From Data to Causes I: Building a General Cross-Lagged Panel Model (GCLM)

แชร์
ฝัง
  • เผยแพร่เมื่อ 12 ธ.ค. 2024

ความคิดเห็น • 54

  • @simonalaba2250
    @simonalaba2250 4 ปีที่แล้ว +5

    Very well explained Micheal. Thanks for taking the time to put this together, it's an excellent accompaniment to your papers.

    • @michaelzyphur2476
      @michaelzyphur2476  2 ปีที่แล้ว

      I wanted to get in touch about the Institute for Statistical and Data Science (Instats), which is now live and providing a range of statistics seminars on a variety of topics at instats.com.au . For our initial June launch, we're giving a 50% discount on all of the Mplus seminars that I'm delivering, which is in addition to the substantial PhD student discounts and the 10% discount on all Structured Programs. We offer workshops in English and Mandarin, and you can see all workshops currently available at instats.com.au/workshops

  • @luisluna11
    @luisluna11 4 ปีที่แล้ว +2

    This is amazing, thank you!

    • @michaelzyphur2476
      @michaelzyphur2476  2 ปีที่แล้ว

      Hi, I wanted to get in touch about the Institute for Statistical and Data Science (Instats), which is now live and providing a range of statistics seminars on a variety of topics at instats.com.au . For our initial June launch, we're giving a 50% discount on all of the Mplus seminars that I'm delivering, which is in addition to the substantial PhD student discounts and the 10% discount on all Structured Programs. We offer workshops in English and Mandarin, and you can see all workshops currently available at instats.com.au/workshops

  • @longzhudong2777
    @longzhudong2777 ปีที่แล้ว

    Thank you so much! Very well explained and really helpful. I have one question, if we have nested data (e.g. students in classes) and we need to control for class characteristics (e.g. sizes), do we need to conduce multiple level cross lagged SEM, OR, the “co-movements” terms 31:26 already take any potential impacts from classes into consideration, so no multilevel analysis will be needed?

  • @ureshitanoshi168
    @ureshitanoshi168 2 ปีที่แล้ว

    Hi Michael, it may be a small matter in the end but since x sub t-2 at the point 27:36 mark is still exogenous isn't it premature to add u sub t-1 then (it makes sense upon the introduction of the random intercept), right?

    • @michaelzyphur2476
      @michaelzyphur2476  2 ปีที่แล้ว

      Hi Ureshi, thanks, that's an interesting point. It seems like we always are in the strange position of having to make some assumptions about exogeneity and initial conditions. I think you're intuition is correct that to understand a model as a whole, the way that I present the model while it's being built should not taken unconditionally and instead the model should be understood only at the end :-) ... By the way, if you want to see more updated workshops or participate, please head to Instats.com where you can find many new resources! Best wishes and thanks again.

  • @Paolo-tu2ft
    @Paolo-tu2ft 4 ปีที่แล้ว

    Prof. Zyphur, this is an amazing explanation. I am a nursing PHD student and I have started approaching cross lagged panel models with my data in MPLUS for a future research on self-care. I have started with a very basic model with two variables measured in three waves. Although I used your EXCEL sheet (I copied and pasted) I didn't reach the convergence of the model. MPLUS told me that I have negative degrees of freedom. In fact, having only 21 elements in the matrix, it is quite easy to run out of degrees of freedom with such a high number of parameter estimations. I tried for example, not to include unit effects, and it worked actually. Although in your model, you used 6 waves and did not run out of degrees of freedom (it was over identified and with 56 degrees of freedom!). Could you explain me why that happens, if I am missing something and how I can overcome this problem? Thanks in advance Prof! Paolo

    • @michaelzyphur2476
      @michaelzyphur2476  2 ปีที่แล้ว

      I wanted to get in touch about the Institute for Statistical and Data Science (Instats), which is now live and providing a range of statistics seminars on a variety of topics at instats.com.au . For our initial June launch, we're giving a 50% discount on all of the Mplus seminars that I'm delivering, which is in addition to the substantial PhD student discounts and the 10% discount on all Structured Programs. We offer workshops in English and Mandarin, and you can see all workshops currently available at instats.com.au/workshops

  • @victoriawingate3435
    @victoriawingate3435 5 ปีที่แล้ว

    Yeah, I'd take a drink, too. Amazing. Thank you.

