Wk 8 - Mediation

Question 1

What is mediation used for? (x2)

Answer 1

To characterise indirect effect of predictor on outcome, via 3rd variable
To explore/infer causal factors underlying an effect

Question 2

Describe the case of Pellagra in US, early 1900s (x4)

Answer 2

Sores, lethargy, vomiting, diarrhea killed 11Ks/yr
Common in prisoners/mental patients, but not guards/nurses
Linked to poverty/poor sewerage/germ, but mediated by diet
*Could be cured with balanced food/higher protein

Question 3

Describe Baron & Kenny’s Causal Steps approach to mediation (x4)

Answer 3

Path A: effect of Predictor on Mediator
Path B: Mediator on Outcome
Path C is Total effect of Predictor/all Mediators on Outcome
Path C’: Is Path C removed/reduced when AxB is accounted for?

Question 4

What 4 key regression analyses are involved in the Causal Steps approach to mediation?

Answer 4

Path C: Outcome = b times Predictor + c
Path A: Mediator = b times Predictor + c
Path B: Outcome = b times Mediator + c

Path C’: Outcome =b1 x Predictor + b2 times Mediator + c

Question 5

How do we interpret Path C’ in Causal Steps approach to mediation? (x2)

Answer 5

If significantly difference form 0, argue partial mediation

If not, argue for full

Question 6

What 2 ways can the strength of the indirect effect be calculated in Causal Steps approach to mediation?

Answer 6

Path A x Path B

Path C - Path C’

Question 7

How is the indirect effect tested for significance in Causal Steps approach tomediation?

Answer 7

Sobel test:
Indirect Effect is divided by SE to give z-score
Which is compared to z-distribution (+/- 1.96)

Question 8

What are 3 limitations of the Causal Steps approach to mediation?

Answer 8

Low power
Traditionally requires significant Path C
Sobel test assumes normal distribution of mediation effects

Question 9

Explain how low power is a limitation of Causal steps approach to mediation (x2)

Answer 9

Huge samples/large effects needed to detect mediation

So only really practical when testing large effects

Question 10

Explain how the requirement of a significant Path C limits the Causal Steps approach to mediation (x3)

But it should be remembered that… (x1)

Answer 10

Suppression effect if Path A and B are positive/negative
(Or multiple indirect effects)
‘Cancel out’ Path C when added to C’

Logically, Path C doesn’t need to significant for an indirect pathway to exist

Question 11

Explain how the Sobel test assumption of normality limits the Causal steps approach to mediation (x3)

Answer 11

z-test assumes normal sampling distribution of indirect effects
But smaller samples often notably skewed/kurtotic
Reducing diagnostic power of test

Question 12

What statistical technique addresses the limitations of the Causal steps approach to mediation? (x1)

Answer 12

Bootstrapping

Question 13

Define bootstrapping (x2)

Answer 13

Resampling technique used to estimate the variability of a sample statistic
Mean, standard deviation, etc. …

Question 14

How does bootstrapping address the limitations of the Causal steps approach to mediation? (x2, x2)

Answer 14

Minimal assumptions
*That sample data represent the population
Recovers important properties of sampling distribution of a statistic
*ie, shape/spread of distribution

Question 15

How is bootstrapping relevant to mediation analysis? (x2, x3)

Answer 15

Estimates sampling distribution of indirect effect
Allowing inference about magnitude/significance of it

Data is resampled eg 5000 times
Summary stats (indirect effect)/mediation model computed for each resample
Giving estimate of variability of indirect effect

Question 16

How does bootsrapping resample data? (x3)

Answer 16

With replacement:
Randomly selected individual’s data copied to resample
And then they’re put back in the pool

Question 17

What use to us is the estimates sampling distribution of the indirect effect in bootstrapped mediation? (x2)

Which we interpret by asking… (x2)

Answer 17

Can compute 95% CI
Of plausible estimates of indirect effect

Does this include zero?
*If so, no significant mediation

Question 18

What are 3 main benefits of bootstrapping for mediation? (x2, x2, x1)

Answer 18

Higher power
*Practical sample sizes for small/medium effect sizes

Gives estimate of variability of indirect approach
*Rather than ‘point estimate’ of Causal steps

Doesn’t erroneously assume normal sampling distribution of indirect effect

Question 19

As SPSS doesn’t provide bootstrapped mediation function, we… (x1)

Answer 19

Use Hayes’ Process macro

Question 20

What kind of variables can be used in the Process macro?

Answer 20

Predictors and outcomes can be continuous or categorical

Mediators must be continuous

Question 21

What output does Process give for mediation? (x3)

Answer 21

Results for each regression in Causal steps method

Bootrapped indirect effect results

Sobel test for indirect effects

Question 22

What Causal steps regression mediation results are produced by Process? (x2)

Answer 22

• b and t-test for each path

* Fit stats (R-square and F-test for each regression model)

Question 23

What bootstrapping results does Process output for mediation? (x3)

Answer 23

• Original sample indirect effect
  
  • CIs for indirect effects in model
  • CIs for contrasts of multiple effects

Question 24

Why might we want to consider multiple indirect effects? (x5)

Answer 24

Maybe theoretically interesting/important
Look at total of combined effects
Examine impact of individual indirect effects
Compare magnitude of indirect effects
Effects of 1 mediator may change with presence of another

Question 25

What do the unidirectional arrows in mediation diagrams represent? (x2)

Answer 25

Correlations!

Not necessarily causal - need multiple studies to infer that

Question 26

3 caveats on mediation?

Answer 26

Always possible for another mediator to exist
Good model doesn’t rule out alternative accounts
Implications ltd by choice of predictors/mediators - don’t overstate!

Question 27

What results must you report for mediation?

Answer 27

All relevant path coefficients, per standard regression
CIs for any bootstrapped indirect effects
If multiple mediators:
*Total indirect effect
*Contrasts of indirect effects