interpreting xtlogit result

[Fathima Salih]

Hi,

I was to look at the impact drought interacted with male has on school enrolment using panel data for 4 years (2010, 2011, 2012 and 2014). The variables are defined as follows:

Drought=1 1 if individual was exposed to a drought and 0 otherwise
male=1 if individual is male and 0 otherwise
enrolled=1 if enrolled in school and 0 otherwise

I use a fixed effects panel model.

. 
xtset pidlink year
xtlogit enrol 1.drought0#1.male  $controls, fe
note: multiple positive outcomes within groups encountered.
note: 6,669 groups (11,755 obs) omitted because of all positive or
      all negative outcomes.


Conditional fixed-effects logistic regression        Number of obs    =  4,615
Group variable: pidlink                              Number of groups =  1,571

                                                     Obs per group:
                                                                  min =      2
                                                                  avg =    2.9
                                                                  max =      4

                                                     LR chi2(12)      = 101.90
Log likelihood = -1620.5449                          Prob > chi2      = 0.0000


--------------------------------------------------------------------------------
      enrolled | Coefficient  Std. err.      z    P>|z|     [95% conf. interval]
---------------+----------------------------------------------------------------
 drought0#male |
       1#Male  |   .4243507   .1396314     3.04   0.002     .1506782    .6980232
           age |  -.1015805   .0450608    -2.25   0.024    -.1898981   -.0132629
   _Iyear_2010 |   -.548134   .3994242    -1.37   0.170    -1.330991    .2347231
   _Iyear_2011 |  -.4042592   .2737263    -1.48   0.140    -.9407528    .1322344
   _Iyear_2012 |   -.364275   .1627985    -2.24   0.025    -.6833542   -.0451958

--------------------------------------------------------------------------------

How exactly do I interpret the coefficient on drought*male? Is this correct: being male and experiencing a drought increases the probability of being enrolled at school, compared to females. Also, what exactly is the probability? I read elsewhere that the -margins- command cannot be used here.

thanks!

[Matthew Berg]

Hello Fathima,

Odds ratios provide a good starting point for interpretation of the coefficients. You could use:

 xtlogit enrol 1.drought0#1.male  $controls, fe or

Maybe this quote from Carlo might be important to you:,

As an aside, please be advised that under -xtlogit, fe- you will get coditional fixed effects, which are different from the ones you can get under -xtreg, fe-.

I don't know which literature you refer to, but margins can be used after xtlogit. You could post your code or error messages so that we can provide more information.

After fitting your model, you can use

margins, dydx(*) post

I hope this helps, and that more experienced users will chime in.

[Joao Santos Silva]

Dear Matthew Berg,

Using margins like that computes partial effects assuming that the fixed effects are zero, and there is no reason to assume that. So, those results are meaningless, as I have illustrated here. I guess this is what Fathima Salih is referring to.

Best wishes,

Joao

[Matthew Berg]

Dear Joao Santos Silva

Thank you very much for this insight!

[Jeff Wooldridge]

A few things to add. First, I think the drought variable needs to appear on its own, otherwise the effect of a drought on attending school that's common across genders is not being allowed. This can easily bias the estimated difference between male and female. There may be none once you include drought.

Second, I suspect the variable "drought" varies at something like a village or regional level. It may be enough to include dummy variables at the village or regional level, rather than removing individual FEs. Then you can just use logit, cluster the standard errors (maybe at the level of the drought measurement), and use margins. If you try this, then male should also appear on its own as it won't be eliminated by using higher-level fixed effects.

[Fathima Salih]

Jeff Wooldridge thank you. The "drought" variable is at the household level given the data is geo-referenced. This is useful as around 50% of the households moved location between the waves. The "drought" variable can be included in the xtlogit estimation.

Is it correct to pool the 4 waves of data and estimate a logit regression when the data are longitudinal, following the same individual in 2010, 2011, 2012 and 2014?

If I use xtlogit, how do I interpret the results and calculate probabilities given that margins cannot be used as noted by Joao Santos Silver?

Thanks a lot!