## 2004-05-12

SPSS had a nice web seminar on their Regression and Advanced Statistics modules. Here are some notes that I took while sitting in.

The Regression module has several advanced regression methods:

• Binary logistic regression, which is used when the outcome is a binary variable (two levels)
• Multinomial logistic regression, which is used when the outcome is categorical, typically with more than two levels.

More basic regression models, such as linear regression, appear in the base SPSS package.

I have a training class on binary logistic regression on my web pages.

Advanced Models has several other methods:

• Repeated measures, which is used when you have multiple measurements of the same outcome on each patient in your study,
• Nested ANOVA, which is used when some factors in a study are contained within other factors. For example, many educational studies look at different schools as one factor and classrooms within each school as a nested factor.
• Ordinal logistic regression, which is used when the outcome is categorical, typically with more than two levels, and where there is a logical ordering to the variables.
• Survival data models, which I will discuss in detail below.

I have a training class on survival data models on my web pages. The example given in class is an interesting one, where people are promoted after a training program. Some are promoted early, others are promoted late, and some left for other jobs before they had a chance to be promoted. People who left might have been promoted eventually. We don’t know when. All that we do know if that they did not get promoted during the time they were here. A survival model will factor that person in to help estimate promotion rates up to and including the day that they left. After they leave, they no longer provide information on promotion rates.

These people who leave before they get a promotion are referred to as censored observations. There may also be censored observations because we have to stop a study before everyone gets promoted. An important assumption for survival data is that censoring is independent of outcome. In this case, it means that people left for reasons unrelated to whether they might get a promotion sooner rather than later. This is worth investigating; perhaps people left because they were frustrated and didn’t expect to get promoted anytime soon.

Survival models are used a lot in cancer studies and you should also investigate whether censoring is independent of outcome. If, for example, patients drop out of a study because they went to Mexico for laetrile treatment, they would be considered a censored observation. But more likely than not these patients dropped out because the treatment they were receiving in the study was not going well.

You can find an earlier version of this page on my original website.