Use of binary explanatory (control) variables in Difference-in-Difference Estimation

[Johannes Wirth]

Hi all,

I am conducting a Diff-in-Diff with the following data, which is copied in here by using the dataex command. The normal DD regression works just fine. However, I would strongly like to include control variables. All the ones that are expressed in percentages (GDP-growth, inflation, remmitances..) work fine too. The binary (dummy) ones, like membership in some supranational bodies, for instance the EU, do absolutely not. Stata always states that they're omitted because of collinearity and I do not get way this is the case, to be honest. The same variables work fine with FE and Random Effects and also in bigger samples than the one displayed below.

* Example generated by -dataex-. To install: ssc install dataex
clear
input int year str7 countryname str12 countryyear double new_avwgi byte(EU_candidate EU_member Intrastate_conflict) double(new_exports new_FDI new_GDP new_inflation new_remittances) byte(time BRI_state) float(did did_EUaccess)
2010 "Estonia" "Estonia.2010"    1.0663436849912007 0 1 0 75.07823021170658  13.302610404251492   2.259075893831181  2.97204512015696 1.8324759753382773 0 0 0 0
2011 "Estonia" "Estonia.2011"    1.0747986833254497 0 1 0 86.53972527282187   4.827558144959612   7.597300209245603   4.9819013145361 1.8919482862267252 0 0 0 0
2012 "Estonia" "Estonia.2012"    1.0621930360794067 0 1 0 85.98943858726155  7.7592920176217595   4.307258987653256  3.93339987296979 1.9998493379228786 0 0 0 0
2013 "Estonia" "Estonia.2013"     1.112437019745509 0 1 0 84.30308150402843   4.370606686465119  1.9365434443487857  2.78056658954996 2.2581279426120306 0 0 0 0
2014 "Estonia" "Estonia.2014"     1.220025469859441 0 1 0   81.802388001185   6.691461144936392  2.8885638447165007 -.106175146521716 2.0430252776096847 1 0 0 0
2015 "Estonia" "Estonia.2015"    1.1950828234354656 0 1 0 77.34500432713354 -3.1189166996698736  1.9002865897060275 -.492326006547338 1.9957658562540663 1 0 0 0
2016 "Estonia" "Estonia.2016"    1.1961500843365986 0 1 0 77.59978592650414   3.916098151538917  3.4890971114797082  .148684907627356  2.027937379229015 1 0 0 0
2017 "Estonia" "Estonia.2017"    1.1927852034568787 0 1 0 76.54409627801454   5.844004555823057   4.856686856443943  3.41723549488057  1.899560156901767 1 0 0 0
2010 "Serbia"  "Serbia.2010"     -.1452512772132953 0 0 0 32.25560609728482   4.049632331111981   .7310464430853898  6.14255360247246  9.847691855465044 0 1 0 0
2011 "Serbia"  "Serbia.2011"    -.10529540820668142 0 0 0 33.00182218001296  10.002289323791814  2.0362757259362922  11.1373976342129  8.036210678703773 0 1 0 0
2012 "Serbia"  "Serbia.2012"     -.1449378045896689 0 0 0 35.84480697786771   2.946768361558177  -.6815421330412903  7.33038589596636   8.19178067430819 0 1 0 0
2013 "Serbia"  "Serbia.2013"    -.09417816003163655 0 0 0 39.85150587567927  4.2555897728095555   2.892634162141732  7.69426362896666  8.317079992436222 0 1 0 0
2014 "Serbia"  "Serbia.2014"     .04181295136610667 1 0 0 42.07752997177007  4.2476586388978035 -1.5895061837666447  2.08244793880151  7.851232973416203 1 1 1 1
2015 "Serbia"  "Serbia.2015"    .058732101072867714 1 0 0 45.27361076131518   5.910584879417411  1.7763209728228588  1.39235822000187    8.5018044003722 1 1 1 1
2016 "Serbia"  "Serbia.2016"  .00017994269728660583 1 0 0 48.61545929836844  5.7955002005828735  3.3403290685937606  1.12231397417765  7.887580830680063 1 1 1 1
2017 "Serbia"  "Serbia.2017"    -.02473145102461179 1 0 0 50.54079107542444   6.556310285630537  2.0493105252801627  3.13106248590119  8.130079194227449 1 1 1 1
end

Can anyone help? Your help would be very much appreciated.

Best,
Hannes

[Jeff Wooldridge]

The dummies likely have no variation over time for any country, and so they drop out in the fixed effect estimation. This is as it should be. There’s a way to bring them back in and still do fixed effects on the other variables.

[Johannes Wirth]

Hi Jeff,

thank you for your quick reply! Do you know the way to bring them back by any chance? This would help me quite a lot!
I can check if there are variations by using a correlation matrix, if I am not mistaken?

Best,
Hannes

[Jeff Wooldridge]

You need to use the Mundlak device by including the time averages (probably created using -egen-) of all time-varying variables (except for year dummies). This will reproduce the fixed effects estimates on the the other variables and give you coefficients on the time-constant variables. I probably can turn up an example.

[Johannes Wirth]

Hi Jeff, that would be great if you can! Thanks for the help and your advise so far!

[Jeff Wooldridge]

This is an exercise I give in a course I teach at UCL in London every May ....

[Johannes Wirth]

Thank you Jeff! I will have a look at it