main regression insignificant, after adding fixed effects significant

[bob van eldik]

Hello,

I am new to this forum and currently writing my Bachelor Thesis for Finance.
I am researching the effect of CSR on Dividend Yield (DY).
In the simple regression, the effect of CSR on DY is positive but not significant at the 10% level:
dep: DY
indep: CSR
control variables: Size, Growth, Debt, EPS, ROA.

But after adding fixed industry effects to this regression, my regressions are positive and significant at a 1% level.
How should i interpret/explain this?

Greetings Bob

[Clyde Schechter]

A fixed-effects regression is a within-panel regression, whereas an ordinary pooled OLS regression does not distinguish between and within effects. So the fixed-effects regression says that within any particular industry, increasing values of CSR are associated with increasing values of DY. The OLS regression, in combination with this, indirectly suggests that there is little or no association between the average level of CSR in an industry and the average DY in that industry. That is industries with higher average CSR do not necessarily have higher average DY, but within any given industry, over time, increasing CSR is associated with increasing DY within that industry.

[bob van eldik]

Okay that sounds logical. Two questions:1. So if I conclude this for my thesis,it would be something like:
CSR does not have a significant effect on DY, but when industry fixed effects are added. We can conclude that within industries, increasing values of CSR are associated with increasing value of DY?

2. After robustness tests for endogoneity and heteroskedecatity the effect was insignificant again..
How should I handle this? My own assumption: The model suffers from omitted variable bias so is not significant and we can't conclude that CSR has an positive effect on DY

Thanks for the help

[Clyde Schechter]

So if I conclude this for my thesis,it would be something like: CSR does not have a significant effect on DY, but when industry fixed effects are added. We can conclude that within industries, increasing values of CSR are associated with increasing value of DY?

Not quite. Absence of evidence of an effect is not evidence of absence of an effect. That would depend on considerations of statistical power, measurement reliability, etc. So I would say it more cautiously, something like this: While we found that within industries increasing values of CSR are associated with increasing values of DY, we could not establish a similar relationship between these variable across industries.

With regard to your second question, as long as you restrict your conclusions to claiming an observed association, endogeneity and heteroskedacticity are not problems. Heteroskedacticity, in particular, affects the specifics of inference (standard errors, p-values), but it does not introduce bias (as long as it is not due to model mis-specification). Endogeneity certainly undercuts any pretense of causal relationships, but a crude association can still be claimed. Missing variable bias is a possible reason for both of these issues, but it is not the only possibility. So while I would raise it as a possible cause and suggest that future research might explore this, I wouldn't make it the centerpiece of my discussion of limitations. We can conclude that crudely CSR is positively associated with DY within industries over time. We just can't strengthen that claim, not even in the slightest.