Lec. 25 Mediation Flashcards
(42 cards)
1
Q
What is an interaction effect?
(review from past exam)
A
The effect of one variable is influenced by another variable
2
Q
Mediation analysis - what is it?
A
- looking at association between one predictor variable and an outcome variable
- trying to divide total effect in a direct effect and an indirect effect
> “what proportion of the total effect is due to indirect effect and direct effect?”
→ based on the answer, we decide whether to account for mediator
→ mediation occurs if direct relationship between the predictor and outcome is reduced by including the mediator
→ perfect mediation occurs when “c” is zero
3
Q
what do the three linear models in a mediation analysis predict?
A
- model 1: predictor on dependent variable
- model 2: predictor on mediator
- model 3: predictor + mediator on dependent variable
4
Q
what are the four conditions of mediation?
A
- the predictor variable must significantly predict the outcome variable (model 1)
- the predictor variable must significantly predict the mediator (model 2)
- the mediator must significantly predict the outcome variable (model 3)
- the predictor variable must predict the outcome variable more strongly in model 1 than in model 3
- b^ is the unit of measurement for predictions
5
Q
pornography study
X: pornography
Y: infidelity
Z: relationship commitment
- how does this example follow the four assumptions?
A
- pornography (P) significantly predicts infidelity (DV)
- pornography (P) significantly predicts relationship commitment (M)
- relationship commitment (M) significantly predicts infidelity (DV)
- relationship between pornography (P) and infidelity (DV) is stronger in model 1 than in model 3 (b^ is larger)
= when including relationship commitment, there is a reduction in relationship between pornography and infidelity
6
Q
pornography study
how do we compute the mediation values in JASP for the pornography example?
A
see image 12, and look at all the boxes that were completed and selected
7
Q
what do the outputs show?
A
- estimate → b^
- R2 → proportion of variance explained by first variable on second variable
- Std. Estimate → effect size
8
Q
pornography study
how can we interpret the output of the path coefficients in the pornography study?
A
- path coefficients:
> pornography does not predict infidelity significantly (when commitment is in the model!)
> pornography significantly predicts commitment
> commitment significantly predicts infidelity
= when looking at std. estimates, the there is a larger relationship bewteen commitment and infidelity (-0.29) compared to pornography and infidelity (0.15)
(image 13)
9
Q
pornography study
how can we interpret the output of the total effects in the pornography study?
A
- total effect: effect of predictor on outcome when the mediator is not present in the model (axb+c) (!)
> when commitment is not in the model, pornography significantly predicts infidelity (b^=0.59)
(image 14)
10
Q
pornography study
how can we interpret the outcome of direct and indirect effects in the pornography study?
A
- direct effect of pornography on infidelity in isolation (0.46) → not significant
- indirect effect of pornography on infidelity when commitment is included as predictor (0.13) → not significant
- C.I. of indirect effect include zero
= relationship between pornography and infidelity is not explained by commitment
(image 15)
11
Q
pornography study
What is the final conclusion of the pornography study?
A
- the relationship between pornography (P) and infidelity (DV) cannot be explained by commitment
(confidence intervals of indirect effect contain zero)
12
Q
what does a negative estimate indicate?
A
- as variable 1 increases, variable 2 declines (and vice versa)
13
Q
what are total, direct and indirect effects?
A
> direct effect: between independent and dependent variable (accounting for mediator)
indirect effect: between independent variable, mediator, dependent variable
total effect: direct + indirect effect
14
Q
what is the simple mediation model?
A
- only one independent variable and one mediatior
- simple mediation: simple relationship between the predictor variable and the mediator
(see image 1)
15
Q
what is a mediator?
A
- single predictor variable that is associated with the other predictor variable
- when present, it explains the effect of the independent variable on the dependent variable
16
Q
What is an example of mediation?
How is the effect explained?
A
~ “does the speed of recovery after sickness improve with the use of alternative medicine or is this effect mediated by a healthy lifestyle?”
- predictor: homeopathic remedies
- mediator: healthy lifestyle
- dependent variable: speed of healing
> homeopathic remedies is strongly associated with healthy lifestyle, and healthy lifestyle leads to faster healing
17
Q
how can we divide total effect in direct and indirect effect?
A
- through mediation paths
> total effect
> mediator effect
> combined effect
(see image 3)
(you can do your own mediation analysis in jasp by clicking button on the slide that is titled “the data”)
18
Q
Total effect
- regression equation
A
- “bt”
- predict outcome variable (speed of healing) by looking at single predictor variable (homeopathic remedies)
> one intercept, and one regression weight
(see image 2)
19
Q
Mediator effect
A
- “a” arrow
- predicts mediator based on original predictor variable
- it assesses relationship between mediator and predictor variable
> effect of predictor on mediator
20
Q
Combined effect
A
- “b+c” arrows
- predict outcome based on predictor and mediator
- two regression weights (predictor variables):
> predictor to outcome (c) (direct effect)
> mediator to outcome (b)
→ two-way regression analysis
21
Q
How can the indirect effect be calculated?
