Nonlinear Models… Fitting Curves
[Pages:19]Stat 102
Fall, 2014
Nonlinear Models... Fitting Curves
ROADMAP
Linear relation: effect on y of changes in x is the same at every value of x.
Nonlinear relation: effect on y of changes in x depends on the value of x.
We can expect nonlinearities in many business applications Diminishing marginal effect (eg, promotion response, manufacturing) Relationships with constant elasticity (eg, price and demand)
Examples Diamond prices Track times Retail sales Auto mileage
diamonds.jmp, more_diamonds.jmp track.jmp display.jmp cars.jmp
Robert Stine , The Wharton School of the University of Pennsylvania, prepared this document. Copyright 2014.
Fitting Curves
Statistics 102, Fall 2014
LINEAR MODELS
Meaning of linearity Equal changes in X associated with equal changes in Y (on average)
Example: diamond prices
Price
$1,600.00
$1,400.00
$1,200.00
$1,000.00
$800.00
0.3
0.35
0.4
0.45
0.5
Weight (carats)
Interpretation: Model parameters, assumptions
Substantively: Does this model make sense?
THE WHARTON SCHOOL OF THE UNIVERSITY OF PENNSYLVANIA
2
Fitting Curves
Add more data, namely larger diamonds
Statistics 102, Fall 2014
Price
$1,400,000 $1,300,000 $1,200,000 $1,100,000 $1,000,000
$900,000 $800,000 $700,000 $600,000 $500,000 $400,000 $300,000 $200,000 $100,000
$0 0 1 2 3 4 5 6 7 8 9 10 11 12 13 Carat
Does a linear model still make sense?
How do these data violate conditions implied by the SRM?
Could we have anticipated these problems earlier?
What to do about these problems?
THE WHARTON SCHOOL OF THE UNIVERSITY OF PENNSYLVANIA
3
Fitting Curves
Statistics 102, Fall 2014
NONLINEAR MODELS... CURVES
Question: What sort of curve captures the pattern in the prior plot?
Finding the right transformation...
(a) Graphically from scatterplot of Y on X
(b) Residual plots
(c) Substantively (what would make sense)
Logs and percentages
Change on a log scale: think percentages
loge (x) ? loge (y)= loge (x/y)
= loge ( (y + x-y)/y )
= loge (1 + (x-y)/y)
(x-y)/y
if x-y is small compared to y1
Modeling prices: How are diamond prices related to the price of diamonds?
1 This only works in such a nice way with natural logs. Base 10 logs (or others) require some messy constants.
THE WHARTON SCHOOL OF THE UNIVERSITY OF PENNSYLVANIA
4
Fitting Curves
On log scales, the fit appears linear...2
Price
2,000,000
1,000,000
700,000 450000,,000000 300,000 200,000
100,000
70,000 4500,,000000 30,000 20,000
10,000
6,000 4,000 3,000 2,000
1,000
600 400 300 200
0.2
0.3 0.4 0.5 0.7 1
2
Carat
3 4 5 6 7 8 10
Statistics 102, Fall 2014
Interpretation of slope: elasticity of price with respect to size in carats. 1% increase in weight associated with 2% (1.97) increase in price, on avg
2 Do this in JMP by double clicking on each axis in the scatterplot and picking the "log" option in the dialog. To get the fitted model using logs, select the Fit special item from the Fit Y by X dialog. An example of this dialog appears later.
THE WHARTON SCHOOL OF THE UNIVERSITY OF PENNSYLVANIA
5
Fitting Curves
Statistics 102, Fall 2014
Fit on original scale...
Price
1,400,000 1,300,000 1,200,000 1,100,000 1,000,000
900,000 800,000 700,000 600,000 500,000 400,000 300,000 200,000 100,000
0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 Carat
"Linear regression"... The fitted model always has a "linear" equation, but the variables X and Y may involve transformations of the original measurements.
Estimated (loge Price) = 8.554 + 1.972 (loge Carat)
Y
X
THE WHARTON SCHOOL OF THE UNIVERSITY OF PENNSYLVANIA
6
Fitting Curves
Statistics 102, Fall 2014
VALUE OF RESIDUAL PLOTS
Nonlinear patterns are not always visible in the scatterplot itself, and only become apparent in the detail offered by the residual plot. Consider the following men's records in track events
Time (seconds)
6000
5000
4000
3000
2000
1000
0 0
5000 10000 15000 20000 25000 30000 Distance (m)
With R2 so large, can there possibly be a problem with this model?
THE WHARTON SCHOOL OF THE UNIVERSITY OF PENNSYLVANIA
7
Fitting Curves
Statistics 102, Fall 2014
ANOTHER NONLINEAR REGRESSION
Common question: what is the optimal amount of promotion for a product?
Specific case: A chain of liquor stores needs to know how much shelf space in its stores to devote to showing a new wine to maximize its profit. Space devoted to other products brings in about $50 of net revenue per linear foot.
The data Display.jmp has weekly sales ($) and shelf-feet from 47 stores of the chain. Should we expect a linear relationship between promotion and sales, or should we expect diminishing marginal gains?
$450 $400 $350 $300 $250 $200 $150 $100
$50 $0 012345678 Display Feet
Visually, what do the data suggest for the optimal shelf space?
Sales
THE WHARTON SCHOOL OF THE UNIVERSITY OF PENNSYLVANIA
8
................
................
In order to avoid copyright disputes, this page is only a partial summary.
To fulfill the demand for quickly locating and searching documents.
It is intelligent file search solution for home and business.
Related download
- learning from data lecture 10 nonlinear transforms
- using your ti nspire calculator linear correlation and
- nonlinear models fitting curves
- nonlinear programming 13
- fitting models to biological data using linear and
- using solver in excel for non linear curve fitting
- 1 multiple regression nonlinear regression regression
- chapter 315 nonlinear regression statistical software
- section i linear regression
- polymath tutorial on non linear regression example 7 4
Related searches
- linear thinking vs nonlinear thinking
- linear and nonlinear thinking styles
- linear and nonlinear model
- linear and nonlinear meaning
- linear and nonlinear regression
- multivariable nonlinear regression
- nonlinear multiple regression
- nonlinear regression coefficients
- nonlinear regression r
- multiple variable nonlinear regression excel
- nonlinear regression stata
- linear regression vs nonlinear regression