Discover the world's scientific knowledge

Hello,

assume I have data from two (and later three) time points that are in a three-level structure. I want to test mediation. The predictor is binary, the mediator is (usually) continuous (in some analyses also binary), the outcome is continuous (and in some analyses binary).

So, it's

- mutilevel

- longitundinal

- mediation

analyses.

Can you recommend how to analyse this data? I was thinking about latent growth modeling with mediation but I am not sure if this is either possible or the best.

Thanks!

I am using R to run multilevel modelling. In my study, each participant completed a survey every day for 5 days. I have set up the data in Excel such that each participant has a row for each survey they completed (representing each day). The columns are sorted by participant ID (i.e., a participant's surveys across the 5 days are lumped together) and then by day (i.e., participant A's first row is from Day 1, second row from Day 2, etc). When I import the file into R, the program no longer maintains the sorting that I used (i.e., Days 1 through 5 occur in random order). Is this problematic when running my analyses? I would think that I need to maintain the temporal order when analyzing my data, but I am not sure how to do so.

Dear all,

I have a burning question regarding mediation analysis for multilevel repeated measures designs. I have a design in which groups of three participants (multilevel) discussed one time FtF and one time via text-based online chat (repeated measures). I performed a content coding on all discussions and let participants fill out a questionnaire after each discussion. Now, I would like to run a mediation analysis where the content coding mediates the effect of condition on the questionnaire results. However, I am really struggling to find out how this works as the condition and mediator are on the group level whereas questionnaire outcomes are on the participant level. Moreover, I am wondering how can I account for the repeated measures effect (all participants and thus groups participated in all conditions). I am using lme4 in R for these analyses btw. I hope someone can give me advise.

Thank you.

Hi. I am using the European Values Survey (3rd wave and 4th wave) and run multilevel models separately with EVS3 and EVS4 (two separate analyses). I chose not to pool the two waves because most previous studies related to my topic don't pool the different waves and just use only one wave or conduct separate analysis for each wave. What is the best practice for pooling or not pooling the different rounds of mass opinion surveys? What is the basis of your judgement?

I have a data set which includes 200 individual's responses in 100 companies. I have collected 2 responses per organization.

With the data set, I have analyzed the data based on multilevel SEM.

multilevel.model <- '

level: within

Aw =~ II1 + II2 + II3 + II4 + II5 + II6 + II7 + II8 + II9

Bw =~ CET1 + CET2 + CET3 + CET4 + CET5

Cw =~ CER1 + CER2 + CER3 + CER4 + CER5

Dw =~ POC2 + POC3 + POC5 + POC6

Ew =~ POJ1 + POJ3 + POJ4 + POJ6

Bw ~ Aw + Dw + Ew + Gender + Age

Cw ~ Aw + Dw + Ew + Gender + Age

level: between

F =~ IIP2 + IIP3

G =~ IIP2 + RIP3

Db =~ POC2 + POC3 + POC5 + POC6

Eb =~ POJ1 + POJ3 + POJ4 + POJ6

Bb =~ CET1 + CET2 + CET3 + CET4 + CET5

Cb =~ CER1 + CER2 + CER3 + CER4 + CER5

Bb ~ Db + Eb

Cb ~ Db + Db

F ~ Bb + Cb + ii1*Indust_1 + ii2*Indust_2 + ii3*Indust_3 + ii4*Indust_4 + ii5*Indust_5

G ~ Bb + Cb + ir1*Indust_1 + ir2*Indust_2 + ir3*Indust_3 + ir4*Indust_4 + ir5*Indust_5

'

I have experienced a convergence problem.

Do you think 2 response per firm may not enough to do multilevel SEM? Or, do you think I should look for other reasons for the problem? Any suggestions?

Thanks in advance.

I've conducted an RCT in which I'm testing the effect of a group mindfulness intervention on depressive symptoms. Only one group was running at a time so there were four study waves, with each wave of participants being randomized to intervention or control. Outcomes were measured bi-weekly for 6 months. I'm testing the effect of intervention using PROC MIXED with bi-weekly assessments nested within participant identified in the repeated statement.

A reviewer has suggested that I include treatment wave as a random factor. However, the interaction between treatment and study wave (as fixed effects) is not even close to significant (p = .99), suggesting that the effect of treatment is the same across waves. Is this sufficient justification to keep my analyses as they are? Thanks!

Hello,

I have a dataset with a three-level structure. Participants reported an outcome of interest for 3-4 aspects of an incident, and they reported 4 types of incidents (1 incident for each level of 2 * 2 factors). I draw the structure in the figure below.

