mediation Flashcards
what is mediation?
- a mediator links w variables
- a mediator is a variable that is affected by the IV
- the mediator also effects the DV
- the IV-DV effect is at least partially dependent on the mediator
- another version of regression analysis
- looking at whether there is an indirect effect of the Iv on the DV through a mediator (3rd variable)
- full mediation- when the effect of the IV on the DV is fully accounted for through the mediator (all variance explained by the mediator)
mediation vs moderation
- A mediator is a variable that accounts for an association between a predictor and a DV
- A moderator affects the strength of a relationship between a predictor and a DV.
- A mediator is a variable the accounts for an association between a predictor and a DV (i.e. explains why X is associated with Y),
- A moderator says in what circumstances associations between X and Y will be apparent
moderation
- A moderator does not have to be associated with the IV or the DV (but can be)
A mediator MUST be associated with the IV AND the DV- We may predict the association between self-reported food palatability and calories consumed would be moderated by diet intention
- i.e.,
○ Participants who are dieting may not show an association between palatability and calories eaten (they may think the food is palatable but resist it)
○ Those not dieting will show an association between palatability and calories eaten - Notably dieting cannot be influenced by the palatability of a food in an experiment (time-order relationship)
mediation- suppression effects
A direct IV-DV association may be in the opposite direction to the IV-M-DV association
- A study looks at the effect of anxiety on drinking
- The study finds anxiety is associated with decreased drinking, i.e. There is a negative correlation between the IV-DV
- May be because they are too anxious to socialise
- However, if we measure drinking motives (reasons people drink) we can get a score on whether people drink to cope with anxiety
- Anxious people generally drink less, as there may be complications with their medication and they may be less socially active due to their disorder
- However, there may be a portion of the sample who drink to cope with anxiety; we will see anxiety positively associated with drinking to cope with anxiety, and a positive association between drinking to cope with anxiety and alcohol consumption
- Because there is a negative direct effect and a positive indirect effect (mediated) the total effect (both pathways) comes at as 0 because they cancel each other out!
- When we found negative association, if we include the mediator the association can flip around the other way i.e. turn into a positive association
mediation- causal steps
- Another key problem with many mediation analyses in the literature is that they are grossly underpowered
- If you expect a medium strength IV-M and medium strength M-DV association and you were to use the causal steps approach you would need 397 participants (if small effects are expected you would need 21,000 participants!!!!)
- By contrast the method we look at today would need 74 participants
- Causal steps: an approach that
- has little or no sensitivity at all (needs huge sample)
- Is mathematically incorrect (IV DV association not necessary)
Is unable to detect suppression effects
bad solutions (Sobel Z test)
- Gives a p value for the indirect effect
- Based upon a product of the coefficient’s calculation (i.e. multiplies the regression coefficients together and tests for the significance of the outcome)
- Assumes the product of the coefficient is normally distributed – this is almost never the case
- This method also requires a more participants that to detect indirect effects than the methods described today
joint significance test
- This method simply ignores the IV-DV association (i.e. it doesn’t have to be significant!)
- If the a-path (IV-M) significant and the b-path (M-DV) is significant then there is evidence of mediation.
- Essentially run a regression analysis for A and B path- of these are significant it suggests there is a mediation going on
- If one of these paths are not significant there is no mediation.
a note on causality
- Mediation analyses do not show causal relationships!
- They test proposed associations between variables
bootstrapping
- Way to get more stringent confidence intervals and p values using your data
- Takes samples from dataset and then calculates the statistic in question e.g. standard error- in this case it is confidence intervals and p values
- It then takes another sample and do this again- can also draw from the first sample again- same value could be taken more than once
- R does this 1000 times and then calculates an average p value and confidence intervals- more accurate estimate as it tried to correct any biases
- If you run this more than once they will change every time because the bootstrapped estimate will differ
total effect (T)
- We get a value called ‘Total Effect’
- This is the overall effect of the IV on the DV it takes into account the direct effect (the association between the IV and DV not taking anything else into account) and the indirect effect (the effect of the IV on the DV through a mediator).
Total Effect = Direct effect + Indirect effect.
average causal mediation effect (ACME)
This measures how much of the effect of the IV on the DV happens through the mediator.
average direct effect (ADE)
- This is the direct effect of the IV on the DV controlling for the mediator.
- If the mediator didn’t exist, what is the effect of the IV on the DV
- Is there an effect of the IV on the DV that occurs even if we ignore the mediator?
- If this is the case, then we would have a partial mediation rather than a full mediation.
proportion mediated (PM)
- Tells us what fraction of the total effect of an IV on the DV is explained by the mediator
- Proportion mediated= Indirect effect (ACME)/Total effect (T)
- Normally express this as a percentage = 0.4 is 40% of the effect is mediated
- If you go over 1, this is a suppression effect and it has flipped around the other way because of the mediator
summary
- Mediation analyses are poorly understood both in terms of-
- Conceptualisation (e.g. confusing it with moderation)
- Analysing (using mathematically incorrect methods)
- Simple mediation can be tested with the joint significance test
- More advanced mediation models are becoming more common in the literature
what to do in R
- load in data
- A path
- B path
- Bootstrapping
