I have used OLS and GWR for validating the dependencies between two distinct database. The Residual Square for GWR is 0.82 and thereby making it the right regression model to be used for determining the relationship between the two datasets.

What I wanted to know that, GWR being Local regression and OLS being global regression, which should be used where and when?

Also what does it really mean if Moran's I for the GWR model is Random?


What these procedures are

Although OLS and GWR share many aspects of their statistical formulation, they are used for different purposes:

  • OLS formally models a global relationship of a particular sort. In its simplest form, each record (or case) in the dataset consists of a value, x, set by the experimenter (often called an "independent variable"), and another value, y, which is observed (the "dependent variable"). OLS supposes that y is approximately related to x in a particularly simple way: namely, there exist (unknown) numbers 'a' and 'b' for which a + b*x will be a good estimate of y for all values of x in which the experimenter may be interested. "Good estimate" acknowledges that the values of y can, and will, vary from any such mathematical prediction because (1) they really do--nature is rarely as simple as a mathematical equation--and (2) y is measured with some error. In addition to estimating the values of a and b, OLS also quantifies the amount of variation in y. This gives OLS the capability to establish the statistical significance of the parameters a and b.

Here is an OLS fit:

enter image description here

  • GWR is used to explore local relationships. In this setting there are still (x,y) pairs, but now (1) typically, both x and y are observed--neither can be determined beforehand by an experimenter--and (2) each record has a spatial location, z. For any location, z (not necessarily even one where data are available), GWR applies the OLS algorithm to neighboring data values to estimate a location-specific relationship between y and x in the form y = a(z) + b(z)*x. The notation "(z)" emphasizes that the coefficients a and b vary among locations. As such, GWR is a specialized version of locally weighted smoothers in which only the spatial coordinates are used to determine neighborhoods. Its output is used to suggest how values of x and y covary across a spatial region. It is noteworthy that often there is no reason to choose which of 'x' and 'y' should play the role of independent variable and dependent variable in the equation, but when you switch these roles, the results will change! This is one of the many reasons GWR should be considered exploratory--a visual and conceptual aid to understanding the data--rather than a formal method.

Here is a locally weighted smooth. Notice how it can follow the apparent "wiggles" in the data, but does not pass exactly through every point. (It can be made to pass through the points, or to follow smaller wiggles, by changing a setting in the procedure, exactly as GWR can be made to follow spatial data more or less exactly by changing settings in its procedure.)


Intuitively, think of OLS as fitting a rigid shape (such as a line) to the scatterplot of (x,y) pairs and GWR as allowing that shape to wiggle arbitrarily.

Choosing between them

In the present case, although it is not clear what "two distinct databases" might mean, it seems that using either OLS or GWR to "validate" a relationship between them may be inappropriate. For instance, if the databases represent independent observations of the same quantity at the same set of locations, then (1) OLS is probably inappropriate because both x (the values in one database) and y (the values in the other database) should be conceived of as varying (instead of thinking of x as fixed and accurately represented) and (2) GWR is fine for exploring the relationship between x and y, but it cannot be used to validate anything: it's guaranteed to find relationships, not matter what. Moreover, as previously remarked, the symmetric roles of "two databases" indicate that either could be chosen as 'x' and the other as 'y', leading to two possible GWR results which are guaranteed to differ.

Here is a locally weighted smooth of the same data, reversing the roles of x and y. Compare this to the previous plot: notice how much steeper the overall fit is and how it differs in the details, too.

Lowess 2

Different techniques are required to establish that two databases are providing the same information, or to assess their relative bias, or relative precision. The choice of technique depends on the statistical properties of the data and the purpose of the validation. As an example, databases of chemical measurements will typically be compared using calibration techniques.

Interpreting Moran's I

It is hard to tell what a "Moran's I for the GWR model" means. I guess that a Moran's I statistic may have been computed for the residuals of a GWR calculation. (The residuals are the differences between actual and fitted values.) Moran's I is a global measure of spatial correlation. If it is small, it suggests that variations between the y-values and the GWR fits from the x-values have little or no spatial correlation. When GWR is "tuned" to the data (this involves deciding on what really constitutes a "neighbor" of any point), low spatial correlation in the residuals is to be expected because GWR (implicitly) exploits any spatial correlation among the x and y values in its algorithm.

  • So in GWR you said that reversing the variables gives differing results but the one, which gives higher residual square, does it not mean that it shows stronger relationship between the two?
    – Sam007
    Jun 12 '12 at 16:14
  • Sam, a low residual square in itself does not indicate a stronger relationship. In particular, when you reverse the roles of x and y, you can't even compare the residual squares--they frequently are in different units. (E.g., one might be a square temperature and the other might be a square amount of rainfall: how do you tell which one is lower?) You can always reduce a residual square by including more parameters in a model, even if they are meaningless: going too far in this direction is called "overfitting." In some sense, GWR with a short spatial range is a form of overfitting.
    – whuber
    Jun 12 '12 at 16:36
  • 2
    You may be thinking of R-squared, Sam: it is a ratio. (I understood "residual squared" to be exactly what it says: the sum of squares of residuals. Most stats software reports this statistic along with R-squared values.) But it's still dangerous, and usually wrong, to compare different models (such as y in terms of x versus x in terms of y) in terms of R-squared: see stats.stackexchange.com/questions/13314. Because GWR is exploratory, it's great for finding patterns and hypothesizing relationships, but (as usually practiced, anyway) it's unsuitable for substantiating claims.
    – whuber
    Jun 12 '12 at 17:25
  • 1
    WOW that was really all statistics. Actually the reason I am finding it hard is becoz i have very poor background in statistics, so i m solely relying on the models results, without understanding what they exactly mean. Most of the stuff on the R2, went over my head. Could you suggest some good beginners book to get me started and build my foundation in statistics?
    – Sam007
    Jun 12 '12 at 17:38
  • 2

Rsquared should not be used to compare models. Use log likihood or AIC values.

If your residuals in GWR are random, or I guess appear to be random (not statistically sig.) than you might have a specified model. It at least suggests that you don't have correlated residuals and should suggest that you do not have any omitted variables.

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.