4.2 Random effects?

When looking at any study in Linguistics (and beyond), we rarely use productions of one vowel, from one speaker and from one item. If this were the case, we are unable to quantify changes in specific languages. To be able to generalise our results, we go for data collected from:

  1. Multiple speakers
  2. Multiple vowels and consonants
  3. Multiple Items (words)
  4. Multiple utterances where words are embedded
  5. Multiple listeners in perception experiments.

In the last case, when designing our perception experiment, we can sometimes use multiple items, coming from multiple utterances and from multiple speakers!

If we do not account for these inter-dependencies in our dataset, we are increasing Type I Error.

Type I error, is when you easily find statistical differences when they are not there. Type II error is when you fail to find statistical difference when it is there. There are other types of errors see this reference for more details.

4.2.1 How to choose fixed and random effects

Fixed effects are those that are part of the experimental conditions. If you have exhausted all levels of an experimental condition, then this goes into fixed effects. Random effects are random selections of the population you have and you want to generalise over them.

E.g., Speakers, listeners, items, utterances are all random effects because you are not using all the population of speakers, listeners, items, or utterances in your data!

BUT.. Can Speakers, listeners, items, or utterances be included as fixed effects? Yes!! When you do this, it means you are interested in this specific population and want to evaluate differences specific to the population!!

4.2.2 What about Random Intercepts and Random Slopes

Random Intercepts are used to obtain averages of your population and these are used in your statistical model to estimate variations.

Random Slopes are adjustments to your participants’ observations as a function of your variables of interest. Usually, any within-subject (or within-item) variable is to be included as a random slope, but you need to use model comparison to evaluate the need to use it.