Sie sind auf Seite 1von 3

Introduction to Linear Regression

http://elearning.algonquincollege.com/coursemat/viljoed/gis8...

Introduction to Linear Regression


(http://gis.nrcan.gc.ca/~viljoen/gis8746/regress)

Purpose of tting equations to data: 1. 2. 3. 4.


To summarize a mass of data in order to obtain interpolation formulas or calibration curves Conrm or refute a theoretical relation Compare several sets of data in terms of the constants in their representing equations Aid in the choice of a theoretical model

In terms of summarizing a mass of data, linear regression coefcients a, b, and R2 provide an idea of the relationship between two data sets (in raster land, two images). However, this relationship can be misleading. That is, data with the same a, b, and R2 may be very different.

A good method of tting should: 1. 2. 3. 4.


Use all relevant data in estimating each constant Have reasonable economy in the number of constants required Provide some estimate of uncontrolled error Make it possible to nd regions of systematic deviations from the equation if any exist

Linear regression analysis steps: Overview


1. Plot and evaluate a scatter diagram 2. Perform linear regression analysis
(give me an "a", give me a "b", what do you have, linear regression, yahoo!)

3. Create a model dependent image using the regression equation 4. Compare the model image to the original dependent image 5. Optional: Eliminate areas that do not t the model and go back to Step 1.

Linear regression analysis steps: In detail


1. Plot and evaluate a scatter diagram to determine whether the relationship between the two data sets is linear. (Remember, this is LINEAR
regression analysis) a. If it is linear, proceed to Step 2. b. If it is not linear, it may be linearizable. i. Compare the scatter diagram with the gure of linearizable functions and select the function that best matches the pattern in the scatter diagram. ii. Apply the appropriate transformation equation to linearize the relationship between the two data sets. Proceed to Step 2. Perform linear regression analysis (e.g. REGRESS in Idrisi) to compute values of the intercept with the "y-axis" (a) and the slope of the line (b). a. If the original data was linear, the values of "a" and "b" may be substituted directly into the general equation: y = a + bx b. If the original data was linearized (i.e. a transformation equation was used to linearize the data), transform the values of "a" and "b" so that they can be used in the general equation for the function selected. For example, if the relationship between the original data was exponential, a linear form of the transformed data would give values of "a" and "b" according to the following equation: y = ln a + bx. A value for "a" to be used in the general equation y = aebx is computed using eln a , since a = exp(ln a)

2.

1 of 3

1/15/13 2:19 PM

Introduction to Linear Regression

http://elearning.algonquincollege.com/coursemat/viljoed/gis8...

3. Create a model dependent image using the regression equation values of "a" and "b". In the equation y = a + bx, y represents a 4. 5.
"dependent" image and x represents an "independent" image. Use Idrisi OVERLAY and TRANSFOR programs to create the "dependent" image. Compare the model image to the original dependent image. This should give you an idea of the model accuracy. An analysis of the residuals is useful in assessing model accuracy. Optional: Eliminate areas that do not t the model and go back to Step 1. If the model image does not seem to work in the areas that you wanted it to, it may be appropriate to mask the areas of interest and repeat the regression analysis. If this still yields unacceptable results, some other form of modelling may be appropriate (e.g. process model).

Linear regression analysis computations: In detail


Equation of a linear model

Best t line from minimizing the sum of squares of residuals

where,

The predicted value of the i'th observation predicted by this model will be:

The correlation coefcient (index of the goodness of t) is dened by:

2 of 3

1/15/13 2:19 PM

Introduction to Linear Regression

http://elearning.algonquincollege.com/coursemat/viljoed/gis8...

An evaluation of the t must also include an analysis of the residuals:

where,

References
Anscombe, F.J. (1973) "Graphs in statistical analysis.", American Statistician, Vol. 27, p. 17-21 Chatterjee, Samprit, Price (1977) Regression analysis by example. John Wiley & Sons. Daniel, Cuthbert, & Wood (1980) Fitting equations to data. John Wiley & Sons.

viljoen@gis.nrcan.gc.ca 03/02/98

3 of 3

1/15/13 2:19 PM

Das könnte Ihnen auch gefallen