I wonder if it is allowed to log a variable which is already logged (transformed). When I do this, the skewness of my model goes to 0.05 which is quite good. Let me know! Thanks in advance :)
Is it allowed to log a already (logged)transformed continuous variable for low skewness? [duplicate]
$\begingroup$
$\endgroup$
6
-
2$\begingroup$ Sure. Everything is 'allowed' as long as you aren't breaking math. But does it actually make sense in the context of your problem? That's not something we could answer with the info provided. $\endgroup$– DasonCommented Dec 18, 2020 at 16:41
-
$\begingroup$ Hi Dason, thank you for your reaction! The context is as following: Your client has commissioned you to conduct a survey study on the determinants of “consumer preferences” (Y-variable) for a new health service. The client wants to know the (potential) indirect effect of consumer “values” (X-variable) on these preferences via the mediator “confidence” in one’s own health (M-variable): a * b. I have 2 continuous variables, V16 (age) and V17 (mood) and they are already transformed, says the assignment. However transforming them again leads to a perfect skewness... $\endgroup$– MmijCommented Dec 18, 2020 at 16:54
-
$\begingroup$ @Dason, am I still allowed to log here? :) $\endgroup$– MmijCommented Dec 18, 2020 at 17:19
-
1$\begingroup$ Can you edit your question to include the context you gave in a comment? $\endgroup$– Ben BolkerCommented Dec 18, 2020 at 18:54
-
1$\begingroup$ Is the already logged variable strictly positive? Why do you consider that a variable having small skewness is "good"? $\endgroup$– Chris HaugCommented Dec 18, 2020 at 18:57
|
Show 1 more comment
1 Answer
$\begingroup$
$\endgroup$
Transforming the predictor variables (independent variables) to achieve a particular distribution does not make sense. We don't need particular distributions of the predictors to do regression. When we make a normality assumption, it is about the error term, not the predictors.
There might be cases where theory says that a transformed predictor variable makes more sense. There might be cases where you determine that a transformed variable results in a model with stronger performance.
But transforming the variable just because you want something that isn't skewed? I'll pass.
