Hi, thank you very much for your videos. I would like to kindly ask you two questions: i) If our data is non-normally distributed what should we do? Bootstrapping method? ii) In this case, sex is "time-invariant covariates (TICs) that do not change in value as a function of time and TVCs that at least in principle can change as a function of time". You presented an awsome video for TICs and what about for TVCs, do you recommend any procedure? Any video that we can follow? Thank you in advance!
Hi Mike, thanks for your very insightful videos. I have a simple question regarding the last model you described (adding sex): - your slope estimate is negative (-.089); this means that there's a general decreasing tendency of the DV over time - however, sex predicts the slope negatively (-.15): your exact wording is "the amount of change that we observed over time for males is lower than those of females". I'm a bit puzzled by the term "amount of change". Does this mean that the decreasing tendency (-.089) is dampened for males? So they actually decrease LESS than females? Or should we interpret it in the sense that males' decrease over time is even larger than that of females? Thanks in advance for your help in these hard statistical topics. All the best, Alessandro
Hi mike, thanks for your excellent videos. I would like to ask 3 questions if you have any time to answer it would be very appreciated. 1) what if I do not estimate all variances freely? In my model, when I freely estimated all variances, slope variance was not significant. But when not freely estimated, slope variance was significant but the model fit was bad. 2) If level 1 model fit not good especially for RMSEA can I add level 2 variable such as personality to estimate intercept and slope. When I do this my model fit improved but I am not sure can I add level 2 variable before level 1 model fit reach to acceptable level? 3) If a negative slope, which is yielded in first model, is negatively predicted by level 2 variable , does it mean that slope's negative direction become flat ? Thank you in advance. By the way a video about piecewise LGCM would be perfect :)
Hi Mike, thanks for your lecture, it is really helpful. I saw you used the mean of perceived math ability to indicate the observed valuable, there are actually three items assessing the perceived math ability, why didn't you use the score of each item as the observed valuable like normal SEM analysis on AMOS? Thank you very much!
Mike is it possible to compare both models? The initial model without sex and the one that includes sex as an exogenous variable. How can we do that on the same layout?
Thank you very much for sharing this useful video. I just start learning this and get a few questions: 1. In this example, test scores at 3 different times (i.e., 3 observed variables) are used. How is it different from ANOVA? 2. Can we include latent variables in latent growth analysis? 3. Can this method be used to analyze test data collected at 2 or 4 times?
Thanks for watching! Regarding your questions: (a) Byrne (2016) indicates that you shouldn't be doing latent growth curve modeling in SEM with less than 3 waves (so two waves is too few); (b) If you are asking about modeling growth curve factors associated with latent factors, there is an approach called the "Curve of factors model" (see e.g., journals.sagepub.com/doi/full/10.1177/0013164416677143); (c) If you are asking how this is different from repeated measures ANOVA, the answer is that you are modeling individual growth curves over time (whereas RM ANOVA is testing for change over time at the group level only; it ignores the individual information). Hope this helps!
Thank you very much for the reply. It really helps. I got some more questions. If I understand it correctly, the model shown in your video assumes that the change over time is linear, right? What if it's not linear? Before doing latent growth analysis, are there any assumptions we need to check first?
You're very welcome. You can model non-linear relationships as well. If you are assuming a quadratic trend, you can add a second slope factor (correlated with the first slope factor and the intercept factor) along with fixed loadings that are the square of those you used with the linear factor. Another alternative is that if you don't know the shape you have in mind you could estimate the shape using the linear factor only and fixing the first loading (associated with time 1 measurement) to 0 and the last loading (for the last time point) to 1 and leaving the remaining loadings free to be estimated. The only downside to this is that it becomes a bit trickier in describing the growth trend if you have a more than a few time points represented. The standard SEM assumption of multivariate normality of your variables applies (this will be the observed variables - repeated measures in your analysis). Sample size should be at least 200 (per Byrne, 2016; www.taylorfrancis.com/books/9781317633136).
Thank you very much. The second slope factor thing is beyond me. Do you have a video about that? I didn't find it in Byrne's book. Can we perform correlation between variables or draw scatterplot to see whether the change over time is linear or not?
I don't have a video on that. However, there is a powerpoint here (psy.mq.edu.au/psystat/CEH/LGM-ls.pdf) that includes a demonstration of what you are asking about. There also appears to be some demonstration of how you can obtain growth curves for each case. You can also check out another pedagogical article here: www.ncbi.nlm.nih.gov/pmc/articles/PMC2888524/. I also demonstrate this in another video illustrating growth curve modeling with HLM in SPSS (ua-cam.com/video/_MjEYz96TA8/v-deo.html). However, to do it the way I show you, the data has to first be in long (vertical) form. Here's a video on restructuring the data to do this: ua-cam.com/video/9jJbAnbW3jI/v-deo.html
Thank you very very much, Mr. Crowson. I find all the videos I've watched statistically significantly helpful.
