Adding independent variables to xtlogit

[Jordi Imbrechts]

Dear all,

For my thesis I have to research the causes of layoffs for (big) Belgian firms. Period 2011-2020: unbalanced panel data, but with gaps. Around 98k observations for layoff variable (dependent var).

Question:
To be able to give an adequate answer to my research question, I investigate 4 hypotheses. For the last 2 hypotheses I use seperate models. But for my first two hypotheses, "a decrease in productivity will increase the chance of a layoff" & "a decrease in profitability will increase the chance of a layoff", I was using just one xtlogit model:

Xtlogit layoff decreaseproductivity1y decreaseROA1y (control variables), fe

I was just wondering if this is the correct methodology that I'm following, or should I use seperate models for productivity and profitability aswell (just like I'm doing for my 3rth and 4th hypothesis)?

Kind regards,
Jordi

[Carlo Lazzaro]

Jordi:
set aside that is -xtlogit- and not -Xtlogit-, I fear that your first code may suffer from endogeneity (reverse causation) if the chance of layoff can actually decreases productivity.
What's your supervisor's opinion about your approach(es)?

[Maxence Morlet]

As an aside to Carlo’s great point, with logit the interpretability of the coefficients is a lot more complicated than with OLS’s Linear Probability Model, seen as here you are modeling the outcome as a log-odds ratio.

Be sure to then report accurate marginal effects, as the coefficient you will obtain on “decreaseproductivity” cannot be interpreted as the impact of a decrease in productivity on the probability of dismissal per se.

[Jordi Imbrechts]

Carlo:

Thanks for the comment. My bad, I meant -xtlogit- ofc. Supervisor did not give me any direct feedback about this. I also can't ask for any more feedback. Final deadline is tommorow (which is ok, because I have everything ready in my do-file), but this is just one last thing that I'm wondering if it is correct. I also feared that my first code is not correct, but I do not completely understand your comment. If I remove 'decreaseROA' from the first code, will the risk of reverse causation be minimalised?
But on the other hand I do understand your remark: from other research I learned that the survivor's syndrome may lead to a decrease in productivity. Still, not quite clear for me to remove 'decreaseROA1y' and to make a seperate model for this hypothesis.

Maxence:

Thanks for the comment. But I'm not allowed to interpret all the coëfficiënt as an odds-ratio (my research is no "exact science" according to my supervisor), the main goal for my thesis is to look at the + or - sign of the coëfficiënt and the corresponding p-values.

[Carlo Lazzaro]

Jordi:
1) removing return on assets (ROA) does not change the risk of reverse causation between layoff and decreased productivity.
That said, as you deadline is tomorrow, I'd prepare to reply to a possible issue concerning what above from discussants. The usual formula is something along the following lines: "More research is needed to delve into the relationship between layoff and decreased productivity so to rule out any endogeneity problem." Please note that endogeneity is a serious nuisance that makes your coefficients unreliable;
A reference concerning the economics of more research is needed can be found (free of charge) at: https://pubmed.ncbi.nlm.nih.gov/11511601/.
2) FWIW, I respectfully disagree with your supervisor on p-value relevance without considering 95% CIs (see, free of charge, https://pubmed.ncbi.nlm.nih.gov/29348701/).

[Jordi Imbrechts]

Carlo,

Thank you very much. My thesis is finished. If you ever come to Belgium, I'll provide you with some cold Stella Artois'.

Kind regards,
Jordi

[Carlo Lazzaro]

Jordi:
I'm teetotaller, but thanks anyway.
Finger crossed for your discussion!

[Jordi Imbrechts]

Thanks!