Interpretation of Fixed effects Linear Mixed Model Dummy coded variables

juliaaline

Dear Community,

 

I have got a question regarding the interpretation of a main and interaction effects of Mixed Model fitted for reaction times. I have to dummy coded variables.

I analyse the reading times of sentences. The first variable repesents the valence of the first part of the sentence (positive or negative) while the other is coded the same but for the second part of the sentence. Like this I have got four sentence types. nn, pn, np, pp. I want to test whether congruent (nn, pp) sentences are processed faster that incongruent ones ( np, pn). This effect should also be significant for comparisions like this: nn<np and pp<pn. 

So my first question is a rather general one. With my coding as it is, can I interpret the main and interaction effects meaningfully? Regarding of the outpout of jmp how would I interpret the results of the tests of the fixed effects? (I am sorry for this basic question, I really already put a lot of thought in that but somehow feel very unsure about the interpretation at the moment)

 

Also would it suffice to interpret the pairwise comparisions? Would it be possible to do a pairwise comparision of two agains two? So like this: nn=pp<pn=np?

 

I really would be very grateful for any hint.

gzmorgan0

Julia, some important details seem to be missing or I do not understand your experiment. Did you have just 4 sentences (nn, np, pn, pp) and each person has a reading time for each? (a randomized block experiment)

 

Did each measurement represent a unique person and sentence? 

 

Were there more than 4 sentences, where sentence might have an effect (ex. 24 sentences, 6 of each type, where one of the 6 nn sentences had a much longer reading time than the other 5 nn sentences).

 

Before looking at comparative tests, you should always check for validity of assumptions, and data errors.  Look at the variability plot of Reading Time for Y and Question, Type(nn,np,pn,pp). If each person reads more than one sentence, the person/subject should also be a grouping column.     

 

Attached is a simulated data set where 30 subjects each read 4 sentences (order was randomized). The simulated effect was Congruent Sentences (20,2) and Incongruent Sentences(26,3).  Scripts for example graphs and analyses are shown.  This might not match your experiment nor coding, but might give you some ideas.

 

Note, this was simulated so that only the interaction is significant. With simple factors such as these, somethimes it is easier to present results using one factor (Type) with levels, nn, np, pn, pp. When the global test is significant (controlling for type 1 errors), then use a multiple comparison test. 

ExampleRandomizedBlock.jmp