R-squared of 0.001, what could be the possible cause?

[Victoria Rogers]

Most of my regressions generate R-squared of 0.001

I know that a R-squared normally doesn't say much, but my topic is about the effect of gender on stock returns.
So one would expect to see better results than that after using returns as the dependent variable and gender as the independent variable.

[Nick Cox]

Sometimes null hypotheses really are about right.

[Rich Goldstein]

if you have a lot of "ties" in your predictor variables, with different values for the outcome, R-squared will (1) have a max below 1.0 and (2) may be very small (in addition to what Nick says); the -maxr2- command is to help discover this situation (for linear regression only); so -search maxr2- and then download it

[Carlo Lazzaro]

Victoria: does your regression suffer from omitted variable bias? Kind regards, Carlo

[Victoria Rogers]

Sometimes null hypotheses really are about right.

There are some papers wherein there's no gender effect, so you could be right in this case.

Victoria: does your regression suffer from omitted variable bias? Kind regards, Carlo

Yes, in some models

[Victoria Rogers]

if you have a lot of "ties" in your predictor variables, with different values for the outcome, R-squared will (1) have a max below 1.0 and (2) may be very small (in addition to what Nick says); the -maxr2- command is to help discover this situation (for linear regression only); so -search maxr2- and then download it

maximum R-square           = 0.0166
relative R-square          = 0.0536
 prob > F =    0.0000
number of covariate patterns = 604
   as ratio of observations = 0.005

I assume you meant 0.01 as a max instead of 1, however, I'm not sure if I do have ties according to this output of -maxr2-

[Rich Goldstein]

I read this as saying that there are 604 distinct patterns and that this is one-half of one percent of the N of your estimates, meaning that your N is approximately 200*604 - is it? In this case, yes, you have lots of ties;

and no, I meant a max of 1 for r-squared; if you have ties in your predictor variables and there is variation in the outcome variable within the sets of ties, then the maximum r-squared for that estimate must be less than 1.0; the output here tells you that for your data and your model, the maximum r-squared is 0.0166 (much less than 1.0) but still rather larger than 0.001 you reported in #1 above

if you want to read the article I originally wrote about this, and obtain some citations, go to http://www.stata.com/bookstore/indiv...lletin-issues/ and download STB 9

[Victoria Rogers]

Yes that's correct. Do I have a lot of ties because I used normal regress to use your command? Normally I use -xtreg- due to my panel data.
I assumed you wanted to say something else by saying a max R-squared of 1 because 1 is always the max (as far as I know)

I'll read your article but is there something I should do in the meantime with the ties? Is there a way to increase the 0.0166 to something like 0.5000?

[Rich Goldstein]

the ties arise both because of the data and because of the model; your response in #8 is the first mention of -xtreg- that I have seen in this thread; please don't hide information

[Sarah Edgington]

I assumed you wanted to say something else by saying a max R-squared of 1 because 1 is always the max (as far as I know)

Note that Rich did not say that the max was 1 in this case. He said the max was less than 1. How much less than 1 depends on the data.

[Victoria Rogers]

the ties arise both because of the data and because of the model; your response in #8 is the first mention of -xtreg- that I have seen in this thread; please don't hide information

I'm sorry about that Rich. I'm trying to solve various problems at the same time in different threads, which are all related to each other but 'cannot' be put in just 1 thread. Besides that, I need to give my boss the new results tomorrow. Therefore, I'm a little bit stressed and therefore I probably assumed that I already said in this thread that I'm working with panel data.

[Victoria Rogers]

Note that Rich did not say that the max was 1 in this case. He said the max was less than 1. How much less than 1 depends on the data.

I understood Rich but I was being vague even though that was not my intend. I meant to say, the part that Sarah quoted, that 'as far as I knew' R-squared always has a max of 1. Therefore, I first misunderstood Rich and now that I've learnt it by reading his manual and posts I now understand him completely (his posts)

[Richard Williams]

I'll assume that the results are correct. If so, I would think this would be a pretty strong counter to any claims that men are better managers than women. It seems like a situation where an insignificant effect is very important from a substantive standpoint.

If, on the other hand, your findings go directly counter to what everybody else in the world has found, you need to double-check your work and/or be prepared to defend the quality of your analysis.

[Victoria Rogers]

There are some papers which also report insignificant effects between men and women, however, most papers report significant effects. I collected most data by hand, so my data has never been reasearched before (as far as I know). I'm trying to work with all the possible models I can think of, so I can better understand my data. At the moment, I'm trying to figure out -xthtaylor- and -xtmixed- with the help of 2 different threads. However, -xtmixed- seems to be a lot easier to comprehend.