I have two questions:

(1) How should I structure the covariance matrix for residuals? I expect that the residuals for each person to correlate. Also, I expect that the residuals within each event correlate stronger. I prefer not to use an unstructured matrix, as the number of parameters would be too large.

In SPSS, I'm using this syntax:

MIXED Y BY Valence Type Aspect

/FIXED = Valence Type Valence*Type | SSTYPE(3)

/REPEATED=Type*Valence*Aspect | SUBJECT(participant_ID*event_ID) COVTYPE( )

(2) As is illustrated in the figure below, one aspect of the outcome is not applicable to the type 2 events (type 2 events only have 3 level-1 categories). I assume that when I want to assess the effect of incident type, I should exclude the number 4 aspect that is only applicable to type 1 incidents so that the results would reflect the difference between types including same level-1 outcomes. But, excluding number 4 or including it results in a negligible difference. Can I just report that, and then include the number 4 in subsequent analyses? Especially as excluding data from the number 4 will cost substantial statistical power.

Thank you in advance.

Dear all,

I have collected some complex data and I have some questions on the best way to analyze them in order to answer my research questions.

I have a sample of 165 preadolescents. They all come from the same school and are divided in 19 classes. Each kid completed 5 scales of a questionnaire; for each kid a parent (usually the mother) has completed the same 5 scales.

My hypotheses on the relationships among variables regard: H1) a relationships among constructs in the group of preadolescent (for example: in the preadolescents group variable A is correlated to variable B, variable C is correlated with variable D) H2) a relationship among constructs in the group of parents (same as in H1) H3) a relationship between constructs measured for kids and for parents (for example: parent-child correlation on variable A, but also parent variable A correlates with child variable B) H4) homogeneity within classes (kids in the same class are more similar than kids in different classes)

My hypotheses are correlational, I would prefer to avoid directional hypotheses (for example "parent variable A predicts child variable A"), so I'd rather avoid multilevel regression analyses.

Can you think of any way to treat this data, accounting for class? Is there any kind of multilevel correlational analysis? If not, how would you treat this?

Thank you for your attention and for any suggestion you can think of!!

I am doing a study about an intervention targeting pupils of schools, in different cities. Thus, it is a multilevel structure and I am using mixed models for regression analyses. Contextual factors, related with cities, could be explanatory variables for the outcome. In this manner, the variable “city” is a proxy of such contextual factors. But, the variable “city” is also a clustering variable of the multilevel model.

Now, the question:

Is it appropriate to include the clustering variable “city” as confounder variable in the multilevel mixed models when “city” is also included as a hierarchical cluster variable?

Thank you very much.

Andrés.

In my current project, I want to answer if various cognition items (ratio, 30+ of them, may get reduced based on a separate factor analysis) predict moral outrage - in other words, do increases in item 1 (then item 2, item 3, etc) predict increases in outrage in a significant way. Normally, this would be a simple regression. But then I complicated my design, and I'm having a hard time wrapping my head around my potential analyses and whether it will actually answer my stated question, or if I'm over-thinking things.

Currently, I'm considering a set-up where participants will see a random selection of 3 vignettes (out of 5 options) and answer the cognition items and moral outrage about each. This complicates matters because 1) there is now a repeated measure component that may (or may not?) need to be accounted for and 2)I'm not sure how my analyses would work if the vignette selection is random (thus, all vignettes will show up the same number of times, but in different combinations to different people). I am anticipating that different vignettes will not be equal in their level of DV (which is on purpose - I want to see if these patterns are general, not just at very high or very low levels of outrage).

When originally designing this, I had wanted to average the 3 vignette scores together for each subject, treating them as single, averaged item values to use in a multiple regression. But I've been advised by a couple people that this isn't an option, because the variance between the vignettes needs to be accounted for (and the vignettes can't be shown to be equivalent, and thus can't be collapsed down in analysis).

One potential analysis to combat this is a nested, vignette-within-individual multilevel design, where I see if the pattern of cognition items to outrage is consistent between vignettes (level 1) and across subjects (level 2), to account for/examine any vignette-by-cognition/MO pattern interactions. And this makes sense, as MLMs can be used to compare patterns, rather than single scores.

But I can't wrap my head around what part of this set-up/the output I would look at to actually answer my question: generally, which, if any, of these cognition items predicts outrage (regardless of vignette, or across many scenarios)? And can this approach work when the vignettes combinations differ between subjects?

Or is this the incorrect analysis approach and another, simpler one would be more fitting? For example, is the averaging approach workable in another format? What if all vignettes were done by all subjects (more arduous on the subjects, but possible if the strength of the analysis/results would be compromised/overly-complicated)?

Confirmation that my current analysis approach will indeed work, help with what part of the output would answer my actual RQ, or suggestions for an alternative approach, would be appreciated.