Attachment | Size |
---|---|

Output_tssem2.txt | 3.43 KB |

Hello everyone,

I am currently conducting TSSEM using correlation matrices from primary mediation studies. I ran into some problems during the process and hoped that someone could help me.

1) I understand that in studies which reported standardized path coefficients, only the a-path can be treated as a correlation and that the b- and c-paths need to be computed with the impliedR()function (see also here: https://openmx.ssri.psu.edu/node/4671). For me, this is the case in one study - however, there is a missing value (a-path) for which I am not sure what to insert into the Amatrix. I tried “NA” which gives me the following error report: “Fehler in strsplit(freePara1, "*", fixed = TRUE) : nicht-character Argument”. When I use zero, the outcome matrix provides me with the same values I already have retrieved from the primary study. I could not find a guideline on how to include missing values into impliedR() and hoped someone might have an idea.

2) tssem1() returned with an OpenMx status of 5 or sometimes 6 at first. I read in another thread (https://openmx.ssri.psu.edu/node/4244) that a solution might be fixing small heterogeneity variances to 0 for tssem1(), which I tried:

RE <- Diag(c(0,0,"0.01*Tau2_3_3"))

RE

; (although all variances were very small: Tau2_1_1 and Tau2_2_2: 0.0000000001, Tau2_3_3: 0.0203500436) and after that, the OpenMx status was 0. Running tssem2() also works, however the fit indices from tssem2() do not seem to make sense (see attached output) and I do not get standard errors or p-values. Is it a possibility that I just do not have enough studies to pool (four studies investigating five mediators) and too many missings?

pattern.na(matrix.list2, show.na = TRUE, type=c("tssem"))

X M Y

X 4 4 0

M 4 0 1

Y 0 1 1

3) How exactly can I include the sample size of a study investigating more than one mediator into the TSSEM? As I need to report a sample size in the n vector for each correlation matrix, I currently split the sample size of the study with two mediators in order not to overrepresent it by putting in the whole sample size two times. Is that correct?

Any help would be greatly appreciated.

Thanks a lot in advance,

Hannah

Dear Hannah,

1) The impliedR() function is used to convert a standardized path coefficients (A matrix) into a model-implied correlation matrix. If we do not know some paths, I do not see how we can impute some values to them.

2) Since you have requested a likelihood-based CI, SE, z, and p-values are dropped to avoid confusion. If you prefer them, you may use the intervals.type = "z" argument.

More importantly, your pattern.na(matrix.list2, show.na = TRUE, type=c("tssem")) indicates that there is no study on the relationship between X and Y. I would not interpret the results.

3) I do not get it. The n refers to the sample size of a correlation matrix. The n should be the same regardless of how many mediators there are. Could you provide an example?

Best,

Mike

Dear Mr Cheung,

thank you very much for your fast reply, it helps a lot.

2) Regarding the pattern.na() matrix: I have five correlation matrices and four studies in my analysis (because two mediators come from the same study). So there is one study which reported the a-path, the remaining three (one with two mediators) did not. Would you still recommend not interpreting the results, since there are quite many missings and only five matrices, or could preliminary interpretations be made?

3) That answers my question already, thank you very much. I was under the impression that the n refers to the study sample size and was unclear about how to include multiple mediators from one study. Thank you for helping me out.

Kind Regards,

Hannah

Dear Hannah,

2) You can fit the model. However, you may have to consider whether one study can be called a meta-analysis.

Best,

Mike

Dear Mr Cheung,

thank your very much for your help and time, it is very appreciated.

I have one last question regarding the fit indices: the RMSEA for the model we talked about above is zero (see attached output from my first entry), which would indicate a perfect model fit. However, this does not really make sense considering there is only one study for the a-path and very few studies with many missings overall. I have this issue in other TSSEMs I ran as well (all with very few studies with many missings). Do you have an explanation for this? Could the fit indices be distorted due to the small number of studies and many missings?

Kind Regards,

Hannah

Please note that the df of the model is 0.

Mike