I would like to investigate, whether a contextual/compositional effect is mediated. I know the paper by Preacher, Zyphur & Zhang (2010) and I tried to adapt their syntax to my data/question. But unfortunately my data bear some difficulties (dichotomous DV, missing data, weights etc.) so I had to give up the plan of a (latent) path model. Now I would like to follow the traditional approach with separate regressions and I would like to use the Mplus approach which decomposes covariates in two latent parts, a within and a between part. This should be analogous to group mean centering of manifest variables. So the compositional effect bc is not calculated by Mplus but should be the difference of bb and bw.

(1) How do I calculate the indirect effect in this case and why? I see two options:

(a) Do I first have to calculate the compositional effects for X-M (coefficient a) and M-Y (coefficient b) and then their product?

(b) Or do I first have to calculate the product of a and b on both levels and then the difference of the products?

Are there differences in meaning of the two options?

(2) How do I calculate the standard errors for the “right” option?

Thank you for any hints or references

Katrin]]>

I have water data for two stations. Station A and Station B (data known as Q), with the independent variable being Date (in Months).

I need to run a monthly regression model, Station A vs Station

B. What is the monthly Q for B given the flow from A? Then, calculate

the standard error of the estimate and see which months regress the

most--to see if there is a seasonal pattern.

Please help!]]>

I have a question in regard to a model I’m trying to run.

It’s a multilevel model where the DV is a categorical variable (4 different categories). It is a very simple model with only one level 2 predictor (no level 1 predictor) at the moment. I constantly get following message in HLM:

“There is a problem in the fixed portion of the model. A near singularity is likely. Possible sources are a collinearity or multicollinearity among the predictors. We suggest that you examine a correlation matrix among the fixed effect predictors.”

Does anyone have an idea why I get this message (multicollinearity can’t be the problem since I have only one predictor)? Oddly enough when I specify the model as multinomial (again with 4 categories), then it works and the programme gives me the estimates...

Any thoughts or ideas on that would be very much appreciated!

Thanks.]]>

I am conducting a study which is examining data from 354 hospitals. It is working with hospital level and geographic data.

However, in order to test for simpson's paradox I need to analyze the effect that the individual level data has on the hospital level.

I was going to use a MLR for the hospital study, but I am unsure what type of analysis I could use to combine the hospital level as well as the individual level data. All variables, but one are continuous.]]>

I am struggling with the understanding of R

Please correct me if I am wrong: Concerning R/B I would formulate that this approach decmposes the variance of Y into a variance within and a variance between the classes. R

Hox (2002, p. 70) says that the interpretation of a S/B R

My questions are:

(1) Following Hox I would formulate that R

(2) How would you describe the meaning of R

I would be grateful for any information you are able to provide me with! Thank you in advance for any efforts!

Katrin]]>

I would like to compare the effect of different covariates in a multi level regression. Especially I am interested in comparing the effects of individual level variables with the effect of their respective group mean or with the effect of other (non aggregated) level 2 characteristics.

I read about the effect size measure suggested by Tymms (e.g. in Schagen & Elliot, 2004). As far as I understand the chapter this measure only quantifies the effect of a L2 variable. So I wonder how I would have to calculate a respective measure for the individual level variable.

In a multiple regression I could at least compare the standardized coefficients (leaving aside the discussion about the comparability of betas ...). Analogously I could take the standardized coefficients that are offered for example in the TWOLEVEL analysis by Mplus. But is this also a suitable practice for cross level comparisons in multi level analysis?

Are there any ideas or references that explain how I can quantify the effects of variables of different levels and how I can assure comparability of these measures?

Thank you for any further hints

Katrin]]>

I try to specify a multilevel model and face the fowllowing questions:

1) Centering

I am interested in testing whether the level 1 coefficients vary over my clusters. From Raudenbush & Bryk (2002, Ch. 5) I know that I should use group mean centering for this purpose. Now I ask myself whether I should also add the group mean of the respective variables on Level 2. In Ch. 5.2 Kreft and de Leeuw (2004) show that the use of the group mean may substantially changes the results - whereas this seems to apply more to the fixed level 1 coefficients. Are there any (statistical) guidelines when to add the respective group mean? Or is this just a matter of theory resp. of the state of research on the topic?

