Cross Section Data Methods

The emphasis of this section is on specialized techniques which are applied mainly (or solely) to cross sectional data: data with a single observation for a number of individuals (or firms or countries, etc.). Data which also have a "time series" component (following each single individual over time) have special characteristics and are covered in the Panel and Grouped Data page.

The textbook replication files for Wooldridge's (2010) Econometric Analysis of Cross Section and Panel Data, 2nd Edition and Greene's (2012) Econometric Analysis, 7th Edition include many examples of the techniques described here and others that are more advanced.

In addition to the specialized methods covered here, cross section data sets are also more likely than time series data sets to require use of weighted least squares, tests for heteroscedasticity and subsample selections.

Cross-Sectional Data

Cross-sectional data sets consist of a single observation for a number of individuals (or firms or countries, etc.). Do not use a CALENDAR instruction for cross-sectional data. Set the number of observations to process either with ALLOCATE or on the DATA instruction or let DATA figure it out for you. For instance, suppose you have data on the number of males and number of females in fifty states. You could either do something like:

allocate 50

open data statepop.dat

data(format=free,org=obs) / males females

open data statepop.dat

data(format=free,org=obs) 1 50 males females

or even

open data statepop.dat

data(format=free,org=obs) / males females

By omitting the CALENDAR instruction, you are telling RATS that your data set has no time series properties. RATS will label entries with the entry number alone. Note that on a LINREG (or other least squares estimator), RATS will still calculate a Durbin–Watson statistic. This is likely to have little value, but if your data set is ordered on a particular variable, the Durbin–Watson may be able to detect a specification error related to that variable. (It can't be "serial" correlation, but there can be correlation in the ordering variable).

The special topics are

Probit and Logit Models

Censored and Truncated Samples

Hazard Models