Log-Linear Analysis Model Specification - Advanced Tab

Select the Advanced tab of the Log-Linear Model Specification dialog box to access the options described here.

Specify model to be tested. Click the Specify model to be tested button to display the Specify Model to be Tested dialog box, in which you enter the numbers of the factors involved.

Test all marginal & partial association models. Click the Test all marginal & partial association models button to produce a spreadsheet containing the results of the test. The partial association between, for example, two factors 1 and 2 (denoted by 12) is computed by comparing the fit (i.e., evaluating the Chi-square difference) of the model that includes all two-way interactions with the model that excludes the interaction between factors 1 and 2. The marginal association between, for example, two factors 1 and 2 is computed by comparing the fit of the model including all main effects (i.e., all effects of lower order than the one of interest) with the model including the interaction between factors 1 and 2.

Note that these tests are actually computed only the first time you click the Test all marginal & partial association models button. Thereafter, unless you return to the Startup Panel, the tests are stored in memory and are available instantaneously. The utility and interpretation of these tests is described in the Introductory Overview.

Automatic selection of best model. Click the Automatic selection of best model button to display the Automatic Selection of Best Model dialog box. The logic of the automatic selection algorithm is described in Log-Linear Analysis of frequency tables Introductory Overview - automatic model fitting.

Structural zeros. Click the Structural zeros button to instruct STATISTICA to "find" all zero frequencies in the table and display them in the Select Structural Zeros dialog box. You can then select the structural zeros from that list.

Delta. The constant Delta (0.5 by default) is added to all frequencies in the observed table before it is submitted to the actual analysis. This is a recommended procedure when the table contains several cells with low frequencies (e.g., less than 10); it is a correction analogous to the Yates' correction for two-way tables, and this correction does not affect the results unless the table contains low frequencies.

Note: Be sure to set Delta to zero if you want to use this module as a tabulation program. Otherwise, you will not obtain accurate cell (raw) frequencies and marginal counts when requesting those tables from the Results dialog box.

Maximum number of iterations. Fitting a model (i.e., marginal tables) to a multi-way frequency table is accomplished via an iterative procedure, the so-called iterative proportional fitting. The default setting of 50 iterations will practically always ensure that this iterative procedure will converge, that is, that the respective model will be fitted to the table with sufficient accuracy.

Convergence criterion. Fitting a model (i.e., marginal tables) to a multi-way frequency table is an iterative procedure (iterative proportional fitting). At each iteration, the fitted marginal tables (i.e., the fitted model) will approximate more closely the observed marginal tables. The iterative procedure will terminate when there is no difference between fitted and observed margins that is greater than the convergence criterion. The default value (0.01) ensures an accuracy that is usually more than sufficient.