Background: I have a very large data set that is an output of specific RF testing for semi-conductor devices. During this RF test there are a variety of differnt conditions that are important. I have 3 outputs that are tested on 16 different channels at 6 different test states which require me to generate 105 differnt graphs.
example: Gain vs Frequency on Channel 1 at Test State 1; Gain vs Frequency on Channel 2 at Test State 1 where i have 15 channels and 6 test states.
I have searched these discussion boards for a solution and the closets i have come is a thread started in 2017 titled "
Summarize( dt, lstChannel = By( :channel ) );
gb = Graph Builder(
invisible,
Size( 1225, 952 ),
Show Control Panel( 0 ),
Variables(
X( :Name( "Freq _(GHz)" ) ),
Y( :Gain ),
Group X( :wafer_num ),
Overlay( :RowCol ) ),
Elements( Smoother( X, Y, Legend( 19 ) ) )
);
For( i = 1, i <= N Items( GroupBy ), i++,
gb <<(
columns(
:wafer_num,
:channel,
:RowCol),
Where( :channel == LstChannel[i] ) ),
// I would like to loop through the teststates as well
Where( :teststate == 1 ),
Mode( Show( 1 ), Include( 1 ) ) );
If( i == 1,
gb <<( "$TEMP/jmp_example2.pptx" ),
gb <<( "$TEMP/jmp_example2.pptx", Append )
);
gb <<;
);
Open( "$TEMP/jmp_example2.pptx" );
Yes, well your skills are stellar and I would take a JSL class from you any day..
Quick question though.. I ran it with the gd_inc.jsl and this is what i get.. I can see that the scale is almost perfect but my 15 data sets have these 2 difference. Not to say that this isn't going to happen in the future.. When you say modify your script with that second one you quoted i get confused.. I am trying to revere into and figure out where the 10.5 and 42.0 is derived from and think it has something to do with the log() function but not sure..
Yng,
You did not specify what you are looking for. I typically use JMP default scaling, and use something like the following if I want the same axis scaling for all subgroups. This script names the variables of interest ylist, but this can be applied to any numeric continuous varaible (column), even Datetime. The trick is to let JMP scale for all, then copy the axis setting to the table column, then any graph using that column will deploy that scale.
Names Default to Here(1);
ylist = {"Sepal length", "Sepal width", "Petal length", "Petal width"};
dt = Open("$sample_data/iris.jmp");
//get common axis
for(i=1, i<=nitems(ylist), i++,
ovl = dt <<(Invisible,
Y( Column(dt,ylist[i]) ), //yaxis
Overlay( 0 )
);
report(ovl)[AxisBox(1)] <<;
ovl <<;
);
//now the same axis will be applied to all sub groups
ovl1 = dt <<( Y( :Petal length ), Overlay(0), By(:Species) );
//all 3 subgroups have the same Y - axis
//ovl = dt << OverlayPlot( Y( eval(ylist) ), Overlay(0), By(:Species) );
I do custom scaling for axes only if:
- there is a known range like control or spec limits,
- need to scale to something like +/- k sigma,
- need to control the grid lines, for example, a grid line at each 0.5 sigma,
- data whose nominal value is large but variation is very small, for example, means value is 1750222 stdev is 0.000023, then I plot a transform Y- mean and figure out a good scale and use the alogorithm, gdinc.jsl
- perform an analyses and determine whether the data is highly skewed or lumpy (clustered).
Regarding outlier detection, I have not found one method that applies to all data. For univariate data with a strong right skewed data (ex. zero inflated Poisson), I like to use a modified Lenth method:
- uscreen = median + k * (upper p percentile -median) / Abs( Normal Quantile(p) )
- lscreen = median - k * (median - lower(p) percentile/ Abs(Normal Quantile(p) )
- where p could be 0.1 ( for 10 and 90 percentiles) or p could be 0.25, then this would be a modified pseudo sigma, and typically k is in [4,6] for summarized data. There are times for raw data, I use a much larger k.
- Also, I typically screen for all data of one variable, but if by groups are quite different, I I screen by groups if there is enough data fopr each group.
Below is an example for computing the screening values by group. If vargroup has 10 groups, the uscreen and lscreen are vectors of size 10 and vg is a list of 10 vargroup names.
//Find global outliers by vargroup for raw measurement data
//Use an assymetric 6 pseudo-sigma, computed from the 10 & 90th percentiles.
summarize( vg=by(vargroup),
c=count,
q10=quantile(varvalue,.10),
q50=quantile(varvalue,.50),
q90=quantile(varvalue,.90)
);
uscr = q50 + 6*(q90-q50)/Normal quantile(0.9);
lscr = q50 - 6*(q50-q10)/Normal quantile(0.9);Please note:
- There is no one correct method, and you should always review the methods with those that are reviewing the data. And outliers shoudl never be ignored, because they are leads to possibly important information.
- I often use 2 levels of screening, especially when I needed to compute control limits, screening for the plotted points and screening for limits ... control limits shoudl not accommodate 5 and 6 sigma data.
- If you are creating weekly analyses, note you can get different screening values each week. You might want to do a mega analysis once an use the same limits each week.
- For multivariate data, you should use a mutivariate screen. See graphic. Two obvious outliers are in control based upon univariate limits. So if data is strongly correlated a correlated screen shoudl be used. JMP multivariate has a nice multivariate screen.
