So I have a very positively skewed distribution of latency data (pasted below). I've read that transforming the data to normality can have some unfortunate consequences in regression models. Any ideas on the distribution that I should use in a generalized linear model?

enter image description here

  • 1
    $\begingroup$ The reciprocal of latency often follows a normal distribution. $\endgroup$ – HEITZ Aug 7 '17 at 22:51
  • 1
    $\begingroup$ The appearance of the marginal distribution is affected by the pattern of predictors (you can get a quite skewed appearance for the raw DV even when the conditional distributions are not skew at all or even skewed the opposite direction, in which case you could end up trying to transform the wrong way). In this case, however, we would expect the conditional distribution of latencies to be right skew. Being times, that's quite common. Speeds (reciprocal of times) are often better behaved and easy to model; sometimes log-times are a reasonable choice. ... ctd $\endgroup$ – Glen_b Aug 8 '17 at 1:26
  • $\begingroup$ ctd... You could instead consider using GLMs which can deal with skewed conditional distributions as well as with the fact that the conditional means cannot actually go negative (even though a linear model could predict exactly that). GLMs would probably have been my first instinct here; you might even want to consider doing both (model inverse of latency with a GLM) $\endgroup$ – Glen_b Aug 8 '17 at 1:26

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.