    • @michaelzyphur2476
      @michaelzyphur2476  5 ปีที่แล้ว

      Pleasure! Glad you liked it... drink up :-)

    • @michaelzyphur2476
      @michaelzyphur2476  2 ปีที่แล้ว

      Hi, I wanted to get in touch about the Institute for Statistical and Data Science (Instats), which is now live and providing a range of statistics seminars on a variety of topics at instats.com.au . For our initial June launch, we're giving a 50% discount on all of the Mplus seminars that I'm delivering, which is in addition to the substantial PhD student discounts and the 10% discount on all Structured Programs. We offer workshops in English and Mandarin, and you can see all workshops currently available at instats.com.au/workshops

  • @DavidLopez-yv6ft
    @DavidLopez-yv6ft 5 ปีที่แล้ว

    really useful. Thanks a lot !!!

    • @michaelzyphur2476
      @michaelzyphur2476  2 ปีที่แล้ว

      Hi, I wanted to get in touch about the Institute for Statistical and Data Science (Instats), which is now live and providing a range of statistics seminars on a variety of topics at instats.com.au . For our initial June launch, we're giving a 50% discount on all of the Mplus seminars that I'm delivering, which is in addition to the substantial PhD student discounts and the 10% discount on all Structured Programs. We offer workshops in English and Mandarin, and you can see all workshops currently available at instats.com.au/workshops

  • @erfanimtiaz9463
    @erfanimtiaz9463 4 ปีที่แล้ว

    Very interesting approach to panel data estimation. I was wondering if it can be used also for a longitudinal "Linear-in-Means Choice Model (peer effects choice model)" study where outcome is a 3 alternative nominal variable and mix of categorical and continues variables are on the right hand side of the equation - with an unbalanced panel data up to 50 time periods. the abstract model can be written as:
    Y(i,t) = alpha + a(i,t)Y(i,t-1) + b(i,t)X(i,t) + SUM_or_SIGMA_over_Choice_Options [ c(i,t,n)Peer_Average_of_X(i,t,n) + d(i,t,n)Peer_Average_of_Y(i,t,n) ] + f(t)

    • @michaelzyphur2476
      @michaelzyphur2476  4 ปีที่แล้ว

      Hi Erfan, I'm not familiar with the "Linear-in-Means Choice Model (peer effects choice model)", but one thing to consider is that if Y is parameterized with a logit or probit link and estimated with ML then you'll have threshold parameters but not residual variances, in which case there are no impulses/residuals as per the GCLM. Also in the GCLM as in VARs everything is endogenous, so all of your right-hand-side variables would also need to be brought into the model. Best of luck!

    • @erfanimtiaz9463
      @erfanimtiaz9463 4 ปีที่แล้ว

      @@michaelzyphur2476 Thank you very much for the explanation.

    • @michaelzyphur2476
      @michaelzyphur2476  2 ปีที่แล้ว

      I wanted to get in touch about the Institute for Statistical and Data Science (Instats), which is now live and providing a range of statistics seminars on a variety of topics at instats.com.au . For our initial June launch, we're giving a 50% discount on all of the Mplus seminars that I'm delivering, which is in addition to the substantial PhD student discounts and the 10% discount on all Structured Programs. We offer workshops in English and Mandarin, and you can see all workshops currently available at instats.com.au/workshops

  • @mroienalm
    @mroienalm 5 ปีที่แล้ว +1

    This presentation is really helpful. Thank you so much, Dr. Zyphur (and your co-authors). I have a question. I downloaded all the supplementary files, but I cannot open the Excel file for generating M+ codes (GCLM Code Generator for Mplus.xlsx). The file seems broken, but I am not sure if I did something wrong (I downloaded the zip file multiple times and tested) or the file in the database is really broken. I would really appreciate it if you can double-check and let me know where I can download a one without any error. Again, thank you so much.

