[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]

Re: st: broken stick (piecewise linear) regression

From   Maarten buis <[email protected]>
To   [email protected]
Subject   Re: st: broken stick (piecewise linear) regression
Date   Thu, 31 Jan 2008 21:05:58 +0000 (GMT)

--- Raphael Fraser <[email protected]> wrote:
> I saw a thread with the above subject line concerning
> piecewise linear regression posted in 2005 on Statalist.
> I am facing the same problem now; that of estimating the unknown
> change point. Mention was made of using -nl- but it uses
> least squares method. The only way to solve this problem
> is numerical optimization. The problem can be solved
> quite easily in SAS using NLIN PROC.
> Is there a Stata solution?

Purely looking at the name my guess would be that both -nl- and NLIN
PROC use the same or a very similar algorithm: nonlinear least squares,
which is not the same as least squares. It is the most common way I
have seen to estimate these kinds of models, though you could also us
-ml- but that would be a bit (but not much) more work. Notice that this
is a tough maximization problem so it really helps to have good
starting values. I would recommend starting with a scatter plot, and
find the likely location for the knot, use that knot to create a spline
and estimate a regression. Use those parameters as starting values when
you use -nl- like in the example below to finally estimate your

*--------------- begin example -----------------
sysuse auto, clear

// first take a look 
twoway scatter price mpg

// 20 seems a nice starting value for
// the knot location
mkspline mpg1 20 mpg2 = mpg, marginal
reg price mpg1 mpg2
local cons = _b[_cons]
local b1 = _b[mpg1]
local b2 = _b[mpg2]
predict initial

// use -nl- to find the knot
nl (price = {cons=`cons'} + {b1=`b1'}*mpg + ///
            {b2=`b2'}*max(mpg -{k1=20},0))
predict nl

// look at the results
twoway scatter price mpg ||          ///
       line initial nl mpg, sort     ///
       legend(order( 2 "first guess" ///
                     3 "estimated"))
*---------------- end example -------------------
(For more on how to use examples I sent to the Statalist, see )

Hope this helps,

Maarten L. Buis
Department of Social Research Methodology
Vrije Universiteit Amsterdam
Boelelaan 1081
1081 HV Amsterdam
The Netherlands

visiting address:
Buitenveldertselaan 3 (Metropolitan), room Z434

+31 20 5986715

Sent from Yahoo! Mail - a smarter inbox

*   For searches and help try:

© Copyright 1996–2024 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index