A
- it is the combination of a and b (a x b)
- regression weight “a” (ba) x regression weight “b” (bb)
> effect of homeopathic remedies on speed of healing through healthy lifestyle
(see image 4)
22
Q
Multicollinearity
- what does it mean when it’s high/low?
A
- association between predictor and another predictor or mediator
> if “a” is zero, it means that there is no multicollinearity, and therefore no association between predictor and mediator
> high multicollinearity is not really a problem, as long as we model associations
23
Q
how can we standardize the indirect effect?
A
- we divide indirect effect (a*b) by standard deviation of outcome variable and multiply it by standard deviation of predictor variable
- we can now use this standardized regression weight for hypothesis testing
> (jasp standardizes it for you)
(see image 5)
24
Q
- how can we compute the proportion of mediation?
- why should we be careful about it?
A
- indirect effect (a*b) divided by total effect
- proportion of total effect that was mediated
> mediation analysis is not really a proportion, and this proportion does not take into account the size of total effect (problematic!)
25
look at image 6:
- what does the venn diagram show?
- 80% of variance in the outcome overlaps with the variance in the predictor
→ show strong association between predictor and outcome
> but what of this variance is explained by mediator?
26
look at image 7:
- what does the Venn diagram show?
- mediation analysis
- you must look at shared variance with mediator
> 75% of variance overlaps with mediator → shows that mediator has strong impact on effect between predictor and outcome
27
How can we calculate these mediation paths on JASP?
- regression models called "Process"
> we can create a regression model per path: one for total effect, one for direct effect, one for combined effect
> we take regression coefficients and through scatterplots, we can see how strong the correlations are (multicollinearity)
28
what is important to remember when using a Process model on jasp?
these models give us correlations, not causations (!)
29
How do we use the Process model on JASP?
1. click blue plus in top right corner, and add "Process (beta)"
2. we selct "Classical Process Model" and add dependent variable and continuous predictors
3. specify model:
3.1 Hayes configuration → input "4" in "Hayes configuration number"
> we get visual model
3.2 we specify what variables fit in which part of the model
! suggestion:
a. always look at statistical models (same as conceptual model but has "a", "b", "c")
b. always use parameter estimates (it gives already regression weights
→ through these we can calculate already direct, indirect and total effect
- see image(s) 8
30
what are the direct, indirect and total effect in image 8.3.c?
- direct effect: 0.547 (c)
- indirect effect: 1.53*0.973 (a*b)
- total effect: 1.53*0.973 + 0.547 (a*b+c)
31
Regression weights vs correlations
- what is the difference?
- they are very closely related
- regression weights can be bigger than 1 and smaller than -1
- they help us to give predictions of the model
32
What do the Parameter Estimates tables in JASP show us?
- gives us the parameter estimates, standard error and z-values (that can be used for hypothesis testing)
- by looking at standard error, we now know the uncertainty with which we are estimating parameters
(see image 9)
33
How do we interpret these findings?
- total effect is strong (extreme z-value)
- direct effect is very weak (C.I. around zero)
- indirect effect is significant
→ high mediation
! p-value, z-value and confidence intervals tell us whether we can conclude that there is a mediation effect
34
Mediation vs Interaction effect
- Mediation: association between predictor variable 1 and predictor variable 2
> is value of one predictor influenced by another predictor?
- Interaction effect: effect of one variable being influenced by another variable
> is effect of one predictor influenced by another predictor?
! just because there is strong association between two predictor variables, does not mean that there is interaction effect !
35
how can we compute the interaction effect?
- we go to Linear Regression analysis and add interaction effect to model 1
(see image 10)
36
what does the interaction effect mean in this context?
- interaction effect is here very weak
- effect of homeopathic remedies depends on how healthy your lifestyle is
- "degree of one predictor variable influences effect of other predictor variable"
37
Mediation analysis - how else could we do it?
- not "Hayes configuration", but "paths"
- specify predictor, dependent variable, process ("mediator") and mediator
- same results
(see image 11)
38
Mediation with two mediators
- what does such model look like?
see image 16
39
How can we plot on jasp the mediation analysis with two mediators?
Image 17
40
How can we interpret the findings in the example study?
(image 18)
I had to include the part of text from the book, because it didn't make sense to rewrite everything in the flashcards.
41
What is the key take-away from the study with two mediators?
- the initial relationship between delusional thinking and belief in fake news (b^=0.24) is diminished to b^=0.15 when the mediators are included
- each of the three indirect pathways are significant
> suggests that part of the relationship bewteen the predictor and the DV operates through lower levels of the mediators
> ("lower" levels because the predictor is negatively associated with mediators, and mediators are negatively related to DV)
42
What is a summary of mediaton?
see image 19 for a complete graph
