Loglinear analysis and modeling
(Press Control while clicking each link and save the file.). Recommended SPSS commands can be found here.
Because the partition into person and situation causes makes most sense for unintentional behaviors, we selected a clearly unintentional behavior (mood) and asked Stanford undergraduate students to explain their own mood (actor perspective) or someone else's mood (observer perspective). To control for mood valence, they were first asked in what mood they were or the other person was (good, so-so, bad) and then explained why they (or the other person) were in that mood.
Explanations were classified by four coders. To test H1, they classified them into person cause (P), situation cause (S), or mixed PS cause (which we'll ignore for the analysis). To test H2 they also classified P causes into specific subtypes: behavioral and internal state (both unstable) and trait (stable). According to the classic literature, two patterns should hold: Observers should cite more person (vs. situation) causes than actors (H1), and observers should cite more stable (vs. unstable) P causes than actors (H2). There were no hypotheses about valence of mood, but one would hope that any detected actor-observer asymmetries do not interact with mood.
Analyze file 1 first. Form a new variable that only analyzes for person vs. situation causes, leaving out mixed causes. Inspect the crosstabulations of perspective (actor/observer) by cause type (person/situation), first across moods, then by valence of mood. What patterns, if any, do you detect? Are there conspicuous interaction? Beyond that, do we have a case of Simpson's paradox here?
Perform a 3-way hierarchical loglinear analysis, including the factors perspective (actor vs. observer), cause type (person, situation), and mood (positive, neutral, negative). Which effects seem to be strong enough to retain? Use partial chi-squares and parameter estimates to answer this question. (Note that the parameter estimates use deviation contrasts by default, so an interpretation of contrasts will be difficult. You can focus on omnibus tests here.)
Perform model testing (using appropriate /DESIGN commands and LR chi-square change calculations) that specifically address the question whether the postulate of an actor-observer asymmetry for types of causes (the PERSPECT x CAUSE interaction) is necessary to account for the data. (a) Use a first approach that is similar to the one taken in the lecture handout: start with a minimal model (e.g., just main effects) and add terms, testing their importance by way of improvement chi-squares. (b) Use a second approach that is basically like backward elimination, in which you start with the saturated model and hierarchically remove terms (starting with the 3-way interaction) until you can't remove any further terms without letting the model fit deteriorate significantly. Justify your final models in both approaches by referencing the appropriate LR chi-square numbers.
Now turn to file 2 and the test of H2. The factors are now perspective (actor vs. observer), mood (negative, neutral, positive), and specific P cause (behavioral, internal, trait). Keep in mind that the first two (behavioral, internal) are unstable, and the third (trait) is stable. Once again, begin with inspecting the crosstabulations (across and within mood) and note what the effects might be. Any reason to worry about Simpson's paradox?
Perform a 3-way loglinear analysis including perspective, cause, and mood. Form contrasts that specifically test whether observers use more stable P causes (vs. other, unstable P causes) than actors do. Note that you must use SPSS's LOGLINEAR procedure (not HILOG) to have control over contrasts. LOGLINEAR does not force you to use hierarchical models, but you might still want to follow a hierarchical approach (if there is an AxB interaction, also consider the main effects A and B). No model testing is needed here.
Besides the a priori hypothesis H2, suggested by the literature, are there other patterns that one should take into account when drawing conclusions about the data in file 2? Use parameter estimate contrasts to answer this question.
Write a brief summary of your results, focusing on the tests of the two actor-observer hypotheses.