OT Quantile Regression Why not ranks or percentiles?

[Art Kendall]

Just curious.

As a general rule of thumb one wants a variable to be as fine grained as is practical in the situation.
However, the few examples I have seen of quantile regression have coarsened to 5 or so values.

Is there a substantive or computational reason for  using this few values?

[Jon K Peck]

Quantile regression builds models of the specified quantiles in the same way as ordinary regression builds models of means, except that they are computationally much more complex.  I doubt that anyone would want to build models of, say, every quantile in (0,1) by .1.  If you are interested in how the coefficients vary by quantile, a half dozen or so points should give a pretty good picture.


Jon Peck (no "h") aka Kim
Senior Software Engineer, IBM
[hidden email]
phone: [phone]




From:        Art Kendall <[hidden email]>
To:        [hidden email],
Date:        10/16/2013 11:48 AM
Subject:        [SPSSX-L] OT  Quantile Regression  Why not ranks or percentiles?
Sent by:        "SPSSX(r) Discussion" <[hidden email]>

---

Just curious.

As a general rule of thumb one wants a variable to be as fine grained as is practical in the situation.
However, the few examples I have seen of quantile regression have coarsened to 5 or so values.

Is there a substantive or computational reason for  using this few values?
--
Art Kendall
Social Research Consultants

Art Kendall
Social Research Consultants

---

View this message in context: OT Quantile Regression Why not ranks or percentiles?
Sent from the SPSSX Discussion mailing list archive at Nabble.com.

[Art Kendall]

That makes sense, the interest would be in n slope coefficients so too many such slopes would be very difficult to interpret.
On 10/16/2013 1:59 PM, Jon K Peck wrote:

Quantile regression builds models of the specified quantiles in the same way as ordinary regression builds models of means, except that they are computationally much more complex.  I doubt that anyone would want to build models of, say, every quantile in (0,1) by .1.  If you are interested in how the coefficients vary by quantile, a half dozen or so points should give a pretty good picture.


Jon Peck (no "h") aka Kim
Senior Software Engineer, IBM
[hidden email]
phone: [phone]




From:        Art Kendall [hidden email]
To:        [hidden email],
Date:        10/16/2013 11:48 AM
Subject:        [SPSSX-L] OT  Quantile Regression  Why not ranks or percentiles?
Sent by:        "SPSSX(r) Discussion" [hidden email]

---

Just curious.

As a general rule of thumb one wants a variable to be as fine grained as is practical in the situation.
However, the few examples I have seen of quantile regression have coarsened to 5 or so values.

Is there a substantive or computational reason for  using this few values?
--
Art Kendall
Social Research Consultants

Art Kendall
Social Research Consultants

---

View this message in context: OT Quantile Regression Why not ranks or percentiles?
Sent from the SPSSX Discussion mailing list archive at Nabble.com.

===================== To manage your subscription to SPSSX-L, send a message to [hidden email] (not to SPSSX-L), with no body text except the command. To leave the list, send the command SIGNOFF SPSSX-L For a list of commands to manage subscriptions, send the command INFO REFCARD

[Andy W]

Actually I remember seeing plenty of examples of the opposite, people estimate the coefficients at various quantiles and plot them in a line (plus area for confidence intervals) for the coefficient at various quantiles between .1 and .9.

See

Britt, Chester L. "Modeling the distribution of sentence length decisions under a guidelines system: An application of quantile regression models." Journal of Quantitative Criminology 25.4 (2009): 341-370. http://dx.doi.org/10.1007/s10940-009-9066-x

Here is a picture taken from Page 360 of the forementioned article



For other online examples see

SAS's procedure - http://support.sas.com/documentation/cdl/en/statug/63033/HTML/default/viewer.htm#statug_qreg_sect033.htm

And some examples from Stata
 - http://personal.stthomas.edu/mehartmann/sentencing_disparity_v14.pdf
 - http://www.decisionsonevidence.com/2011/10/wonkish-statistical-tool-choices-make-a-difference/

A cynic might say most articles only report the coefficients for the "most interesting" quantiles!

[Peter Spangler]

Quantile regression in SPSS using the R module plots the coefficient for each quantile with a confidence interval, including an output table for each coefficient chosen for the model.

On Wed, Oct 16, 2013 at 12:43 PM, Andy W <[hidden email]> wrote:

Actually I remember seeing plenty of examples of the opposite, people
estimate the coefficients at various quantiles and plot them in a line (plus
area for confidence intervals) for the coefficient at various quantiles
between .1 and .9.

See

Britt, Chester L. "Modeling the distribution of sentence length decisions
under a guidelines system: An application of quantile regression models."
Journal of Quantitative Criminology 25.4 (2009): 341-370.
http://dx.doi.org/10.1007/s10940-009-9066-x

Here is a picture taken from Page 360 of the forementioned article

<http://spssx-discussion.1045642.n5.nabble.com/file/n5722587/QuantReg_Britt.png>

For other online examples see

SAS's procedure -
http://support.sas.com/documentation/cdl/en/statug/63033/HTML/default/viewer.htm#statug_qreg_sect033.htm

And some examples from Stata
 - http://personal.stthomas.edu/mehartmann/sentencing_disparity_v14.pdf
 -
http://www.decisionsonevidence.com/2011/10/wonkish-statistical-tool-choices-make-a-difference/

A cynic might say most articles only report the coefficients for the "most
interesting" quantiles!



-----
Andy W
[hidden email]
http://andrewpwheeler.wordpress.com/
--
View this message in context: http://spssx-discussion.1045642.n5.nabble.com/OT-Quantile-Regression-Why-not-ranks-or-percentiles-tp5722584p5722587.html

Sent from the SPSSX Discussion mailing list archive at Nabble.com.

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[Art Kendall]

The x-axis label on the picture from page 360 is not very clear on my monitor. 
Would you be so kind as to post what it says?
On 10/16/2013 3:43 PM, Andy W [via SPSSX Discussion] wrote:

Actually I remember seeing plenty of examples of the opposite, people estimate the coefficients at various quantiles and plot them in a line (plus area for confidence intervals) for the coefficient at various quantiles between .1 and .9.

See

Britt, Chester L. "Modeling the distribution of sentence length decisions under a guidelines system: An application of quantile regression models." Journal of Quantitative Criminology 25.4 (2009): 341-370. http://dx.doi.org/10.1007/s10940-009-9066-x

Here is a picture taken from Page 360 of the forementioned article



For other online examples see

SAS's procedure - http://support.sas.com/documentation/cdl/en/statug/63033/HTML/default/viewer.htm#statug_qreg_sect033.htm

And some examples from Stata
 - http://personal.stthomas.edu/mehartmann/sentencing_disparity_v14.pdf
 - http://www.decisionsonevidence.com/2011/10/wonkish-statistical-tool-choices-make-a-difference/

A cynic might say most articles only report the coefficients for the "most interesting" quantiles!

---

If you reply to this email, your message will be added to the discussion below:

http://spssx-discussion.1045642.n5.nabble.com/OT-Quantile-Regression-Why-not-ranks-or-percentiles-tp5722584p5722587.html

To start a new topic under SPSSX Discussion, email [hidden email]
To unsubscribe from SPSSX Discussion, click here.
NAML

[Andy W]

The X axis is labeled as "tau" in text and it represents quantiles. The Tau symbol is often used to represent quantiles, but it seems an oversight to use it as a textual label.

The points in the graph are coefficient estimates of the same model at various different quantiles, from the picture I would guess every at every .02 quantile from between .1 to .9 (the paper doesn't say exactly). The different plots each represent a different coefficient.