More Control Groups Gone Wild

We had a client, come in to consulting recently, who was studying people receiving treatment in mental health clinics. He classified these patients into three groups; Group 1) individuals who had a personal history of depression, group 2) individuals with a family history of depression and Group 3) individuals with no history of depression. The last group was the control group. The outcome variable was a binary indicator, whether or not they had experienced a depressive episode since their last visit to the clinic.

The problem was that the control group did not experience any depressive episodes. This, in turn, creates a problem for logistic regression. There was an error message indicating that group 3 not equal zero predicts failure perfectly. And, instead of two degrees of freedom for group there was only one degree of freedom (comparing Group 1 versus Group 2) and no coefficient for Group 3 versus Group 1.

This could be dealt with by changing one response score in Group 3 at random from zero to one. There was a further complexity however. Each individual in each of the three groups was measured on 12 occasions, i.e., once a month for a year. And during those twelve months none of the individuals in the control group ever experienced a depressive episode. Since change over time was one of the research questions, it didn't seem right to randomly chance responses in the control group to one.

In the end, there just wasn't not any useful information available from Group 3. It was clearly one more case of control groups gone wild.

pbe