# RE: st: Dependent continuous variable with bounded range

 From "Pavlos C. Symeou" To statalist@hsphsun2.harvard.edu Subject RE: st: Dependent continuous variable with bounded range Date Tue, 15 Apr 2008 17:26:46 +0100

Dear Nick, Anders, and Jay,

thank you very much for your responses. First, I need to agree with you that this problem can not be handled with interval regression. But as Anders recommends, I need to give further information about my variable in question, "reputation". The variable is a construct consisted of three other variables. These three variables take discrete values that range from 0-10 on a 11-point likert scale, namely 0,1,2,3...,10. These are responses to a survey questionnaire. These three variables' values are first added together and divided by 3 to yield my focal variable "reputation" (unfortunately I only have access to the final variable and not its components). Values for "reputation" closer to 0 suggest lower firm reputation. Values closer to 10 suggest higher firm reputation. A histogram shows a relatively normal distribution as you can see from its Skewness below. I also provide you with statistics on max and min values to respond to Anders' inquiry.

Nick advised that I transform my response so that is unbounded using a link function e.g. logit on response/10. (I am not sure how I can do this) Yet, Jay's suggestion is that this might not be that beneficial. Can you please advise on how to proceed?

Min 2.95 Max 8.32
Mean 6.267978
Variance .5091705
Skewness -.057921
Kurtosis 3.642905

Yours truly,

Pavlos

I find both Nick Cox's and Jay Verkuilen's comments very reasonable.
But Pavlos does not mention how the variable "reputation" is created.
How would reputation get a value of, say, "9.6"? Where does the
boundary 0-10 come from, e.g., from the sample's population or from a
questionnaire
? Is reputation a scored variable or does it represent
original data?

Nick Cox outlined several strategies, to which I will add just a bit:

(1) How close to the boundaries are your observations? If the distribution looks reasonably symmetric, you probably won't gain much from using a specialized model that "knows" about the boundaries. If you have some skew due to the ceiling or floor but not a truly L- or J-shaped distribution, the logit transformation will probably normalize your errors enough to do ordinary panel regression models. If the distribution is truly L- or J-shaped, no transformation will fix things up.
(2) -betafit- (by Nick, Maarten, et al) with clustered robust SEs is a viable alternative that uses the linking strategy. This uses the beta distribution as an error model. It won't adjust for the autoregressive nature of panel data, though, but maybe that'll work well enough.
(3) If you want to use the GLM approach and are willing to move to different software (SAS or winBUGS), I can give you examples to do random effects and AR-adjusted beta regression. (One of these days, I want to port a random effects and GEE betareg over to Stata; no time right now.)

Jay

This does not sound like an interval regression problem to me. Interval
regression is for when at least some individual values are known as intervals not
points. Here values are known but must lie in a prescribed interval.
Divide by 10 and the problem has exactly the same form as that often met
for
proportions.
There are correspondingly various options:
that there is no guarantee that predicted values will lie in [0,10] certainly for some
values of the predictors
and possibly for your observed data. Arguably that is quite wrong in
principle and it may be a real
nuisance for your project. There are implications all the way downstream
in terms of assessing fit.
2. Transform your response so that is unbounded. Model in terms of that
and then back-transform.
3. A variant of 2, and in many ways preferable, to use a link function
e.g. logit on response/10.
4. No doubt others.
Nick
n.j.cox@durham.ac.uk

See -help xtintreg-. It does random-effects interval data regression
models.

Pavlos C. Symeou <p.symeou@jbs.cam.ac.uk> wrote:

``` I have panel data for 100 firms for five years and I want to examine
```
```the
```
``` effects of various variables on "reputation". My variable
```
```"reputation"
```
``` takes  continuous values in the range of 0-10. Namely, it can take
values of 1,2,3 but also of 2.5, 9.6, etc. The values that the
```
```variable
```
``` "reputation" takes in my sample range between 2.6 - 8.3. Can you
```
```please
```
``` advise if I can still use panel OLS estimation for panel data or
```
```should
```
``` I use a different model? In essence, my main concern is the
```
```limitations
```
``` of the bounded range of my variable.
```
```
*
*   For searches and help try:
*   http://www.stata.com/support/faqs/res/findit.html
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/
```