I have heard that SEM is used for a sample size of larger than 200 data points. How can we use it for smaller samples
Hi,
small samples lead to biases of the chisquare test towards false rejections but at the same time reduce the power to correctly reject a false model. For this case, Herzog and Boomsma have developed a correction for the chi-square and fit indexes, and provide a R-code for creating a nice little function that is feed by the outcomes of your model (chisq and indexes) and prints the corrected versions.
That having said, the parameter estimates still may have some problems as ML estimates are asymptotically unbiased and thus biased in in small samples.
Best,
Holger
Herzog, W., & Boomsma, A. (2009). Small-sample robust estimators of noncentrality-based and incremental model fit. Structural Equation Modeling: A Multidisciplinary Journal, 16(1), 1–27. Journal Article. http://doi.org/10.1080/10705510802561279
https://www.researchgate.net/publication/48872552_Small-Sample_Robust_Estimators_of_Noncentrality-Based_and_Incremental_Model_Fit
http://www.ppsw.rug.nl/~boomsma/swain.pdf
Article Small-Sample Robust Estimators of Noncentrality-Based and In...
That depends on the data and the model.
If the conditional distribution of the data is relatively symmetric, as few as 5 to 10 values per coefficient in the model usually give sufficiently good estimates of the standard errors.
Hi,
if you have no possibility to extend your data due to some methodological considerations, you can use SmartPLS...it can be suitable
Hi,
Not sure if you want to use SEM for small sample size. But you can also plan on the numbers of constructs. If you have less than 5 latent constructs with three measuring items for each constructs, the minimum sample size required is 100. 7 or less latent constructs also with 3 measuring items for each constructs, you need minimum of 150 samples.
Hair et al (2010) Multivariate Data Analysis. Englewood Cliffs.
HI,
If your sample size is less, then you can try using SMART-PLS. Again the concern is you should have at-least 10 responses for a path between a exogenous and endogenous construct. For example if you have 2 exogenous constructs and 1 endogenous construct there will be two paths from exogenous to endogenous. Then the required sample size 2X10=20 samples. This you can refer from Hair et al book on Smart-PLS. But again there are few concerns in using SMART-PLS, which you can see from the articles on SMART-PLS. For Clarity you can check in the Forum on smart-PLS.
- Badges
- Science topic
- Similar topics
- Psychology