1

I am running regressions where the dependent variable is continuous between -1 and 1. Do I need to somehow transform the dependent variable or can I just work with the variable as it is?

Nick Cox
  • 56,404
  • 8
  • 127
  • 185
user111237
  • 51
  • 3
    So are you using some tan-sig expression as the general form, or a logistic form? What is the nature of the inputs? Every story has an intro, build-up, climax, resolution and epilogue. This question does not; and some folks here need more around a question than that. What is the back-story, the motivation, the purpose. That ephemeris can help inform amazing answers, not just monosyllabic ones. – EngrStudent Apr 08 '16 at 11:16
  • 4
    Short question, so short response. With the variable as it is, the problem with regression is that predictions are not obliged to respect the bounds and indeed linear functional form is less likely to be suitable. Also, the error assumptions are likely to be off. As here http://stats.stackexchange.com/questions/206095/alternative-to-glmm-for-normalised-ratio-bounded-1-to-1-response-variable note that a transformation from $y$ to $(y + 1)/2$ will take you into the realm of fractional regression with logit, probit or other link. – Nick Cox Apr 08 '16 at 11:34
  • 1
    Also consider ordinal regression, e.g., proportional odds logistic regression model. – Frank Harrell Apr 08 '16 at 12:37
  • 1
    As the first comment suggests, knowing something about how this dependent variable ends up restricted to between -1 and 1 would help provide a better answer. In the particular case like this that is linked by @NickCox, there seems to be a very simple answer. – EdM Apr 08 '16 at 14:58

0 Answers0