Creating an Attrition Dummy in Panel Data

[Charles Ackah]

Dear all,

I am working with an unbalanced panel data comprising five waves. I intend to generate a dummy variable identifying individuals (pid) who were observed only in the first two waves (Waves 1 and 2) but not in any subsequent waves (Waves 3, 4, or 5). The dummy variable takes a value of 1 for respondents interviewed in the first two waves who attrited from the panel thereafter, and 0 otherwise.

* Example generated by -dataex-. To install: ssc install dataex
clear
input long pid float wave
301011 1
301011 2
301011 3
301011 4
301011 5
301012 1
301012 2
301012 3
301012 4
301012 5
301013 1
301013 2
301013 3
301013 4
301013 5
301014 1
301014 2
301014 3
301014 4
301014 5
301015 1
301015 2
301015 3
301015 4
301015 5
301016 1
301016 2
301016 3
301016 4
301016 5
301017 1
301017 2
301017 3
301017 4
301017 5
301018 1
301018 2
301018 3
301018 4
301018 5
301019 1
301019 2
301019 3
301019 4
301019 5
301020 1
301020 2
301020 3
301020 4
301020 5
301022 1
301022 2
301022 3
301022 4
301022 5
301023 1
301023 2
301023 3
301023 4
301023 5
301024 1
301024 2
301024 3
301024 4
301024 5
301025 1
301025 2
301025 3
301025 4
301025 5
301026 1
301026 2
301026 3
301026 4
301026 5
301027 1
301027 2
301027 3
301027 4
301027 5
301028 1
301028 2
301028 3
301028 4
301028 5
301029 1
301029 2
301029 3
301029 4
301029 5
301031 1
301031 2
301031 3
301031 4
301031 5
301033 1
301033 2
301033 3
301033 4
301033 5
301034 1
301034 2
301034 3
301034 4
301034 5
301035 1
301035 2
301035 3
301035 4
301035 5
301036 1
301036 2
301036 3
301036 4
301036 5
301038 1
301038 2
301038 3
301038 4
301038 5
301040 1
301040 2
301040 3
301040 4
301040 5
301041 1
301041 2
301041 3
301041 4
301041 5
301043 1
301043 2
301043 3
301043 4
301043 5
301045 1
301045 2
301045 3
301045 4
301045 5
301047 1
301047 2
301047 3
301047 4
301047 5
301048 1
301048 2
301048 3
301048 4
301048 5
301049 1
301049 2
301049 3
301049 4
301049 5
301050 1
301050 2
301050 3
301050 4
301050 5
301051 1
301051 2
301051 3
301051 4
301051 5
301052 1
301052 2
301052 3
301052 4
301052 5
301053 1
301053 2
301053 3
301053 4
301053 5
301054 1
301054 2
301054 3
301054 4
301054 5
301055 1
301055 2
301055 3
301055 4
301055 5
301056 1
301056 2
301056 3
301056 4
301056 5
301057 1
301057 2
301057 3
301057 4
301057 5
301058 1
301058 2
301058 3
301058 4
301058 5
end

[Andrew Musau]

Assuming only observed waves are included in the dataset:

bys pid (wave): gen wanted= wave[1]==1 & wave[_N]==2

[Charles Ackah]

bys pid (wave): gen wanted= wave[1]==1 & wave[_N]==2

. ta wanted

wanted | Freq. Percent Cum.
------------+-----------------------------------
0 | 214,835 100.00 100.00
------------+-----------------------------------
Total | 214,835 100.00

Thanks Andrew Musau , but it doesn't appear to work.

[Carlo Lazzaro]

Charles:
if the -pid- does not have observations from wave 3 onwards, it will be simply omitted from the -e(sample)-.
So why bothering yourself about creating that dummy?

[Carlo Lazzaro]

Charles:
elaborating a bit on your data excerpt, you may want to consider instead:

bysort pid (wave): drop if _n>2 & pid==301011
bysort pid (wave): egen check=max( wave) if _n==_N
list pid check if check!=.

[Charles Ackah]

Thanks Carlo Lazzaro. I am trying to estimate the attrition probability for individuals who were successfully interviewed only in waves 1 and 2. Hence the need to create the dummy only for those who were successfully interviewed in the first two waves out of the 5 waves.

[Andrew Musau]

bys pid (wave): gen wanted= wave[1]==1 & wave[_N]==2

. ta wanted

wanted | Freq. Percent Cum.
------------+-----------------------------------
0 | 214,835 100.00 100.00
------------+-----------------------------------
Total | 214,835 100.00

Thanks Andrew Musau , but it doesn't appear to work.

As I stated, the code in #2 assumes that only the waves that are present are included in the dataset. If you have the identifier and wave variables present for all waves for each individual, but other variables are missing for unobserved waves, then you will first need to drop the unobserved waves before running #2. As Carlo states, the unobserved waves do not enter the estimation sample, so dropping them does not affect anything. Assuming that your outcome variable is named outcomevar (replace with the name of your outcome variable), you can start by running:

drop if missing(outcomevar)

followed by #2.