I want to run a linear regression in SPSS

N = 1400

Outcome variable = rating from 0 to 800 (participants saw or heard a Mandarin speaker and had to rate how pleasant the speaker was feeling)

Predictors =

  1. "Status of Mandarin", i.e. 3-level nominal variable (whether the participants are native speakers / learners / non-speakers of Mandarin)
  2. "Condition", i.e. 4-level nominal variable (whether the participants saw a video recording / heard and audio recording / saw & heard a video recording with sound / only heard an unclear [i.e. low-pass filtered] audio recording)
  3. Culture, i.e. 10-level nominal variable (1O cultural clusters, some participants not being categorised in any of these clusters)

I dummy-coded all variables to include them in a linear model

Frequencies Culture

As you see from this table, some levels of the "Culture" variable have too few observations, so basically I think that can only include the Confucian, the Anglo, and the Latin Europe groups. How do I deal with that with dummy coding? Should I include all 9 dummy-coded Culture variables in the model (excluding Confucian as I want that group as baseline) and only interpret the significance value and coefficients for the Anglo dummy variable and for the Latin Europe dummy variable?

I am also planning to look at interactions between Status of Mandarin and Condition (and ideally also between Status of Mandarin and Culture + between Condition and Culture, but I am not sure whether I can).

I am using SPSS at the moment but might switch to R.

In response to a comment by @rolando2, I am adding a picture of the instrument showing the gliding scales used to collect the ratings. I indeed will not consider a difference of, say 20, as meaningful. Instrument

  • $\begingroup$ Could you explain what you mean by "too few"? Too few for what purpose or according to what criterion? $\endgroup$ – whuber Nov 4 '18 at 16:22
  • 1
    $\begingroup$ Actually you caught me here: I thought that it was a generally assumed rule of thumb that each dummy variable should have minimum 15% of the sample, but a quick search on the internet shows me that this rule is not assumed by everyone (and confuses me even more). $\endgroup$ – Pernelle Nov 4 '18 at 16:46
  • $\begingroup$ The problem stays the same though, as some participants are not categorised in any of these categories (should I have an additional dummy variable "Other" with the rest of the sample?) $\endgroup$ – Pernelle Nov 4 '18 at 16:46
  • $\begingroup$ A 15% rule would have no basis in any legitimate theory. Possibly in certain applications with certain standard objectives and standard sample sizes such a rule could be helpful, but otherwise it ought to be ignored. $\endgroup$ – whuber Nov 4 '18 at 17:17
  • $\begingroup$ So I could simply include all dummy variables in the analysis (and do I need to include an "other" dummy variable to include the part of the sample that has not been categorised in any of these culture groups?) $\endgroup$ – Pernelle Nov 4 '18 at 18:04

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.