explain what you mean by mixed? How should I do it in stata? Thanks

[Carlo Lazzaro]

Shane:
see -mixed- entry and related references in Stata .pdf manual.