    • @michaelzyphur2476
      @michaelzyphur2476  5 ปีที่แล้ว

      Hi SeungYong, thanks for letting me know about this. Some people have had trouble downloading the large zip file. If you only want the Mplus code generator Excel file, you can download it here: www.dropbox.com/s/ruzsqtgg3ytlxvf/GCLM%20Code%20Generator%20for%20Mplus.xlsx?dl=0

    • @michaelzyphur2476
      @michaelzyphur2476  2 ปีที่แล้ว

      Hi, I wanted to get in touch about the Institute for Statistical and Data Science (Instats), which is now live and providing a range of statistics seminars on a variety of topics at instats.com.au . For our initial June launch, we're giving a 50% discount on all of the Mplus seminars that I'm delivering, which is in addition to the substantial PhD student discounts and the 10% discount on all Structured Programs. We offer workshops in English and Mandarin, and you can see all workshops currently available at instats.com.au/workshops

  • @sunchaoran8606
    @sunchaoran8606 3 ปีที่แล้ว

    Hi Michael, very appreciate your video and it has been very informative for my study. Just have a question that do two-wave measurement count as panel data and if RI-CLPM work on that?

    • @michaelzyphur2476
      @michaelzyphur2476  2 ปีที่แล้ว +1

      Hi, yes two-wave data are panel data. However, you can't do much with only two waves. See our first and second GCLM paper where we discuss the identification conditions for panel data models that account for the latent stable part. This requires at least 3 waves of data to distinguish it from the AR term.

    • @michaelzyphur2476
      @michaelzyphur2476  2 ปีที่แล้ว

      I wanted to get in touch about the Institute for Statistical and Data Science (Instats), which is now live and providing a range of statistics seminars on a variety of topics at instats.com.au . For our initial June launch, we're giving a 50% discount on all of the Mplus seminars that I'm delivering, which is in addition to the substantial PhD student discounts and the 10% discount on all Structured Programs. We offer workshops in English and Mandarin, and you can see all workshops currently available at instats.com.au/workshops

  • @liamjwright
    @liamjwright 4 ปีที่แล้ว +1

    This is great. Thank you!
    I'm a little confused about the relationship between your unit-effect AR(1)MA(1)CL(1)CLMA(1) model and Hamaker et al.'s (2015) RI-CLPM model. They restrict unit effect loadings to 1 and within-centred x and y variables transmit AR and CL effects (lavaan code here: ellenhamaker.github.io/RI-CLPM/lavaan.html). In the code from your supplementary information, it is the manifest variables that transmit AR and CL effects. Is this just a difference in coding or is it substantive?

    • @michaelzyphur2476
      @michaelzyphur2476  4 ปีที่แล้ว

      The difference is not only in coding. The difference is in whether observed vs. latent W variables are used for the AR and CL terms. Really the only difference except for the invariant factor loadings over time is how the stable B part which is typically called a fixed-effect (or 'unit effects' in our paper or 'random intercept' in Hamaker et al.) is understood. In Hamaker et al. it really is the stable over-time average but in our model it is adjusted by the magnitudes of the AR and CL terms. However, this interpretability comes at a cost of model identification when AR or CL terms approach unity in the case of an unstable process such as a random walk. In this case our model is still identified and estimable. We may have more to say on this later but for now I'd just mention this as being potentially relevant. For the most part the difference does not matter.

    • @liamjwright
      @liamjwright 4 ปีที่แล้ว +1

      @@michaelzyphur2476 That's very helpful. Thank you. (Also thank you for the other videos which are similarly excellent.)