Hi, thank you very much for your videos. I would like to kindly ask you two questions:
i) If our data is non-normally distributed what should we do? Bootstrapping method?
ii) In this case, sex is "time-invariant covariates (TICs) that do not change in value as a function of time and TVCs that at least in principle can change as a function of time". You presented an awsome video for TICs and what about for TVCs, do you recommend any procedure? Any video that we can follow?
Thank you in advance!
Thanks so much for this!
You're welcome! Thanks for watching!
Hi Mike, thanks for your very insightful videos. I have a simple question regarding the last model you described (adding sex):
- your slope estimate is negative (-.089); this means that there's a general decreasing tendency of the DV over time
- however, sex predicts the slope negatively (-.15): your exact wording is "the amount of change that we observed over time for males is lower than those of females". I'm a bit puzzled by the term "amount of change".
Does this mean that the decreasing tendency (-.089) is dampened for males? So they actually decrease LESS than females?
Or should we interpret it in the sense that males' decrease over time is even larger than that of females?
Thanks in advance for your help in these hard statistical topics.
All the best,
Alessandro
Great information!
How do I run the model over 2 time points?
I have been model identification problems
If I only have 6 participants but almost 50 period of observation (repeated measures), is it justifiable to do LGCM?
Hi mike, thanks for your excellent videos. I would like to ask 3 questions if you have any time to answer it would be very appreciated. 1) what if I do not estimate all variances freely? In my model, when I freely estimated all variances, slope variance was not significant. But when not freely estimated, slope variance was significant but the model fit was bad. 2) If level 1 model fit not good especially for RMSEA can I add level 2 variable such as personality to estimate intercept and slope. When I do this my model fit improved but I am not sure can I add level 2 variable before level 1 model fit reach to acceptable level? 3) If a negative slope, which is yielded in first model, is negatively predicted by level 2 variable , does it mean that slope's negative direction become flat ? Thank you in advance. By the way a video about piecewise LGCM would be perfect :)
Hi Mike, thanks for your lecture, it is really helpful. I saw you used the mean of perceived math ability to indicate the observed valuable, there are actually three items assessing the perceived math ability, why didn't you use the score of each item as the observed valuable like normal SEM analysis on AMOS? Thank you very much!
Mike is it possible to compare both models? The initial model without sex and the one that includes sex as an exogenous variable. How can we do that on the same layout?
Thank you very much for sharing this useful video. I just start learning this and get a few questions:
1. In this example, test scores at 3 different times (i.e., 3 observed variables) are used. How is it different from ANOVA? 2. Can we include latent variables in latent growth analysis? 3. Can this method be used to analyze test data collected at 2 or 4 times?
Thanks for watching! Regarding your questions: (a) Byrne (2016) indicates that you shouldn't be doing latent growth curve modeling in SEM with less than 3 waves (so two waves is too few); (b) If you are asking about modeling growth curve factors associated with latent factors, there is an approach called the "Curve of factors model" (see e.g., journals.sagepub.com/doi/full/10.1177/0013164416677143); (c) If you are asking how this is different from repeated measures ANOVA, the answer is that you are modeling individual growth curves over time (whereas RM ANOVA is testing for change over time at the group level only; it ignores the individual information). Hope this helps!
Thank you very much for the reply. It really helps. I got some more questions. If I understand it correctly, the model shown in your video assumes that the change over time is linear, right? What if it's not linear? Before doing latent growth analysis, are there any assumptions we need to check first?
You're very welcome. You can model non-linear relationships as well. If you are assuming a quadratic trend, you can add a second slope factor (correlated with the first slope factor and the intercept factor) along with fixed loadings that are the square of those you used with the linear factor. Another alternative is that if you don't know the shape you have in mind you could estimate the shape using the linear factor only and fixing the first loading (associated with time 1 measurement) to 0 and the last loading (for the last time point) to 1 and leaving the remaining loadings free to be estimated. The only downside to this is that it becomes a bit trickier in describing the growth trend if you have a more than a few time points represented. The standard SEM assumption of multivariate normality of your variables applies (this will be the observed variables - repeated measures in your analysis). Sample size should be at least 200 (per Byrne, 2016; www.taylorfrancis.com/books/9781317633136).
Thank you very much. The second slope factor thing is beyond me. Do you have a video about that? I didn't find it in Byrne's book.
Can we perform correlation between variables or draw scatterplot to see whether the change over time is linear or not?
I don't have a video on that. However, there is a powerpoint here (psy.mq.edu.au/psystat/CEH/LGM-ls.pdf) that includes a demonstration of what you are asking about. There also appears to be some demonstration of how you can obtain growth curves for each case. You can also check out another pedagogical article here: www.ncbi.nlm.nih.gov/pmc/articles/PMC2888524/. I also demonstrate this in another video illustrating growth curve modeling with HLM in SPSS (ua-cam.com/video/_MjEYz96TA8/v-deo.html). However, to do it the way I show you, the data has to first be in long (vertical) form. Here's a video on restructuring the data to do this: ua-cam.com/video/9jJbAnbW3jI/v-deo.html