Problem with choosing which regression (panel-data)

[Anouk Hond]

Hi,

I am having trouble with deciding what regression to use for my data.
I've got panel data (data over multiple years) and want to do a regression with moderator. I am doing research on the effect of reputation on the relationship between distinctiveness and performance for craft breweries.
My dependent variable is relative_performance, independent variable dist_brewery and moderator is reputation.

I am currently running the fixed effects model as the Hausman test i did was significant and my regression with moderator included looks like this: xtreg relative_performance c.dist_brewery##c.dist_brewery##c.reputation Gold Silver Bronze total_reviews numb_beers number_years sd_consumerscore i.Year, fe vce(robust).
I want to have a look at if there is an inverted u-shape or u-shape in this model and if it changes from inverted-u with low reputation to u shape when reputation is high.

I do find that nothing is significant, but when i solely use regress instead of xtreg everything is significant. Is that possible to use or not?

Thanks in advance!

[Carlo Lazzaro]

Anouk:
welcome to this forum.
-xtreg,fe- and -regress- cannot be compared.
As per FAQ, please provide more details about what you typed and what Stata gave you back. Thanks.

[Anouk Hond]

Hi Carlo,

Thank you so much for your response. I am sorry if I wans't clear in my earlier message.
I think I found the right regression, xtpoisson, as I am using panel data with a count dependent variable.

Now I have a new problem I am struggling with, which you may know the answer to.
After the regression, I want to have a look at the curvilinearity in the model, and see if at different values of the moderator the shape changes (from inverted u to u-shape).

In order to do so, I need to check first whether the squared term of my independent variable is significant (which is dist_brewery and it is significant). For the second conditions I need to check whether the minimum and maximum of the slopes at those 3 different levels of the moderator (reputation 5th median and 95th percentile 2.015277 3.403288 4.39309) are significant and one is negative and one is positive.
Now I struggle a lot with what margins command to use and how i can then see the min and max value of the slopes for all 3 values of the moderator. See the screenshots for more info about my output in stata.

Thanks a lot in advance and let me know if you need more information to be able to provide an answer.

Kind regards,
Anouk

[Carlo Lazzaro]

Anouk:
1) I cannot tell if -xtpoisson- is the right way to go here as I do not know about your -depvar- (count? Continuous?). Be as it may, -xtpoisson- gives back conditional fixed efefct due to incidental parameters bias (PII: S0304-4076(99)00044-5 (brown.edu));
2) you seem to have a maximum at

. di .1812873/(2*.0069012)
13.134477

3) see example #5, -margins- entry, Stata .pdf manual.

As an aside, for reasons that are well explained in te FAQ, please use CODE delimiters instead of screenshots to share what you typed and what Stata gave you back. Thanks.

[Anouk Hond]

Hi Carlo,
Thank you for your response. I now have another struggle: I want to check the turning points at 3 different values of my moderator, but i can't find a clear code in stata which does that. Without the moderator included in the regression I used this formula:

nlcom _b[dist_brewery]/(-2*_b[c.dist_brewery#c.dist_brewery])

.

I now want to look at the turning points for reputation at the values

2.015277  3.403288  4.39309

.

A similar question I have is how can I do this for just adding the moderator to the regression? Can I then do the regression and use this formula

nlcom _b[dist_brewery]/(-2*_b[c.dist_brewery#c.dist_brewery])

. as the regression was different anyway with the moderator added?

Thanks in advance.

[Carlo Lazzaro]

Anouk:
see How can I calculate the turning point of a moderator at given values in Stata? - Statalist

[Anouk Hond]

Thanks! With my variables I used the code a bit differently, is this correct:

foreach s of numlist 2.015277 3.403288 4.39309 {
    local turning_point = (-_b[dist_brewery] / (2 * _b[c.dist_brewery#c.dist_brewery])) - `s' * (_b[c.dist_brewery#c.reputation] / _b[c.dist_brewery#c.dist_brewery])
    display "`s'", "`turning_point'"
}

And one more question: how do I then see the significance, std err, 95% interval etc of these turning points?

[Carlo Lazzaro]

Anouk:
I'd think of -margins- or -nlcom- but I do not have the time to delve into this issue at the moment.