    • @michaelzyphur2476
      @michaelzyphur2476  2 ปีที่แล้ว

      I wanted to get in touch about the Institute for Statistical and Data Science (Instats), which is now live and providing a range of statistics seminars on a variety of topics at instats.com.au . For our initial June launch, we're giving a 50% discount on all of the Mplus seminars that I'm delivering, which is in addition to the substantial PhD student discounts and the 10% discount on all Structured Programs. We offer workshops in English and Mandarin, and you can see all workshops currently available at instats.com.au/workshops

  • @olafsimonse
    @olafsimonse 4 ปีที่แล้ว +1

    I can use a drink now, indeed. Can't help thinking: "how would you model needing a drink after a talk?".

    • @michaelzyphur2476
      @michaelzyphur2476  4 ปีที่แล้ว +1

      Hah! Some things require modeling. Other things merely require doing :-)

    • @olafsimonse
      @olafsimonse 4 ปีที่แล้ว

      @@michaelzyphur2476 On a more serious note. I have a couple of questions:
      - I would like to apply a CL-model to a dataset with quite some outliers, heavy tailed distrubutions and multicollinearity. In other words, some of the basic assumptions of linear regression are not met. Can your method handle that?
      - I have a model with four variables, and would like to include their second- and third-order interactions. So in total that would include 14 variables/interactions. Can that be done?
      To be honest, I haven't read your papers yet, so "go and read the papers first" would be a completely acceptable answer. Any tips in addition are a bons. Thanks.

    • @michaelzyphur2476
      @michaelzyphur2476  4 ปีที่แล้ว +1

      @@olafsimonse there are a few issues here: 1) for non-normal variables you can use an ML estimator that is robust to these but probably better would either be normalization or instead using a skew-normal or other distribution which Mplus allows (see some discussion here tinyurl.com/y4hdmj3q); 2) interactions with in models like these requires latent interactions among the impulses. Reading the two-paper series to familiarize yourself with the basic modeling setup and downloading the online Excel spreadsheet in the supplemental materials to first get a model running would be a good starting step. Then you could modify the input to allow/test for skewness and consider latent interactions. Online materials are here tinyurl.com/ya5spuuw

    • @olafsimonse
      @olafsimonse 4 ปีที่แล้ว +1

      Thank you very much, Michael!

    • @michaelzyphur2476
      @michaelzyphur2476  2 ปีที่แล้ว

      I wanted to get in touch about the Institute for Statistical and Data Science (Instats), which is now live and providing a range of statistics seminars on a variety of topics at instats.com.au . For our initial June launch, we're giving a 50% discount on all of the Mplus seminars that I'm delivering, which is in addition to the substantial PhD student discounts and the 10% discount on all Structured Programs. We offer workshops in English and Mandarin, and you can see all workshops currently available at instats.com.au/workshops

  • @arnatw8908
    @arnatw8908 3 ปีที่แล้ว

    How can I apply this method to perform a Random intercepts cross-lagged panel model? I have three waves of data.

    • @michaelzyphur2476
      @michaelzyphur2476  3 ปีที่แล้ว

      Hi, if you want an RI-CLPM why not just use that? If you want something like it, just eliminate the MA and CLMA paths. Good luck!

    • @michaelzyphur2476
      @michaelzyphur2476  2 ปีที่แล้ว

      I wanted to get in touch about the Institute for Statistical and Data Science (Instats), which is now live and providing a range of statistics seminars on a variety of topics at instats.com.au . For our initial June launch, we're giving a 50% discount on all of the Mplus seminars that I'm delivering, which is in addition to the substantial PhD student discounts and the 10% discount on all Structured Programs. We offer workshops in English and Mandarin, and you can see all workshops currently available at instats.com.au/workshops

  • @沈卷
    @沈卷 5 ปีที่แล้ว

    thank you so much,this is really helpful. I have a question.how to add
    Two or more Moderators into this model .

    • @michaelzyphur2476
      @michaelzyphur2476  5 ปีที่แล้ว

      Hi, thanks. We discuss this in the online supplemental materials. There are various kinds of moderators than can be used -- same-level and cross-level moderators. Either way, latent variable interactions are needed, either among impulses or unit effects. Good luck!