2) Reliability of random level-1 coefficients

In the HLM output I saw that HLM calculates reliability estimates of the random level-1 coefficients. I have two questions regarding these estimates:

a) Raudenbush & Bryk (2002, p. 125) say that "whenever the reliability of a random level-1 coefficient drops below 0.05, that coefficent is a candidate for treatment either as fixed or nonrandomly varying". So I conclude that this is a second kind of significance test for the variance components. But which criterion is the more important one? In my case the chi-square test and the reliability estimate would partly lead to different conclusions.

b) All reliability estimates in my model are very low - especially those of the random slopes (.20 or less). So I asked myself whether I have to deal with these estimates like with ordinary reliabilities of a scale for example - especially after reading about the above mentioned (very low) criterion value. Could you give me any advice how to inetrpret reliabilities like these?

I would really appreciate any comments and ideas!

Katrin]]>

my question is how to model a certain multilevel data structure.

Specifically, I have survey data with persons (i) nested in regions (j), and these surveys have been collecetd over several time points (k).

So for persons, it is a repeated cross-section, not a panel study. Person-level predictors for the dependent variable aside, in my model

- a first predictor (Z1) varies only across time points (k).

- another predictor (W1) refers to the regions (j). W1, however, takes on different values for each region j at each year k.

This means that W1 varies across both time points (k) and regions (j).

Initially, I thought it would be adequate to modelt his data structure as a cross-classified model. This was motivated by the ideat that persons living in any of the j regions could have been interviewed at a certain time point k, while respondents interviewed at a certain time point k could also have lived in any of the j regions.

So I used cells (= cross-classifications) of regions j (= rows) and time points k (= columns) at the highest level of analysis.

But there seems to be a problem involved. Because W1 takes on different values across both time points and regions, doesn’t this mean that, conceptually, W1 neither represents a unique row respectively column variable? Or is it possible that a predictor variable in a cross-classified model can account simultaneously for row AND column variance???

Many thanks for your comments!]]>

Now I ask myself: a) How does it happen that the variance of a cluster sample is larger than the variance of a random sample of equal size - I had assumed that the higher similarity of the cluster members not only leads to smaller intra-group variances but also to a smaller total variance. b) Why are the standard errors underestimated - if not due to a smaller variance of the cluster sample?

Do you have a recommendation for a reference that gives an explanation? - I know that, concerning sampling questions as these, I should probably consult classics like Kish or Cochran. But unfortunateley my statistical education is not so profound that I could understand them. So I would be very glad to begin with some rather introductory references.

Thank you in advance for any comments! Katrin]]>

currently I am working on multilevel data (students within classes). I would like to specify a 2-level-model and I assume that certain class characteristics (e.g. class size etc.) will influence the betas of level 1. But I don't want to explicitly specify this within the model. I would rather like to find out whether there are "natural" groups of classes and explore afterwards by which class characteristics they differ.

Unfortunately I am a real beginner in multilevel modeling. So, I would be glad for some feedback: Is this - for whatever reasons - a stupid idea? And if not: Could please someone give me advices where to read more about this and how to do this in Mplus?

Thanks in advance for any comments and advices!

Katrin]]>

For my master thesis i wanted to investigate the effect of different reward schemes on cooperation and knowledge sharing and whether these effects were moderated by individual characteristics such as: risk aversion, overconfidence and social value orienation. I investigated my research topic by means of a factorial survey (vignette study). In order to analyse my results i want to run a multilevel mixed effects analysis in SPSS via the mixed-model procedure.

However I am not sure whether to use the random or repeated function, and whether i need to check for random effects on the level 2 (individual characteristics).

Could someone please help me out?

Thanks alot in advance!

Nicole]]>