I am considering purchasing the statistics toolbox to help plan a set
of simulations in which several parameters are varied, with the goal
of identifying which parameters give the biggest effect on some chosen
variables.  The parameters aren't independant: When several parameters
are increased, the resulting effect may be larger than the cumulative
effect of increasing each individual variable.  A colleague said that
Design of Experiments (DoE) theory can be useful for this sort of
problem (I have no background however), and I noticed that the
Statistics Toolbox contains some DoE routines.  I need to decide
whether to purchase the toolbox to access these routines.

If you use these DoE routines, do you find them useful, or do you
prefer your own code?

On 27 Jul, 18:31, ke <kateedward...@hotmail.com> wrote:
> I am considering purchasing the statistics toolbox to help plan a set
> of simulations in which several parameters are varied, with the goal
> of identifying which parameters give the biggest effect on some chosen
> variables. =A0The parameters aren't independant: When several parameters
> are increased, the resulting effect may be larger than the cumulative
> effect of increasing each individual variable. =A0A colleague said that
> Design of Experiments (DoE) theory can be useful for this sort of
> problem (I have no background however), and I noticed that the
> Statistics Toolbox contains some DoE routines. =A0I need to decide
> whether to purchase the toolbox to access these routines.
>
> If you use these DoE routines, do you find them useful, or do you
> prefer your own code?

Using a toolbox to design an experiment is like using a hammer
to design a building - plain nonsense.

You can use the hammer to *implement* the design - build the
house - but it is up to an architect or similar to come up with
a plan of what to build.

Designing the experiment is *your* responsibility.

Rune

On 7/27/2010 1:01 PM, Rune Allnor wrote:
> Designing the experiment is *your* responsibility.

Rune, I imagine that ke was referring to tools for creating things like 
D-optimal designs, or doing power analysis, or making effects plots. 
While these are just tools to implement or analyze a chosen design, they 
are what is traditionally known as (statistical) "Design of 
Experiments", and the Statistics Toolbox has a variety of functions 
available:

<http://www.mathworks.com/access/helpdesk/help/toolbox/stats/bq_w_hm.html#brj0q33-1>

I will leave it up to others to answer ke's question, which was, "Do you 
find these useful?"  If the answer is no, I'd be interested in hearing more.

- Peter Perkins
   The MathWorks, Inc.

Peter Perkins <Peter.Perkins@MathRemoveThisWorks.com> wrote in message <i2pc5l$47l$1@fred.mathworks.com>...
> On 7/27/2010 1:01 PM, Rune Allnor wrote:
> > Designing the experiment is *your* responsibility.
> 
> Rune, I imagine that ke was referring to tools for creating things like 
> D-optimal designs, or doing power analysis, or making effects plots. 
> While these are just tools to implement or analyze a chosen design, they 
> are what is traditionally known as (statistical) "Design of 
> Experiments", and the Statistics Toolbox has a variety of functions 
> available:
> 
> <http://www.mathworks.com/access/helpdesk/help/toolbox/stats/bq_w_hm.html#brj0q33-1>
> 
> I will leave it up to others to answer ke's question, which was, "Do you 
> find these useful?"  If the answer is no, I'd be interested in hearing more.
> 
> - Peter Perkins
>    The MathWorks, Inc.

Yes, exactly what I had in mind.  (Rune, I also had never heard of formal 'Design of Experiments' before my mechanical engineering colleague mentioned it).
If anyone uses these tools in the Statistical Toolbox, I would like to know they have been helpful.

Ke

I work in an industry were DOE's are a regular practice.  The majority of my peers use Minitab since it is very simple and if you are familar with excel then this may be an option.

As Rune points out (maybe too much) that users simply looking for shortcuts are asking for trouble, in this case I agree.

I have both Minitab and the Statistics Toolbox and the DOE's in the toolbox can be used to help set up experiments.  The learning curve for implementing, designing, analyzing and refining experiments using Matlab is a whole lot steeper but far more comprehensive.  

Setting up experiments in Minitab, filling in the data and analyzing can take a day to learn but I personally have seen the analysis along with the use of the output commonly misused.  Mostly because Minitab limits the analysis to simpler models (linear in the fitting coefficients) and users typically have no idea about the calculations behind the analysis.  Alot of times, I see users fitting data that is poorly conditioned however, Minitab will force it out.  An error message or notice may be displayed but users constantly ignore and typically "forget" to report those facts.  

If you have factors with more then 2 levels (ie 4 temperatures, 3 pH values), both packages limit DOE designs to full factorial.  The Fractional Factorial designs are limited to factors with 2 levels only.  So fractional factorials with multiple factors and multiple levels are not built in.  I was told that Matlab has a fancier toolbox with this feature but that was out of my price range.

Your limited explanation suggests that your factors influence each other, which both packages can handle (see covariance).  But as Rune points out, implementing designs and then analysis of the data can be very tricky.  In Minitab, the behind the scenes calculations are not readily available so I have seen a lot of analysis that simply was wrong by inexperienced users simply not understanding the covariance feature and what it was doing in the calculation.

In Matlab, all of this is possible and you can even tailor it further beyond anything Minitab can offer but that takes a very thorough understanding of what you are doing and the math behind the theory.  There is no place in Matlab, that I have found, where you build your DOE, enter in your data and basically one click analyze the data, spitting out which factors are and are not relevant.  There is more heavy lifting in Matlab to get that information but once you figure all of that out, Rune may agree that you will be a better person for it....

