1/28
Looks like no tags are added yet.
1: understanding moderation
what is moderation?
variable that changes the strength or direction of the predictor-to-outcome association
association between a predictor + an outcome depends on another variable
simple terms: it depends
what is a moderator variable?
moderator is also a predictor variable
however, it is a predictor that affects the IV-DV association
acts like a dial that can be tuned to alter the effect of the predictor on the outcome
what is an example of moderation?
hours spent studying (IV) may be associated with exam scores (DV), but the strength of this association depends on phone use (moderator)
when phone use is low, studying may have a stronger positive association with exam scores
when phone use is high, studying may have a weaker association with exam scores
what is a main effect?
effect or association of one independent variable (predictor) on a dependent variable (outcome)
describes direct relationship between an IV + DV
example: more hours spent studying are associated with higher exam scores
what is an interaction?
occurs when the effect of one IV (IV1) on the DV depends on the value of another IV (IV2 or moderator)
the effect of one IV is not consistent across different levels of the moderator
eg association between hours spent studying + exam scores depends on phone use
what is the difference between a main effect + an interaction?
main effect: direct association between a predictor + an outcome
interaction: association between a predictor + an outcome changes depending on the level of another variable (moderator)
example:
main effect: more studying is associated with higher exam scores
interaction: more studying is associated with higher exam scores to a greater extent when phone use is low
what is an interaction term?
mathematical product of the predictor + moderator (predictor × moderator)
used to test whether slope of the relationship between IV + DV changes depending on the moderator
if the interaction term is statistically significant, this indicates that the relationship between the IV + DV changes across levels of the moderator
how is an interaction presented mathematically + graphically?
mathematically: product term (predictor × moderator)
graphically: slopes of regression lines differ across different moderator values
2: understanding interaction graphs
what do parallel lines suggest in an interaction graph?
parallel lines suggest there is no interaction
effect of the predictor on the outcome remains consistent across moderator levels
there may still be main effects of both variables
example:
more studying is associated with higher exam scores for both light + heavy phone users
light phone users may have higher scores overall
however, the association between studying + exam scores is the same for both groups
what do divergent lines suggest in an interaction graph?
divergent lines suggest a spreading (also called quantitative or ordinal interaction)
lines do not cross, but strength of relationship between predictor + outcome differs across levels of the moderator
example: hours spent studying are positively associated with exam scores for both light + heavy phone users, but the relationship is stronger for light phone users
what is a crossover (qualitative/ disordinal) interaction?
occurs when regression lines cross
direction of effect reverses depending on moderator
example:
more studying is positively associated with exam scores for light phone users
more studying is negatively associated with exam scores for heavy phone users
what is a spreading (quantitative/ ordinal) interaction?
occurs when regression lines diverge (spread apart)
lines do not cross
effect of one IV changes in magnitude across different levels of the moderator, but does not reverse direction
example:
more studying improves exam scores for both light and heavy phone users
however, the effect is stronger for light phone users
3: testing moderation using interaction terms
how is moderation tested in regression?
using interaction terms
interaction term is the product of the IV and moderator
if the interaction term is significant, this indicates that the relationship between the IV + DV changes across levels of the moderator
what does a significant interaction term tell us?
effect of the predictor on the outcome depends on moderator
strength or direction of the IV-DV association changes across different moderator values
example: if phone use is high, association between hours spent studying + exam scores may be weaker or even change direction
what does a non-significant interaction term tell us?
there is insufficient evidence that the relationship between the predictor + outcome changes depending on the moderator
the interaction does not explain a statistically significant additional contribution to the model
4: centering variables
what is centering in moderation analysis?
before the interaction term we should ‘centre’ our variables
involves subtracting the mean from every observation of the predictor + moderator
both the predictor + moderator should be centred before creating the interaction term
example:
find the mean of phone use
subtract mean from each participant's phone use score
repeat for hours spent studying
why do we centre variables before creating an interaction term?
1) Multicollinearity:
interaction term is mathematically related to the main effects
without centering, predictors + interaction terms can be highly correlated, this can lead to unstable estimates
2) Interpretability:
without centering: the b value represents effect when the moderator is at 0 (which is often meaningless)
with centering: the b value = effect when the moderator is at the mean (interpretable)
important notes:
centering does not affect the significance of the interaction
only changes how the coefficients are interpreted
what is multicollinearity + why is it relevant to moderation analysis?
occurs when predictor variables are highly correlated with one another
interaction term is mathematically related to the main effects
high correlations can lead to unstable regression estimates
centering helps reduce this issue
5: using hierarchal regression to assess moderation
why is hierarchical regression useful for assessing moderation?
allows us to clearly separate main effects + interactions into separate ‘blocks’ or ‘steps’
we put the interaction term into the final ‘step’ as this allows us to see how much additional variance it explains beyond main effects
what are blocks?
groups of predictors entered into a hierarchical regression in separate steps to assess how much additional variance they explain in the outcome
interaction term is entered in the final block to determine whether it explains additional variance beyond the main effects.
what are the assumptions of multiple/ hierarchal/ moderation regression?
normally distributed (ish) residuals for continuous outcome
independent data
interval/ratio predictors
nominal predictors with two categories (dichotomous)
no multicollinearity
careful of influencing cases
linearity
homogeneity of variance (heteroscedasticity)
what is a simple slope analysis?
so the interaction is significant
that tells us that the effect of the predictor (IV) on the outcome (DV) depends on the moderator
but it does not tell us what type of interaction we have
we need to do some further analysis
first thing we can do is a simple slope analysis
this assesses whether the association is significant at different levels of the moderator
what these ‘levels’ are can vary, one common way is to assess the ‘slopes’ when the moderator is at -1 SD, the mean and +1SD (low, medium + high)
summary of moderation analysis
moderation is when the effect of a predictor (IV) on the outcome (DV) depends on the value of another predictor (moderator)
main effect is the direct association of a predictor on an outcome
an interaction is when the effect of one variable on another depends on the level of a third variable
to conduct a moderation analysis, we should centre the variables to aid interpretability + reduce multicollinearity
an interaction term is used which is the product of the predictor + moderator
we can use hierarchical regression to assess whether the interaction term explains unique variance
significant interaction can be probed further using simple slopes