Thanks, will do! The COVID-19 situation is improving, so we will soon be able to return to to office and I get back to shooting videos. I have a few interesting things on my list.
Hi Dr Rönkkö, thanks for the great video! Apart from dimensionality and tau-equivalence, do we have to take into account other factors? For instance, I have two questionnaires which are both two-dimensional and each factor is not tau-equivalent. But my data (n=35) are not normally distributed. Do I have to take this into consideration? Then, what coefficient should I use? Thanks in advance!
Your sample size is so small that it limits what you can do. Normal distribution or lack of it has generally little consequences for reliability coefficients. I would go for Composite reliability (or omega, which is the same thing) as a default alternative unless you have a good reason to go for something else.
Could you please tell, can we calculate the mean and standard deviation from the Likert scale data(meeting unidimensionality and tau-equivalence assumptions)?
I am not sure if this is the answer that you are looking for, but you calculate mean and SD like for any data. Mean is the sum of items divided by their number and estimated SD is the square root of sum of squared differences from mean divided by the number of items minus one. www.khanacademy.org/math/statistics-probability/summarizing-quantitative-data/variance-standard-deviation-population/a/calculating-standard-deviation-step-by-step
@@mronkko Thank you for the response. I actually asked about item response theory and classical test theory. For example, if I am asking 400 respondents about mobile service quality. Can I add all individual items and calculate mean, standard deviation? Furthermore, can I apply the calculated mean score of each variable to calculate correlation or regression?
@Laxman Pokhrel Using scale scores is never ideal, but it might be a good practical alternative. I would check what researchers in your area do to see what the standard is. That being said, there are two things to consider: 1) in an IRT model, you are assuming a nonlinear item response curve. Does it make a difference if you replace this with a linear one? If you have five or more scale points and (if I remember correctly) three or more items per scale, the differences in factor correlations are negligible. See this paper for details Rhemtulla, M., Brosseau-Liard, P. É., & Savalei, V. (2012). When can categorical variables be treated as continuous? A comparison of robust continuous and categorical SEM estimation methods under suboptimal conditions. Psychological Methods, 17(3), 354-373. dx.doi.org/10.1037/a0029315 2) Does it make sense to use scales scores instead of latent variables after you have made the linearity assumption? This depends on your level of expertise and whether you can trust that the measurement model is correctly specified. Rhemtulla, M., van Bork, R., & Borsboom, D. (2020). Worse than measurement error: Consequences of inappropriate latent variable measurement models. Psychological Methods, 25(1), 30-45. doi.org/10.1037/met0000220 In practice, aggregating likert scales as scores is a common and solid research practice but I would personally hope that researches would justify their choices on methods (scale scores instead of latent variables) a bit more.
exactly the reasoning for citation that I needed in my thesis. Great video.
You are welcome
thank you so much!
keep making great videos
Thanks, will do! The COVID-19 situation is improving, so we will soon be able to return to to office and I get back to shooting videos. I have a few interesting things on my list.
Hi Dr Rönkkö, thanks for the great video! Apart from dimensionality and tau-equivalence, do we have to take into account other factors? For instance, I have two questionnaires which are both two-dimensional and each factor is not tau-equivalent. But my data (n=35) are not normally distributed. Do I have to take this into consideration? Then, what coefficient should I use? Thanks in advance!
Your sample size is so small that it limits what you can do. Normal distribution or lack of it has generally little consequences for reliability coefficients. I would go for Composite reliability (or omega, which is the same thing) as a default alternative unless you have a good reason to go for something else.
@@mronkko Thank you very much! Highly appreciated 😊
🔥🔥🔥❤️❤️❤️
Could you please tell, can we calculate the mean and standard deviation from the Likert scale data(meeting unidimensionality and tau-equivalence assumptions)?
I am not sure if this is the answer that you are looking for, but you calculate mean and SD like for any data. Mean is the sum of items divided by their number and estimated SD is the square root of sum of squared differences from mean divided by the number of items minus one.
www.khanacademy.org/math/statistics-probability/summarizing-quantitative-data/variance-standard-deviation-population/a/calculating-standard-deviation-step-by-step
@@mronkko Thank you for the response. I actually asked about item response theory and classical test theory. For example, if I am asking 400 respondents about mobile service quality. Can I add all individual items and calculate mean, standard deviation? Furthermore, can I apply the calculated mean score of each variable to calculate correlation or regression?
@Laxman Pokhrel Using scale scores is never ideal, but it might be a good practical alternative. I would check what researchers in your area do to see what the standard is. That being said, there are two things to consider: 1) in an IRT model, you are assuming a nonlinear item response curve. Does it make a difference if you replace this with a linear one? If you have five or more scale points and (if I remember correctly) three or more items per scale, the differences in factor correlations are negligible. See this paper for details
Rhemtulla, M., Brosseau-Liard, P. É., & Savalei, V. (2012). When can categorical variables be treated as continuous? A comparison of robust continuous and categorical SEM estimation methods under suboptimal conditions. Psychological Methods, 17(3), 354-373. dx.doi.org/10.1037/a0029315
2) Does it make sense to use scales scores instead of latent variables after you have made the linearity assumption? This depends on your level of expertise and whether you can trust that the measurement model is correctly specified.
Rhemtulla, M., van Bork, R., & Borsboom, D. (2020). Worse than measurement error: Consequences of inappropriate latent variable measurement models. Psychological Methods, 25(1), 30-45. doi.org/10.1037/met0000220
In practice, aggregating likert scales as scores is a common and solid research practice but I would personally hope that researches would justify their choices on methods (scale scores instead of latent variables) a bit more.
@@mronkko Thanks! This is a lot to me. I am sharing your very nice course among Nepali researchers.