PRACTICAL SOCIAL INVESTIGATION:

USING SPSS FOR WINDOWS (Versions 8 and 10) TO ANALYSE MULTI-WAY CROSS-TABULATIONS WITH LOG-LINEAR MODELS

As was shown in an example in Chapter 9, the gender difference in the relationship between occupational class and self-rated class needs to be tested for significance using a log-linear model. To do this, one clicks on Statistics (or Analyze if using Version 10), then on Loglinear, and then on Model Selection... The next step is to click in turn on rrg2, src2 and rsex ('Respondent's sex') in the alphabetical variables list, and move each of these across to the Factor(s): box by clicking on the little triangle. SPSS then needs to be told that each variable has two categories, 1 and 2. This is achieved by clicking on one of the variables, and then clicking on Define Range.... The next step is to type 1 in the Minimum: box, and then to click on the Maximum: box and to type 2 in it. After clicking on Continue these steps can be repeated for the other two variables. The default settings are appropriate, so the model can then be fitted by clicking on OK.

Log-linear model output in SPSS is quite extensive. (In fact, you may need to double click within the output window {SPSS Viewer} in order to be able to scroll down the whole of the output). For our purposes the most important pieces of output are where it is stated that "The final model has generating class RRG2*SRC2*RSEX" and where the output says "If Deleted Simple Effect is...", followed by "RRG2*SRC2*RSEX" and these headings and corresponding figures: "df" = "1"; "LR Chisq Change" = "6.045"; "Prob" = ".0139". What this shows is that SPSS has attempted to simplify the model of the three-way cross-tabulation by removing the RRG2*SRC2*RSEX interaction, i.e. by hypothesising that the relationship between occupational class and self-rated class does not vary according to sex. However, the fit of the model, as judged by the change in the Likelihood Ratio Chi-Square measure (6.045 for 1 degree of freedom), deteriorates significantly (P = 0.0139 < 0.05) if this interaction term is removed. In other words, there is a statistically significant gender difference in the relationship between occupational class and self-rated class.

By default SPSS uses backwards elimination to fit a series of log-linear models to a specified multi-dimensional cross-tabulation, starting with the (most complex) saturated model, which contains all the possible relationships between the variables, and ‘throwing out’ relationships which are not found to be statistically significant. If the variables prsoccl ('Parents' social class [self rated]') and partyid2 ('Partyid.compressed A+B only derived') are recoded into appropriate dichotomous (i.e. two-category) versions, and inserted (together with rrg2 and src2 and instead of rsex ('Respondent's sex') ) into the Factor(s): box within the Model Selection Loglinear Analysis box, then the model selection process which led to Table 9.17 can be reproduced.

The above example can be reproduced using the commands in the file catex.sps which can be downloaded from these web pages and read into an SPSS syntax window. Syntax windows are described in the next page (synwins.html).