Hi all,
I have a manipulated IV (2 store types), with a mediator (4 items on Likert scale) and DV (3 items on Likert scale).
I also have a moderator within those manipulations (every respondent gets to see a leading brand vs non-leading brand).
Which test is most suitable for me?
Thanks in advance all!
Kind regards,
Maayke
Hello Maayke,
It sounds to me as if the first step would be to determine how to treat the scores from the four items associated with the mediator variable and the scores from the three items representing the dependent variable. Assuming all items are posed in the same direction/polarity, the classic method for Likert-type scales, is to use summated scores (e.g., mediator item 1 + mediator item 2 + ...) (see the monograph by Likert himself: https://legacy.voteview.com/pdf/Likert_1932.pdf).
However, before embracing summated scores as the way to develop a composite score for mediator and one for DV, I'd suggest: (a) looking at internal consistency reliability estimates for each item set (via Cronbach's alpha)--these should be high (far closer to 1 than to 0); and/or (preferably) (b) factoring the two item sets to assure that you have two distinct, and coherent (latent) variables being tapped by the seven items; and/or (c) using item response theory (perhaps the graded response model) to evaluate the coherence of the two item sets as scales. Any of these methods could yield empirically defensible composite scores that could be used for analysis.
The analysis could be undertaken via regression or SEM (structural equation model) software. Classic mediation implies an indirect effect of store on Y via the mediator (product of two paths: mediator regressed on store type; Y regressed on mediator) relative to the magnitude of the direct effect of store on Y.
Good luck with your work.
Hi Maayke,
beyond what David said, the following is hopefully helpful
1) If M is measured, the relationship between M and Y can be biased by unmeasured common causes, a reverse effect of Y on M or simultaneity. Hence, even if your treatments are randomized, the indirect effect will be biased.
Hence, you should avoid OLS regression (e.g., as implemented in PROCESS). This holds for any mediation analysis. Instead use SEM. If your 4 measures of M are reflective indicators of M, you could specify a latent variable measurement model and test it (!). If not, adding together the items is surely a practical way (I do that by myself) but leaves it unclear, how many real attribute are "in there" and what exactly the causal mechanism is.
2) If you test for a full mediation model (assumed a sufficiently large N and strong treatment effects), a non-significant chisquare test provides some evidence whether the M - Y effect is confounded or some reverse effects are present. If the chisquare test is significant, having more then one treatment variable (=instruments) should enable you to find out whether the problem is due to the aforementioned problem or one or more of the treatments having a direct effect. All the best, --Holger
David Morse and Holger Steinmetz Thank you so much for your help. It is more clear now!
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Graphs