In addition to least squares and absolute deviation regression, weighted least squares estimation is probably the most commonly used technique. Ordinary least squares techniques assume that the residual variance around the regression line is the same across all values of the independent variable(s). Put another way, it is assumed that the error variance in the measurement of each case is identical. Often, this is not a realistic assumption; in particular, violations frequently occur in business, economic, or biological applications. (Note that weighted least squares parameter estimates can also be computed via the Multiple Regression module.)

For example, suppose we wanted to study the relationship between the projected cost of construction projects, and the actual cost. This can be useful in order to gage the expected cost overruns. In this case it is reasonable to assume that the absolute magnitude (dollar amount) by which the estimates are off, is proportional to the size of the project. Thus, we would use a weighted least squares loss function to fit a linear regression model. Specifically, the loss function would be (see, for example, Neter, Wasserman, & Kutner, 1985, p. 168):

Loss = (Obs-Pred)2 * (1/x2)

In this equation, the loss function first specifies the standard least squares loss function (Observed minus Predicted squared; i.e., the squared residual), and then weighs this loss by the inverse of the squared value of the independent variable (x) for each case. In the actual estimation, the program will sum up the value of the loss function for each case (e.g., construction project), as specified above, and estimate the parameters that minimize that sum. To return to our example, the larger the project (x) the less weight is placed on the deviation from the predicted value (cost). This method will yield more stable estimates of the regression parameters (for more details, see Neter, Wasserman, & Kutner, 1985).