    • @沈卷
      @沈卷 5 ปีที่แล้ว

      Michael Zyphur thank you very much😻

    • @michaelzyphur2476
      @michaelzyphur2476  2 ปีที่แล้ว

      Hi, I wanted to get in touch about the Institute for Statistical and Data Science (Instats), which is now live and providing a range of statistics seminars on a variety of topics at instats.com.au . For our initial June launch, we're giving a 50% discount on all of the Mplus seminars that I'm delivering, which is in addition to the substantial PhD student discounts and the 10% discount on all Structured Programs. We offer workshops in English and Mandarin, and you can see all workshops currently available at instats.com.au/workshops

  • @jithesha3214
    @jithesha3214 5 ปีที่แล้ว

    Hi, I have a question. If I have a model as below: There is a mediated relationship between X and Y through M. Also, all the relationships are moderated by W. I have run this model with a cross-sectional data, and everything is good. Now will it be possible to run the specified model with a panel data (say, a data for 10 years). Will this video help me do that? If yes, may you let me know if it is possible to do this is R? I have to also mention that X, M, Y, and W are not latent variables. All of them are observed variables.

    • @michaelzyphur2476
      @michaelzyphur2476  5 ปีที่แล้ว

      Hi Jithesh, it sounds like you want to estimate moderated mediation (i.e., moderated indirect effects) of X-->M-->Y variety with W moderating at least one of these two paths. This is technically possible, but in a cross-lagged model you must ensure that only the within-person dynamic parts are used for the moderation analysis. This requires latent interactions with the residuals or 'impulses' as we describe them. Check out the Organizational Research Methods papers that this talk is based on. In the special materials for the first paper, I think, we discuss this. Good luck!

    • @jithesha3214
      @jithesha3214 5 ปีที่แล้ว

      @@michaelzyphur2476 Thank you very much for the reply. Will go through the resources you suggested.

    • @jithesha3214
      @jithesha3214 5 ปีที่แล้ว

      @@michaelzyphur2476 Another question is regarding the following constraints I have and wish to know if this method is suitable in that scenario.
      Constraints are as follows:
      1) Only three years of data is available ( 2012, 2014, and 2016)
      2) There is one IV (X), one DV (Y), one mediator (M), and one moderator (W)
      3) Number of data points for one year is 125
      May you kindly suggest which is the best method to run the above model? Thank you sir.

    • @michaelzyphur2476
      @michaelzyphur2476  5 ปีที่แล้ว

      @@jithesha3214 , in this case of T=3 and N=125 you will need to watch out for model non-identification. Primarily I would watch out for the T=3 constraint because with this you cannot estimate a GCLM. Read my two papers, including the special materials, and we discuss model identification conditions. The GCLM requires T=4. A more constrained model with only AR and CL paths and a latent 'unit effect' with t=2 and t=3 factor loadings constrained to = 1 is identified. Read my two papers as linked in the video's description and you'll see. Good luck!

    • @michaelzyphur2476
      @michaelzyphur2476  2 ปีที่แล้ว

      Hi, I wanted to get in touch about the Institute for Statistical and Data Science (Instats), which is now live and providing a range of statistics seminars on a variety of topics at instats.com.au . For our initial June launch, we're giving a 50% discount on all of the Mplus seminars that I'm delivering, which is in addition to the substantial PhD student discounts and the 10% discount on all Structured Programs. We offer workshops in English and Mandarin, and you can see all workshops currently available at instats.com.au/workshops

  • @k_o_r_a_y
    @k_o_r_a_y 4 ปีที่แล้ว +2

    Maybe it would be safer to say, what does not remain is unlikely to be causal.

    • @michaelzyphur2476
      @michaelzyphur2476  2 ปีที่แล้ว

      I wanted to get in touch about the Institute for Statistical and Data Science (Instats), which is now live and providing a range of statistics seminars on a variety of topics at instats.com.au . For our initial June launch, we're giving a 50% discount on all of the Mplus seminars that I'm delivering, which is in addition to the substantial PhD student discounts and the 10% discount on all Structured Programs. We offer workshops in English and Mandarin, and you can see all workshops currently available at instats.com.au/workshops