Say you have IQ and Academic Achievement (AA) and you create a partial correlation with both of those variables and covary them with Socioeconomic Status (SES).

From this video ( https://www.youtube.com/watch?v=GUNPXLRk_60 ), I learned that, if you want to graph that partial regression in SPSS, then you can use the residuals of two regressions, IQ with SES and AA with SES.

Thus, if I am understanding this correctly, the residual of each of those regressions is the value of the dependent variables (IQ and AA) controlled for SES, correct?

This leads me to my current question. Can you control for variables in a linear discriminant analysis? In SPSS, there is no option to, which makes sense because of the nature of the classification function. However, what if I regressed each of the individual dependent variables in the LDA on typical covariates, such as age, sex, and race, then used those residuals of the dependent variables for the discriminant function?

Would that technically control for the variables or is there some problem with this that I am not seeing?

  • $\begingroup$ Discriminant analysis controls for all the variables that are used to create the discriminant functions. Perhaps you mean to stratify by some variable? $\endgroup$
    – Peter Flom
    Sep 7, 2017 at 15:52
  • $\begingroup$ why not to just add the confounding variables as additional covariates? $\endgroup$
    – Krrr
    Sep 7, 2017 at 15:53
  • $\begingroup$ As I understand it, adding the confounding covariates is problematic. For example, by adding covariates such as sex to a discriminant analysis, we are creating a misleading increase in classification accuracy for discriminating diseases and disorders that are more prevalent in men than in women. This paper Adjusting for matching and covariates in linear discriminant analysis goes into more detail than I understand about this dilemma. I am currently trying to understand and compare their method to mine to see if I can resolve this issue further on SPSS. $\endgroup$ Sep 7, 2017 at 20:23
  • $\begingroup$ I am not sure if I understand this correctly (haven't looked at the paper yet). But if you have prior knowledge about any biases in the data then why not to use appropriately weighted performance measures and then tune any parameters accordingly? $\endgroup$
    – Krrr
    Sep 8, 2017 at 11:16
  • $\begingroup$ I am using cognitive measures in my LDA. Those cognitive measures can be calibrated for age, sex, and race. However, we are also putting olfactory measure in the LDA, which are affected by cigarettes per day. My understanding is that I would have to control for all variables (age, sex, race, Cigs per day) evenly across all cognitive measures and olfactory measures in order to stay consistent. Though, I could be misunderstanding my problem. How would you weight the cognitive variables? I am unfamiliar with that technique? $\endgroup$ Sep 8, 2017 at 15:27


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Browse other questions tagged or ask your own question.