Fixed Effects looking at ESG on Revenue/PnL

[Tyler Zabel]

Hi All,

I am struggling with STATA as a newbie undergrad trying to finish my dissertation looking at the 'Financial impact of ESG on 81 clubs in the English Football League.' As ESG in football is quite recent, I only have a panel of two years (I know not ideal but was hoping for a third year's worth of data which never came). I am starting with a pooled OLS and then move on to looking at FE to control for heterogeneity. I have two dependent variables - revenue and profit and loss, and independent variables Governance, Financial Sustainability, Environment, and Fan engagement with an aggregated ESG score based on those four components. I am first looking at the aggregated ESG score on its own as these variables quite intuitively have high levels of colinearity so I don't include multiple independent variables in my baseline model. I have My initial code looks something like this, where there is an interaction dummy term to determine league specific impacts and attendence and leagueposition are control variables.



// Step 1: OLS regression with interaction term

reg pretaxpnl fairgameesgscore averageattendence leagueposition c.leaguedummy#c.fairgameesgscore


// Step 2: Fixed Effects (FE) regression

xtreg pretaxpnl fairgameesgscore averageattendence leagueposition c.leaguedummy#c.fairgameesgscore, fe


// Step 3: Random Effects (RE) regression

xtreg pretaxpnl fairgameesgscore averageattendence leagueposition c.leaguedummy#c.fairgameesgscore, re


// Step 1: OLS regression with interaction term

reg revenue fairgameesgscore averageattendence leagueposition c.leaguedummy#c.fairgameesgscore


// Step 2: Fixed Effects (FE) regression

xtreg revenue fairgameesgscore averageattendence leagueposition c.leaguedummy#c.fairgameesgscore, fe


// Step 3: Random Effects (RE) regression

xtreg revenue fairgameesgscore averageattendence leagueposition c.leaguedummy#c.fairgameesgscore, re

HAUSMAN

My questions are as follows:

- Do I set up tests for heteroscedasticity/autocorrelation before I run my findings (Breusch Pagan, F test, etc.) to know whether to use cluster or robust SE?
- When I run this code, I get R² values of around 0.5 but super high p values so basically no conclusions can be made from my model - don't know how to get around this.
- As a newbie, just looking for a general bit of help with standard approaches to FE and pooled OLS and what tests to use.

Thank you.

[Carlo Lazzaro]

Tyler:
welcome to this forum.
If you have a N>T panel data with a continuous regressand, you should consider -xtreg- as your first choice (in addition, your OLS codes consider your observations as independent, ignoring the panel structure of your dataset).
There are community-contributed modules to test for heteroskedasticity and autocorrelation for -xtreg,fe- (see -search xttest2- ; -search xttest3-).
Cluster-robust standard errors can be invoked via the options -robust- or -vce(cluster clusterid)-, as they do the very same job under -xtreg- (while this does not hold for -regress-);
As -hausman- do not support non-default standard errors, you should switch to the community-contributed module -xtoverid-, that, in turns, does not support -fvvarlist- notation (see -xi- prefix as a workaround).
As an aside, you cannot expect to run a panel data regression without knowing the building blocks of this tricky section of econometrics: see references in -xtreg- entry of Stata .pdf manual.

[Tyler Zabel]

Hi Carlo,

Thanks for the reply, appreciate it can be frustrating dealing with STATA novices like myself.

I utilise xtreg and find heteroscedasticity but no serial correlation. I use robust SE. I use the community-controlled alternative to the Hausman and still have FE as a superior model to RE. I also do a model specification test using sq_fitted as per one of your responses to another forum which gives me a p-value of 0.3788.

When I run OLS/FE for both my dependent variables separately using this code:


// Step 1: OLS regression with interaction term
reg pretaxpnl fairgameesgscore averageattendence leagueposition c.leaguedummy#c.fairgameesgscore, robust

// Step 2: Fixed Effects (FE) regression
xtreg pretaxpnl fairgameesgscore averageattendence leagueposition c.leaguedummy#c.fairgameesgscore, fe robust

// Step 1: OLS regression with interaction term
reg revenue fairgameesgscore averageattendence leagueposition c.leaguedummy#c.fairgameesgscore, robust

// Step 2: Fixed Effects (FE) regression
xtreg revenue fairgameesgscore averageattendence leagueposition c.leaguedummy#c.fairgameesgscore, fe robust


I compile this results table which gives no significant results for anything meaningful.

Table 10 - OLS/FE regression results

Outcome Variable	Model Type	Coeff. of ESG Score	Coeff. of Interaction Term (League Dummy * ESG Score)	R² (Within for FE)	F-Statistic	Prob > F
Pre-tax P&L	OLS	230.44 (p=0.465)	-246.88 (p=0.465)	0.1231	F(4, 75)=2.30	0.0662
	FE	-77.99 (p=0.926)	-349.60 (p=0.435)	0.0165	F(4, 40)=0.25	0.9105
Revenue	OLS	2062.93 (p=0.002)*	-797.26 (p=0.162)	0.7485	F(4, 75)=40.80	<0.0001**
	FE	-720.15 (p=0.385)	-790.97 (p=0.077)*	0.2258	F(4,35)=2.55	0.0562

Clearly, this model is not ideal for estimating any sort of causal relationship but I am deep into this dissertation now so don't really know where to go with this. Any help would be appreciated as well as any feedback on mistakes you may spot. Could this lack of relationship potentially be because I only have a 2 year panel ?

Thanks,

Tyler

[Carlo Lazzaro]

Tyler:
1) as alredy detected, your OLS codes ignores ths panel structure of your dataset: you shoould cluster your standard errors on the -panelid- variable (please note that, under -regress-, -robust- takes heteroskedasticity only into account);
2) I would not consider pooled OLS as my first strategy;
3) if your -fe- results do not to support any panel-wise effect, you should go back to pooled OLS with clustered standard errors. Othet things being equal, a two-wave panel dataset may have some role in that;
4) I agree that any result that you may obtain does not suppport any causal relationshio, but only correlation (see -didregress- and -xtdidregress- for causal inference);
5) if I may, do not be overstressed by non-significant results, but try to explain that in the Discussion section of your dissertation.