Hello,
I want to compose two latent variable-model.
One latent variable is "socioeconomic factors" in a study. Possible indicators are "life expectancy at birth", "number of doctors per capita" etc. which are obtained from Worldbank database.
The point is I plan to use secondary data. I want to know that if it is required to mind reliability and validity?
Thnx for your help in advance,
Hello Ceren,
in your case, socio-economic factors is either a simple aggregate of your primary variables (which your use of the plural in factorS suggests) or a latent formative variable which is affected by these primary variables.
In contrast, a measurement model with a common factor model makes sense as this would propose that a latent variabel that reflects the position of a country on a world-wide socioeconomic dimension (keep in mind the inherent country-level nature of your variables) affects the indicators and thus causes their values and their correlations. But perhaps I am wrong. So if you down load the data, you could specify this model and test it. The N in most worldbank variables is not huge but I guess sufficient to refuse the model ;)
All the best,
Holger
The distinctive feature of your data is not that it is secondary but that is aggregate and about countries not individuals. you therefore need to be careful about the ecological fallacy and not to infer to individuals but stick to countries.
Holger Steinmetz Kelvyn Jones Thank you for your contributions. Let me ask another way. Let's assume I have reflective latent (F) and indicators x1,x2 and x3. Let's assume I am pretty sure about the reflective nature of latent F, and I have strong theoretical reasons to connect X's with F. In this case, are validity and reliability calculations necessary/possible?
Thank you on advance
Hi Ceren,
every scientific piece of information supporting a claim or view adds to the pile of information. That creates a continuoum of evidence strength.
With regard to common factor models, at the lowest level you would have 3 indicators and one latent variable. This model is just identified and has no testable implications. Therefore, the only piece of support is the strenght of the factor loadings.
Adding one (or more ) other variables to the model would--depending on the structure, add testable implications of the whole model and your embedded factor model. A successful test would again increase the strength of evidence.
If this new variable (call it Z) is theoretically meaningful as it --according to your theory--would have a relation to your latent variable of interest, then the estimated relationship would add again to the pile of evidence.
If you conduct an experiment where you intend (and manage) to manipulate/change the latent variable by some supposed change process, than again this would add up. The difference between case #3 and this one would be that even if the relation with Z is as it expected, it could be due to confounding which the experimental scenerio hopefully eliminates.
I hope you get the point.
Best,
Holger
- Badges
- Science topic
- Similar topics
- Methodology
- Research Methodology
- Research Papers