Good morning all,
I post here everything that has been exchanged by email on this subject to centralize our action.
Dated : Thursday, February 10, 2022 10:12 a.m.
Of : Stephane
I allow myself to come back to you in order to prepare a little in advance our next videoconference (March 17) for our webinar (June 6) on the proposal for bivariate modeling assistance in the form of a Check -List.
I offer you a first draft of this Check-List in the attached Excel file (see file 1) (Tab #2 titled “Bivariate Modeling”. Disregard tab #1 titled "Data Preparation")
Your inputs are welcome! Don't hesitate to add steps, correct nonsense, add comments (PS: Please indicate your initials), or even columns (eg: "JMP platform to use"). FYI, I only focused on the case where Y and X are continuous variables, did you have in mind to also deal with the other cases?
Don't worry about the form, it's only a working document. We will see this point later.
If we could reach a consensus on the content before our videoconference, that would be perfect!
We could therefore focus on:
- The definitive content review
- Choice of broadcast format
- Webinar organization
- Definition of the scenario/progress
- Choice of examples
- Choice of presenters
Dated : Friday, February 11, 2022 3:01 PM
Of :Victor
Thank you for this great comprehensive work!
Indeed, depending on the use of the "Fit Y by X" or "Fit Model" platform, the number of options is not the same...
I think it can also be interesting to separate the two, the options or tests mentioned in the table not necessarily being possible on both platforms, so it all depends on the dataset and the intended purpose.
- Start with an exploratory approach with “Fit Y by X” (and other platforms like “Distributions”, “Screening>Explore Outliers”, “Multivariate”,… )
- Approach a modeling part with "Fit Model" and the available tests that you mention.
Some additions from me (some perhaps rather in part 2 of the Modeling topic?):
- For the R² coefficient, there is not really a necessary minimum, everything will depend on the context: if we are on yield/production data with a process based on factors, we will seek to model the system as well as possible, so an R² of less than 0.6 – 0.7 for example will be quite disappointing: there are potentially other influential factors not taken into account… For sensory/preference/taste studies, managing to obtain an R² of 0.5 to explain people's preferences according to different attributes is already very instructive (personal example on JMP Public, an analysis of beers which explains 45% of Someone's preferences is quite interesting: ).
- For the lack of fit/overfitting (and the presence of an outlier distorting the analysis), I really like to use the R² PRESS: calculated by averaging R² calculated with n-1 points each time , this can help detect a problem of generalization of the model if R² and R² Press are very different. Thanks Damien for the tip
- For the residuals, displaying the “Studentized Residuals” is also quite instructive, and allows you to identify a problem at a glance.
- Regarding outliers, the outlier screening platform is very good, but in the case of a problem with a large number of dimensions I like to do an outlier analysis by the “Multivariate” platform -> Jackknife distances .
On the "Fit Y by X" platform, there is also a lot to say in the case of categorical data: ANOVA (with comparison of means, equivalence test of means and standard deviations/variance, etc.), analysis of contingency…
Finally, I am attaching two cheatsheets that I had sent to Florence, and which are quite practical for carrying out statistical tests: smiling_face_with_laughing_eyes: (see files 2 and 3)
Dated : Friday, February 11, 2022 5:18 PM
Of : Andrew
Thank you Stéphane and Victor for this work and your thoughts.
The proposed table is very good. I believe that we can bring more by indicating in the column means to implement the command, the associated JMP option and by relying on the help to deepen to distinguish itself from already existing files (see file 4) (black belts, Olivier's method sheets, etc.).Our cheat sheet should be for using JMP. I added an F column with the JMP menu (see file 5) . This column will be our help to have the right graph, the right value, the right p-value to interpret our data.
As an example, I took the list of options from the Fit Least squares platform (see file 6) . An option can be associated with 1 point to be checked. You can see my preferences below.
Then, I am very careful about the threshold levels for R² and other correlation coefficients. These are random variables and you have to do a test to be sure of their significance. One can be satisfied with a weak but significant linear correlation coefficient (idem for R²).
Perhaps the lines to be worked on should be divided up in order to move forward. What do you think ?
Thanks again for starting a thread. This makes me think that the time I had planned for the slide on our committee has been used for something else!
Dated : Friday, February 17, 2022 2:55 p.m.
Of : Anna
Thanks for all the suggestions on the subject!
I think we risk quickly filling (or even overflowing!) the session: smiling_face_with_laughing_eyes:
I wonder if we could not focus on the Fit Y by X platform in this first session and leave the Fit model for the second session on modeling scheduled for July.
In my opinion, it could be good to deal with the different scenarios depending on whether the variables are quantitative or qualitative.
And actually stay on something very practical oriented JMP, to know where to look and find the corresponding information at each step.
