EViews.com

Hello,

I am trying to find out if I need a one-way or two-way fixed effects model. I am tring to estimate profitability via fixed effects estimation. I have data for 392k firms over the span of 10 years. My research goal is to find out the effect of low interest rates on firm profitability. I have some controls included as well, like age, size and debtratio. I have also an interaction term included: interestrate*cashratio. Cashratio is defined as cash/total assets. The rationale is that large cash position have a negative effect on profitability because of the low interest rates.

I am told by my promotor that I have to determine first if one-way or two-way fixed effects is the way to go. The Hausman test indicates that fixed effects is in order.
When estimating a two-way fixed effects model however, I have estimation output but I am encountering following problem: I receive a 'near singular matrix'-error when performing the redundant Fixed effects test (likelihood ratio)*.
When I leave interest rate as a separate variable out, there is no error and the test delivers output that 2 way-FE is significant (significant F for bot cross-section as period).

Why do I get the near singular matrix error with interest rates in the model? And does this mean I have to let interest rate as a separate regressor out of the model? Is there a way of keeping interest rates in the model?

The redundant fe-test delivers no singular matrix error for a one-way FEM and interest rates included.



Thank you

Not quite sure what you did, but if you include time fixed effects and the interest rate is constant for each time period then the two are indeed perfectly multicollinear.

I have indeed interest rates for each year. These are the same for each company as I am using the 10Y bond rate. Just to make it clear, I am going for panel estimation, where I select fixed effects for both cross-section and period. With 2-way I get estimation output, with a sign.coefficient for interest rate close to 1. Afterwards when doing the redundant fixed effects- likelihood ratio test, there the 'near singular matrix' error appears.

Does this mean that interest rate is part of the time fixed effect and I have to leave it out when performing the 2-wayFE estimation?


My question is: how do I really know which one to choose, 1-way or 2-way? To recap, 2-way equation without interest rate as separate regressor but left inside the model as part of the interaction term works fine
and/or 1-way with interest rate (here sign coef is 0) and interaction term?

I have indeed interest rates for each year. These are the same for each company as I am using the 10Y bond rate. Just to make it clear, I am going for panel estimation, where I select fixed effects for both cross-section and period. With 2-way I get estimation output, with a sign.coefficient for interest rate close to 1. Afterwards when doing the redundant fixed effects- likelihood ratio test, there the 'near singular matrix' error appears.

Does this mean that interest rate is part of the time fixed effect and I have to leave it out when performing the 2-wayFE estimation?


My question is: how do I really know which one to choose, 1-way or 2-way? To recap, 2-way equation without interest rate as separate regressor but left inside the model as part of the interaction term works fine
and/or 1-way with interest rate (here sign coef is 0) and interaction term?

Right, leave out the interest rate when you include time fixed effects.
I believe you can just do an F-test comparing the sum of squared residuals from the two regressions. Probably needs to be done manually and might need a little thought.

I have a follow up question:

I want to control if industry has a significant effect on profitability. As the code doesn't change over time for a certain firm I can't include it just like that. I am using therefore another proxy variable based on the median value of that industry for that year. I created an extra series medianindustry using following command @mediansby(profitability score(ROE in this case);ind.code,year).

The coefficient for this variable is 0.30 and is highly significant. My question is how should I interpret and report this value? My dilemma stems from the fact that I have an abstract value for a series (0.30) but not for the separate industries. Again this is a two-way FEM.

Saying that industry has a significant effect and no more seems a bit brief, or are these the limitations of the software and estimation technique?

Seems like you're showing that profitability of the industry matters for profitability of the individual firm. Not too surprising.

None of this is a limitation of either the software or the estimation technique though.

Seems like you're showing that profitability of the industry matters for profitability of the individual firm. Not too surprising.

None of this is a limitation of either the software or the estimation technique though.

No, I guess not, it is just a control variable. I was only wondering if coeff for he separate sectors were available, but that isn't necessary here.
Thank you Richard for your help!