The strengths of Matlab, being able to work with matrices and vectors makes fitting data very easy but knowing what steps to take beyond the math can be tricky.

I am not sure the $$ for Minitab but if you are a student then Matlab is a no brainer for the price because the Statistics toolbox comes with it.  If you already have Matlab then the statistics toolbox add on is probably similar in $$ to Minitab.

I personally use Matlab for my work but I enjoy reading and learning this stuff.  I find that my peers at work use Minitab only if they have too.

One (of many) books that will help is "Statistics for Experimenters: An Introduction to Design, Data Analysis, and Model Building"  by George E. P. Box , William G. Hunter, J. Stuart Hunter.  Good overview.

Hope this helps a little.  

On 14 Aug, 06:35, "Marc " <marc.schre...@uop.com> wrote:
> Ke
>
> I work in an industry were DOE's are a regular practice. =A0The majority =
of my peers use Minitab since it is very simple and if you are familar with=
 excel then this may be an option.
>
> AsRunepoints out (maybe too much) that users simply looking for shortcuts=
 are asking for trouble, in this case I agree.
>
> I have both Minitab and the Statistics Toolbox and the DOE's in the toolb=
ox can be used to help set up experiments. =A0The learning curve for implem=
enting, designing, analyzing and refining experiments using Matlab is a who=
le lot steeper but far more comprehensive. =A0

Mark,

After having used matlab for a couple of decades, I am now starting
to work alongside people who use minitab as their main tool. Leaving
semantic quibbles aside, how would you compare the two toolsets,
matlab + stat toolbox vs minitab, for statistical process control
tasks?

My main concern right now (haven's seen minitab up running yet) is
that minitab might not have any scripting / macro capacities, meaning
that a lot of time is wasted selecting menu options and / or entering
data names during repeated analyses.

Rune

On 28 Jul, 15:43, Peter Perkins
<Peter.Perk...@MathRemoveThisWorks.com> wrote:
> On 7/27/2010 1:01 PM,RuneAllnor wrote:
>
> > Designing the experiment is *your* responsibility.
>
> Rune, I imagine that ke was referring to tools for creating things like
> D-optimal designs, or doing power analysis, or making effects plots.
> While these are just tools to implement or analyze a chosen design, they
> are what is traditionally known as (statistical) "Design of
> Experiments",

Well, I tend to rather conservative what language is concerned:
I prefer to use terms and words that actually describe what
is going on. Humpty Dumpty is an example to be avoided, not
emulated.

Rune

Rune Allnor <allnor@tele.ntnu.no> wrote in message <46fe4481-5218-4e7f-aeb6-e067b9c355ab@u26g2000yqu.googlegroups.com>...
> On 14 Aug, 06:35, "Marc " <marc.schre...@uop.com> wrote:
> > Ke
> >
> > I work in an industry were DOE's are a regular practice.  The majority of my peers use Minitab since it is very simple and if you are familar with excel then this may be an option.
> >
> > AsRunepoints out (maybe too much) that users simply looking for shortcuts are asking for trouble, in this case I agree.
> >
> > I have both Minitab and the Statistics Toolbox and the DOE's in the toolbox can be used to help set up experiments.  The learning curve for implementing, designing, analyzing and refining experiments using Matlab is a whole lot steeper but far more comprehensive.  
> 
> Mark,
> 
> After having used matlab for a couple of decades, I am now starting
> to work alongside people who use minitab as their main tool. Leaving
> semantic quibbles aside, how would you compare the two toolsets,
> matlab + stat toolbox vs minitab, for statistical process control
> tasks?
> 
> My main concern right now (haven's seen minitab up running yet) is
> that minitab might not have any scripting / macro capacities, meaning
> that a lot of time is wasted selecting menu options and / or entering
> data names during repeated analyses.
> 
> Rune

Rune

One of the things I run into with SPC is that if the data is not normal than people tend to misuse CpK and PpK because Minitab likes to use average moving range for CpK if you do not have historical data (ie. your historical data is the data you are fitting).  Obviously with MATLAB the ability to fit data to multiple PDFs or user created PDFs with MLE function makes my arguments for data analysis hard to argue with.

I was recently dealing with data that has a hard lower spec limit but the upper spec limit is more for keeping costs in line and making sure our manufacturing site doesn't pile on the raw material.  

Since historically this has been poorly controlled, we end up with a set of data that appears to be bi-modal.  The MATLAB demo in the statistics toolbox for fitting user created PDFs essentially walks one through this problem.

Long and short, I do not believe there is anyway to handle this type of data in Minitab.  

Also I have become fond of ksdensity as it turns historical data into x,y data which now I can use all of the tools in MATLAB to try and fit.  

For other SPC tools, control charting in MATLAB is a little weak and those who are used to Excel tend to gravitate towards Minitab.  Minitab I think is limited to 99 plots per project.  Which surprisingly enough, I have run into that error.

As for dealing with the tabs, repeating plots can be an issue in Minitab if you have multiple worksheets, as the tabs work on active worksheets.  So if you modify your data to new columns or new worksheets, figuring this out can be time consuming.

Overall, I like Matlab but I like to know how things work.   Unfortunately, I find I am in the minority most of the time.

Sorry for the late reply...

Rune

As I forgot to mention... If you have any other questions, feel free to contact me.

Thanks

Marc