Reliability and correlation in Within-item multidimensional model

[shentong]

Hello,

I was running a within-item multidimensional model with a 23-item scale. It has two subscales and each item is loaded on both subscales in the within-item model. The variable map and item parameter estimates looked good, but the WLE Person separation RELIABILITY were shown to be 1.000 for both subscales... This looks really weird to me.  Could anyone explain what has happened?

Also, in the within-item model, the correlation between the two subscales is 0.445, which is much lower than the correlation (coefficient =0.865) in the between-item multidimensional model (first 13 items loaded in the first dimension, next 10 items loaded in the second dimension).  Why the correlation is lower in the within-item model? Any help would be greatly appreciated!

Best,

Tong

[shentong]

I have a follow-up question about this analyses. Using the same data and within-item model, I found that the person estimates (WLE) were not generated correctly. There seems to be some error and neither estimates of the two dimensions were provided. Does anyone know what has happened?


The WLE file looks something like this:"

1  1  35.00000  36.00000  35.00000  36.00000   0.00000  -1.#IND0   0.00000  -1.#IND0
2  2  34.00000  36.00000  34.00000  36.00000   0.00000  -1.#IND0   0.00000  -1.#IND0
3  3  35.00000  36.00000  35.00000  36.00000   0.00000  -1.#IND0   0.00000  -1.#IND0
4  4  33.00000  36.00000  33.00000  36.00000   0.00000  -1.#IND0   0.00000  -1.#IND0
5  5  28.00000  36.00000  28.00000  36.00000   0.00000  -1.#IND0   0.00000  -1.#IND0
6  6  34.00000  36.00000  34.00000  36.00000   0.00000  -1.#IND0   0.00000  -1.#IND0
7  7  26.00000  36.00000  26.00000  36.00000   0.00000  -1.#IND0   0.00000  -1.#IND0
8  8  20.00000  36.00000  20.00000  36.00000   0.00000  -1.#IND0   0.00000  -1.#IND0
9  9  24.00000  36.00000  24.00000  36.00000   0.00000  -1.#IND0   0.00000  -1.#IND0
...."

Thanks a lot!

Tong

[Alvin Vista]

Hi Tong,
Please include your data and command files so that we can replicate the errors you've encountered.
Cheers,
Alvin

[shentong]

Hi Alvin,

Thanks for your reply! The data and command files are attached below.

Thanks again,
Tong

[Alvin Vista]

Hi Tong,
I don't understand your model. It seems you have all items load on two dimensions, but the two dimensions are exactly the same.
This does not make sense to me, please clarify what you are trying to do.
Alvin

[shentong]

Hi Alvin,

Thanks for the reply! Yes, actually this is the part that I'm concerned about, too. The instrument that I'm working on is a two-dimension scenario type scale. Each item measures two dimensions at the same time but with different levels. E.g., item 1 is a scenario where a person with high levels of dimension 1 and low levels of dimension 2 would do. Together all items are ranked in a hierarchical order. Participants respond by comparing themselves to each scenario and then rate themselves on a 4-point scale. In this case, each item is measuring two dimensions at the same time, so I employed the within-item model and had all items load on two dimensions.
Is it because participates only give one score for each item, it is then hard for the software to untangle the two latent traits...? Or is there some other problems? Would you suggest a better way of modeling it?

Thanks,
Tong

[Alvin Vista]

"Is it because participates only give one score for each item, it is then hard for the software to untangle the two latent traits...? "

Yes, under the current setup, there is no way to disentangle the two latent traits.
I can't assist you in the design of your study but I suggest that you revise 1) the design of your items and/or, 2) the design of how your case subjects take the items depending on their hypothesised traits.