Lag variable with gaps in timevar

[Federico Ricca]

Hi,

I am currently using an unbalanced panel with gaps in the timevar (it's a panel on parties in cabinets, with timevar the start_date of the cabinet recorded in %td, hence I have no consecutive timevar observation, they are all separated by gaps). I would like to create a simple lag, but it keeps generating all missing values. Is there a way to lag with respect to the previous, non missing value available?

Thanks,
Federico

[Nick Cox]

What defines a panel here? It's perfectly possible to create your own pseudo-time variable as here, so each time occurs just once for each panel,

clear 
input id time 
1 2 
1 3 
1 5 
2 7
2 11
2 13
end 

bysort id (time) : gen counter = _n 

tsset id counter 

list, sepby(id)

-- but there is a price to be paid for this in ignoring gaps and even asynchrony of panels.

[Federico Ricca]

Thanks, that could be a way out. I was hoping to do that without losing the (proper) time dimension. Do you think there is no way to lag back to the last non missing value (ignoring somehow the gaps)?

[Nick Cox]

You can always refer to the last non-missing value using subscripting [_n-1] If you have missing values, see also http://www.stata.com/support/faqs/da...issing-values/

[Phil Bromiley]

With xtset, the lag operator won't lag across panels. If you go the [_n-1] etc route, you can lag across panels, so you'll probably want to add an if panel==panel[_n-1] condition.

[Nick Cox]

bysort id (counter) : gen change = whatever - whatever[_n-1]

is a common pattern to insist on all comparisons being within the same panel.

[Radhouene DOGGUI]

Hi Nick, please what is the difference between creating manually a lagged variable by doing this bmilag1 = bmi[_n-1] and use this opeartor L1.bmi ?

[Radhouene DOGGUI]

bysort id (counter) : gen change = whatever - whatever[_n-1]

is a common pattern to insist on all comparisons being within the same panel.

Hi Nick Cox , please what is the difference between creating manually a lagged variable by doing this bmilag1 = bmi[_n-1] and the use this opeartor L1.bmi ?

[Nick Cox]

It's the difference between the previous value and the value one time unit before, which is no difference if there are no gaps.

[Radhouene DOGGUI]

It's the difference between the previous value and the value one time unit before, which is no difference if there are no gaps.

Thank you Nick Cox

[Pandelis Andreou]

Nick Cox

I have a similar problem: unbalanced panel dataset, T = 120, n = 3.

There is only one variable, say X, that I need it in lagged values. This has the same value for all n, at each t.

Therefore, it's possible to calculate lags in each n.

However, since n =2, 3 have gaps, I lose about half of observations.

Is it possible to calculate the lagged value of X for n 2, 3, even if there are gaps?

I guess I can do it manually in Excel, but that would take a lot of time.

Is there ANY way to do this in stata?

[Nick Cox]

Sorry, but I can't follow #11 at all. Please give a data example.

[Pandelis Andreou]

Year	Order	Age	AEA
2010	First	69	10
2010	Second	58	10
2010	Third	71	10
2011	First	59	15
2011	Second	65	15
2012	First	62	20
2012	Second	62	20
2013	First	61	25
2013	Second	60	25
2013	Third	60	25
2014	First	59	30
2015	First	58	35
2016	First	57	40
2016	Second	57	40
2017	First	56	45
2018	First	55	50
2018	Second	55	50
2019	First	54	55
2019	Second	53	55
2019	Third	53	55

Random values. The dataset refers to the age of winners in an prize where every year can be 1 to 3 winners.

Therefore time variable (t) = Year. Cross-section variable (n) = Order.

Say I want to estimate the following panel data model using the data above:

Age n, t = a + b * AEA n, t-1

t = year
n = order

As it is, using the lag operator loses a lot of observations because e.g. for observations where order = second and there was not a second observation in the previous year (e.g. 2018 and 2017) then there would be a missing observation. AEA 2, 2017 = .

In that manner, I lose a lot of observations.

I ask if there is a way for Stata to register AEA 2, 2017 = 45.

[Nick Cox]

Thanks for the data example. Treating this even as panel data seems a real stretch, but if there was no awardee in a given year, then there is nothing to impute, interpolate, infer, or estimate.

[Pandelis Andreou]

Thank you for the reply.

There is always at least an awardee in every year.

I treat this as an unbalanced panel data set, that can be either modelled as a random effects model or a pooled OLS regression (I will use.Breusch-Pagan and Hausman tests for the final selection).

However, no matter what model I would choose, the problem of losing observations will remain: I would always lose observations if use the lagged value of AEA.

What I want is for lagged value of AEA to be recorded also for observations of second and third. awardee, not just the first (in every year).

Is it possible to that in Stata? Or do I have to compute them manually in Excel?