Dated : Friday, February 23, 2022 1:28 p.m.
Of :Victor
To follow up on the various comments on this first modeling part, and the need to focus on a reduced number of analyzes (and platforms), I offer you a small summary of the possible analyzes and statistical tests on univariate and bivariate distributions in JMP, using the Distributions and Fit Y by X platforms (see file 7) .
I don't think there's any need to see everything (in a one-hour webinar it might be short…), but we could pick up application cases on some of these analyses, and create a flowchart for the cheatsheet which allows users to navigate to the right platform according to the data they have and their goals (modelling, statistical test/comparison, visualization, etc.).
I would be of the opinion to push some examples of statistical tests in this webinar, because they are important in exploratory analysis to better understand its dataset and are also found in more complex platforms/analysis (analysis of experimental designs, MSA , …), so being able to understand how they work on a simple example can make it possible to better understand the logic behind more complex analyses: experimental plans (and the hypotheses involved) or measurement system analyzes (and the analysis variances depending on equipment, operators, etc.)… :smiling_face_with_laughing_eyes:
Dated : Friday, February 23, 2022 2:22 p.m.
Of : Stephane
I was just going to send you what I had been working on for a few days for our group.
I wanted to test the Check-List prototype I sent you on a (made-up) data case. (see files 8 and 9) . This gives an idea in addition to the frame that can be unrolled during the Webinar.
It only deals with the case of bivariate modeling (I admit that's more what I'm used to) using:
- Not only, as Anne recommended, the Fit Y by X platform (which I rediscovered for the occasion: smiley_face_with_laughing_eyes: )
- But also the Fit Model platform (to which I actually go instinctively every time)
Reading Victor's email and document, I realize that there is so much to say that we will have to... split the Webinar or make drastic choices!
(André, Florence, that will teach you how to put a Victor and a Stéphane together: face_qui_tire_la_langue_et_plisse_les_yeux: )
Dated : Friday, February 23, 2022 4:36 PM
Of :Victor
Thank you for the slides, it's an interesting step-by-step approach to solving a problem.
For my part, the Excel file does not help too much to represent the process and I was not necessarily very clear, but I wanted to accentuate the cheatsheet side according to the data, distributions, and what we want to do (see example in PJ) (see file 10) in JMP like the flowcharts that I had sent you previously on the statistical tests (degree of precision of the flowchart to be considered so as not to make it too busy…).
The two approaches are quite complementary I think: Stéphane shows what can be done and the different precautions to take when analyzing a dataset, while the cheatsheet presents a "range of possibilities" from which to draw according to needs: smiling_face_with_laughing_eyes:
Regarding the documents and following the email from Florence, perhaps we can upload all our documents (Excel, ppt, etc.) on our community space, in order to have them in the same place (and not through several emails)?
- To see where we can put them (we create a Knowledge base article grouping all the current work files, and the following will come in response to the initial post?)?
As Stéphane points out, it can be dangerous to associate a Stéphane and a Victor together: face_qui_tire_la_langue_et_plisse_les_yeux:
Dated : Friday, February 23, 2022 4:53 p.m.
Of : Florence
What work !!! well done !!!
I think the basics of the webinar are really there!
I would just have a remark: Let's not forget to stay on the fact that JMP is simple! so I think it's going to have to be simplified. And above, I join you Victor. Your excel file is super interesting but too dense. It should be offered in a more hierarchical structure, like the PPT, but more visible, because the PPT is also dense.
I think we can create a post in the mentioned group. And suddenly, reacting at this level, commenting etc. We will see with use if it is easier or not in the end! I let the leader of the modeling group create the first post!
See you on March 17 to put all this content in place and, big challenge, get it back in less than an hour!!!!
Report of our preparation Webinar.
Date : 03/17
Attendees : Florence, Anne, Isabelle, André, Victor, Thomas, Stéphane
Organization of the June 9 Webinar
- Introductory presentation of a flowchart to direct users to use cases according to the nature of their data. Make this part attractive and informative to capture the attention of the audience
- Presentation N°1: Case of categorical data (ANOVA) (20 min), Victor
- Presentation N°2: Case of continuous data (FitXbyY, FitModel) (20 min), Stéphane
- Presentation N°3: Regulatory framework applicable to the pharmaceutical field, Thomas
Recommendations
- Keep it simple! Make presentations accessible and attractive... put some glitter!!
- Think of an attractive title for each part of the Webinar
- Orient the "business", "industrial concept" presentations, while keeping in mind that any user in the audience must feel concerned
- Be careful not to insist too much on the transformations of variables which may scare inexperienced users
Next meeting
April 13, 10-11 a.m.
Feel free to comment on this post if I forgot something (especially Thomas, I'm not really comfortable with what was said about your part)
