That's an interesting question. We first aggregate the points that are closest - so that would be a and b. You then say a and c are distant. Relative to what? We really need a point d do decide that, I think. Then it's a question of whether c is closer to a&b or closer to d. But there are a couple of further details that we need to think about, and that weren't discussed in this video (I hope I will address these in some new year videos!). First, when we ask about distance of c from a&b what do we mean? We have to define distance more carefully. In the video I used the average location (think of a centre of mass). So we are not comparing the distance of c to a, but the distance of c to the midpoint of a and b. There are alternative distance metrics that we can use - we could take the closest point in a cluster, in which case we would look at the distance of c to b; or we could take the furthest point in a cluster, in which case we would look at the distance of c to a. Different distance metrics allow us to describe different shapes of clusters (think of spherical clusters versus elongated string-like clusters). In all these cases we would compare these distances against the distance from c to d. Any finally, perhaps we want to stop the agglomeration process and keep d and or c separate from a&b because we think that we have reached the optimal number of clusters. To do that we need to define what we mean by "optimal" - but a rule of thumb is to look for the point where we transition from small distances to large distances - and this is something I will discuss in my next video (probably early January because I am decorating the room that I use for recordings!).

You're welcome :)

You're welcome!

I was using JMP Software.

@PegaAnalytics thank you so much for your reply. Is this software available in online?

It's commercial software, produced originally by SAS: www.jmp.com/

Sorry that it wasn't helpful for you.

@PegaAnalytics it was my short comings not your.

Well I can see that you also produce educational videos. One of the joys of teaching is to try and come up with new ways of explaining concepts. Different explanations work well for different people - ourselves included!

Congratulations on getting to the last part of the series. I'm packaged the videos and the associated code into a Udemy course. You'll need to enrol for the course the get access to the files, but it doesn't cost you anything. Plus also, it should give you direct access to me if you have any future JSL questions. Here's the link: www.udemy.com/course/jsl-decoded/

For me personally, I do a lot of process capability analysis, which is very sensitive to the assumption of normality of the data. For many other statistical methods (regression, two-sample t-test, etc) we construct a t-test and derive a p-value; these statistical methods assume that our error terms (residuals) are normally distributed.

This you so much for taking the time to leave this comment, and for your kind words. This type of feedback gives me huge motivation to make further videos.

Thanks for your positive feedback. My next video will discuss the optimal number of clusters

Thank you!

That's great to hear. Thanks!

Glad it was helpful!

Hi Mark, thanks for your contribution. I totally agree, these are two perfect use-cases for journals. It's interesting that you export a problem-solving journal to MS Word, once you have completed the work. I think that's probably a very helpful step. Something I will look try-out myself!

@@PegaAnalytics Why MS Word? Journal Text Sledgehammer add-in is an improvement, but still nowhere near the features/ease of use of MS Word. Also, customers may not have JMP. MS SharePoint can't index JMP files, but can index Word files.

Funny you should ask that ... the following paper is next on my reading list: "Pulsed-field gel electrophoresis (PFGE) analysis of Listeria monocytogenes isolates from different sources and geographical origins and representative of the twelve serovars" www.academia.edu/111312006/Pulsed_field_gel_electrophoresis_PFGE_analysis_of_Listeria_monocytogenes_isolates_from_different_sources_and_geographical_origins_and_representative_of_the_twelve_serovars

I've looked at this in a bit more detail, and to be honest, handling these type of data is beyond my area of expertise. I did find some general information that I found helpful: A guide to interpreting electrophoresis gels: bento.bio/resources/bento-lab-advice/interpreting-electrophoresis-gels-with-bento-lab/#:~:text=The%20smallest%20bands%20are%20at,is%20up%20the%20ladder%20scale. (pulsed-field addressed larger DNA molecules but I presume the principles on interpretation remain the same). Any analytical technique requires digital data. I found this: Data processing of pulsed-field gel electrophoresis images www.ncbi.nlm.nih.gov/pmc/articles/PMC6940661/ The data processing would seem to me the critical step, which will ultimately result in the generation of tabulated data that would be amenable to cluster analysis. The columns of this tabulation would correspond to metrics that describe the banding, which each sample being represented by a row. I would guess that this data processing is integrated into most laboratory systems that produce pulsed-field electrophoresis gel?

Happy to hear the video helped :)

Thank you so much!

I appreciate your feedback. I'll make a note to make a video about identifying the optimal number of clusters - thanks for the suggestion.

Thanks!

Glad you liked it!

Thanks Mark, I always value your feedback.

Thank you so much, that is very generous of you!

Hi, thanks for your feedback. I can't think of a good reason to not use the 'fit separately' option. Perhaps I might use it if I am doing a demonstration and want to focus on model visualisation and interpretation without dwelling on the details of model building. But in a real-world scenario, I would always want to spend my time building the best model I can for each response, and for that I would use 'fit separately'. The 'fit separately' option allows you to refine model terms individually for each response, whereas if not selected you refine model terms en-masse across the responses. Prior to version 14 this was the only way of handling multiple responses - apart from running Fit Model separately for each model. And the benefit of specifying multiple responses rather than running multiple Fit Models, is that you can view all the models simultaneously in the prediction profiler and contour profiler, as well as perform numerical optimisations that take account of all the responses.

You should be able to access it here: www.pega-analytics.co.uk/blog/how-to-write-a-jsl-script/

@@PegaAnalytics Thank you!

Glad it was helpful!

You are welcome!

I had the preference "Ppk Capability Labeling" turned off. This is one of the preferences for the Distribution Platform( File> Preferences> Platforms> Distribution ).

@@PegaAnalytics Thank you

Thanks for your questions. The median value of 84 is based on fitting a Lognormal distribution to the data then taking the median of fitted distribution. That median calculation can either be done using a simple script based on the shape and scale parameters for the distribution (see my video on using the script editor as a calculator) or directly from JMP (after fitting a distribution, from the associated red triangle select Profilers> Quantile Profiler, and enter 0.5 for the quantile). Yes, you are correct, the value of 5.8 is specific to these data. Whereas for a Normal distribution we can always say that the width is +/- 3 sigma, that is not true for other distribution types. The value of 5.8 was determined by calculating the interval from the 50% percentile (the median) to the (100-0.135)% percentile. Similar to the median calculation, this can be done with a simple JSL script or by getting the numbers from the Quantile Profiler for the fitted distribution.

Thank you@@PegaAnalytics

Thanks for your feedback Mark.

I'm glad you liked it. I appreciate the feedback. Thanks!

Thanks!

Thanks for your positive feedback :)

Glad you liked it

I'm glad you are enjoying the videos. It's really helpful to get requests, so I know what is of interest to people. I'll add the app builder to my list of topics!

@@PegaAnalytics Thanks!

That's wonderful. Thanks Mark.

Glad it was helpful!

Thanks Steffen. I appreciate the encouragement!

Thanks for the feedback. Part 2 tomorrow!

That's a good point. My version of the table was slightly different to the one that you can find in the sample data. On the left of the table, where it lists all the column names, you can select multiple column names and then right-click: you will then see an option to "group columns". JMP will assign a default name to the group e.g. "Mill Time etc.", but you can click on the name and edit it to be something more descriptive. ( You will also find differences in the order of the rows. For illustration purposes I wanted a condition where a control chart had recently gone out of control. To achieve this I think I sorted the data by descending order of mill time. )

Thanks! I really appreciate your feedback.

RSM CCD design and its analysis to predict the response.