logisitic regression for bankruptcy prediction with time fixed effects

[Ali Rezaei]

Hello everyone,

I have a unbalanced sample of 2096 firms for 30 years (around 24000 firm-year observation). Number of bankrupt firms is 288. I am trying to estimate bankruptcy probability using conditional logit model.
When I control for both firm and time effects p-value are equal to 1. But controlling for firm effects alone produces consistent results.
I appreciate any suggestion on how to control for time fixed effects in xtlogit model.

[Carlo Lazzaro]

Ali:
you may want to try:

xtlogit <depvar> <otherpredictors> i.time, fe

Please note that:
1) you cannot add -panelvar- among the set of predictors, as it

outcome does not vary in any group
r(2000);

2) with 2096 firms, default standard error is not appropriate (you did not provide any detail about this issue, so I make a bit of prevention here);
3) last but not least, due to incidental parameters bias (http://www.econ.brown.edu/Faculty/To...meters1948.pdf), -xtlogit,fe- gives back conditional -fe-.

[Ali Rezaei]

Thanks Carlo Lazzaro ,

Aside from default standard errors problem , I have a set of problems in model :

Outcome variable is equal to 1 if a firm went bankrupt otherwise it is zero. By construct, xtlogit , fe keeps those firm with outcome variable changed at least one time. In my data, this means, it only uses bankrupt firms' observation.
This, in turn means, limited number of firms per year as shown in the table below.
My question is do I require to control for year fixed effect when there are limited number of observations per year (This also applies to state fixed effects)?


The beginning year of sample is 1991 and end year is 2021. Firms enter to and exit form sample in any year from 1991- to 2020.

ear	# of firms	State	# of firms
1991	18	AL	2
1992	13	AZ	2
1993	7	BC	2
1994	13	CA	1
1995	9	CO	12
1996	7	CT	3
1997	9	DE	14
1998	11	FL	3
1999	10	GA	7
2000	18	IL	8
2001	14	IN	18
2002	18	LA	4
2003	14	MA	2
2004	9	MD	3
2005	9	MI	1
2006	8	MN	25
2007	11	MO	7
2008	8	NC	20
2009	13	NH	7
2010	1	NJ	1
2011	3	NV	14
2012	4	NY	1
2013	4	OH	17
2014	2	OK	21
2015	1	ON	2
2016	2	PA	1
2018	2	QC	9
2019	4	SC	2
2020	0	TN	6
2021	1	TX	1
		VA	16
		VT	4
		WI	1
		WY	5

[Carlo Lazzaro]

Ali:
1) that's the way (conditional) fixed effect estimator works: time invariant variables are wiped out;
2) including -i.year- in the right hand side of your regression equation is actually required when gong -fe-;
3) if -i.industry- is a time- nvariant predictors, what's the gain with including it among your predictors?

[Ali Rezaei]

Carlo Lazzaro Thank you for your comments on this topic.
You are right when time fixed effects are not varying during the sample period their inclusion in the model is none-sense.