space, land area, age, and number of units. In an apartment rent prediction model based

on 4,500 apartment complexes in eight markets, Valente et al. (2005) use the log of

asking rent as the dependent variable. For regressors they use square footage per unit,

number of floors, property age, submarket dummies, and year of sale dummies.

 Building age, age squared, the square footage of improvements, building foot

print, lot size, number of units, and number of floors, are some of the most common

structural characteristics used in commercial property price or rent equations (see, for

example, Colwell, Munneke and Trefzger, 1998, and Saderion, Smith and Smith, 1993).

The choice of functional form is also very important in order to ensure that the model is

correctly specified. Weirick and Ingram (1990) provide an excellent analysis of various

approaches to functional forms in hedonic regressions, when the dependent variable is

selling price. In particular, the authors compare three standard functional forms:

* A linear model
* A semi-log model which uses the logarithmic transform of the dependent variable
 (selling price)
* A log-linear model, which uses logarithmic transforms of both the dependent
 variable as well as independent variables

 As Weirick and Ingram (Ibid.) point out, the linear form has "serious deficiencies

from a market theory standpoint." Such models force the value of an extra square foot of

improvement for a 2,000 sq. ft. property to be the same as the value of an extra square

foot for a 10,000 sq. ft. property. The semi-log and log-linear models take into account

nonlinearities in the data. In addition, by using quadratic transformations of explanatory

variables (such as square footage and lot size) I can capture property value relationships

that are concave or convex in certain characteristics (Ibid.).