FIXED EFFECT MODEL, small saple size.

[Yasmin bzd]

Dear all, I am regressing the impact of Netflix subscriptions on theatrical admissions in 16 countries from 2012 to 2017.
I have heterogeneity, hence I run

xtreg logadmissions lognetflixsubscribers loggdp averageTV, fe vce (robust)
xtreg logadmissions lognetflixsubscribers loggdp averageTV, re vce (robust)

and xtoverid test to decide which one is a better fit.

The result of the test shows that FE model is a better fit p<0.05. My question is, as you can see I don't use i.year and i.country variables in my model because of small sample size and low degrees of freedom. If I don't add time fixed effect in to my model as a result of this constraint, would be wrong? What can I do as an alternative? or shall I keep my model?

(Here is my FE regression results.)

Thank you.

[Jordan Sydenham]

As far as I am aware there is no need to include time fixed affect variables unless you have some reason to believe that the relationship may have changed over time. As far as I can tell there is no reason that this should be the case within the short time frame of your data. So I believe that it is appropriate to keep the model you have already estimated.

[Yasmin bzd]

Thank you for your response Jordan

[Clyde Schechter]

I'll just add that because you have -xtset country-, the use of -xtreg, fe- automatically incorporates country fixed effects into the model. If you tried to add i.country to the list of regressors, you would find that they are all dropped due to colinearity with the automatically included country fixed effects. (The time variable, however, is not automatically added.)

[Yasmin bzd]

Thank you very much Clyde. That helped me I would like to ask another question. Something bothers me in my regression. The time period that I use is from 2012-2017. I didn't take the previous year as 1) I didn't have data 2) Netflix entered to many European markets (I have 15 European countries and the USA in my sample) from 2012 and on. I try to capture the relationship between Netflix subscription numbers and movie theatre admissions, to see if Netflix causes a stimulation or cannibalization on admission numbers.

My question is do I make a technical and logical mistake as I don't have older data before Netflix? (On a different note, in my data (16 countries, t:2012-2017) the year of Netflix's entry to countries didn't occur simultaneously ( in France in 2014, in Spain 2015 etc). Not to leave Netflix subscription numbers in these countries empty, instead of leaving it "zero" I gave them a low subscription number in my list which 4200 (inspired by Austria's Netflix subscription number in 2014, 4300, which is the actual lowest Netflix subscription number in my list of countries).

Thank you.

[Clyde Schechter]

Your question now is about the substance of modeling the effect of Netflix on movie theater admissions. I have no knowledge at all, let alone, expertise in that area, and can't advise you.

I will say, from a general statistical perspective, that using 4200 to indicate that Netflix has not entered the market yet is likely to be wrong. If Netflix isn't in the market, then the number of subscriptions should be 0. The question is whether additional variables reflecting before-after Netflix entry as a dichotomous 0/1 variable, or perhaps a variable counting the number of years since Netflix entry (and 0 before entry), or some transformation or spline based on the number of Netflix subscriptions or some other device is needed to properly capture the effects.

I think you need to discuss this with somebody who has expertise in this content area. It's really not a Stata or statistics question: it requires knowledge of the substance of the entertainment industry. Perhaps somebody who follows this Forum has such knowledge and respond. If not, you need to find some source of expertise on entertainment economics.

[Yasmin bzd]

Ok, thank you Clyde

[Jordan Sydenham]

Before conducting any data collection or research a good idea is to see what literature is already out there and try and understand their methodology. I would suggest that you read the following paper; Parlow, A., & Wagner, S. (2018). Netflix and the Demand for Cinema Tickets-An Analysis for 19 European Countries. Found at ;
https://mpra.ub.uni-muenchen.de/8975...aper_89750.pdf

You can then follow this paper with a few adaptations that you see fit to make.

[Yasmin bzd]

Thank you Jordan, I have already seen this paper. Yet I started to work on my Netflix paper before the publication of this article. I still had some technical questions that I carried here.

Regarding Clyde's recommendations, having zero values in a continuous data, and taking their logs, poses problems in stata. Then, there are alternative ways of transforming such data. This is why instead of leaving them zero, I gave them a small number (4200) (when logs taken hence there wouldn't be a problem). On a different note, I added a dummy variable, for each country and eacy year , for the "Netflix entry", and verified that the entry of Neflix raises the box office.