Preselected Roles in Graph Builder?

[#1]

It appears that if you preselect a role (e.g. weight or frequency) for a column that Tabulate or Table Summary use that role in the analysis, but Graph Builder does not.  Can someone confirm that Graph Builder does not recognize pre-selected roles?  And, if that is correct, should we consider that a glitch or is there a good reason for that? 

On a somewhat separate issue, I still find the choice of weight or frequency for a variable confusing - perhaps unnecessarily so.  In many computations, both designations provide the same results, but in many analyses they differ.  In particular, in simple linear regressions either weight or frequency seem to provide the same linear model, but Weight appears to give the correct standard error for the regression coefficient but the Root Mean Square Error, while using Frequency has the reverse characteristics.  Is there a reason for that disparity?

As an example, I am looking at airfare data for different domestic air routes.  Each route is an observation and I have data on the average fare and the average number of passengers per day (and the distance (miles) for the route).  For a weighted average fare, using passengers as either weight or frequency gives the same results.  Doing a regression of average fare on distance, using passengers as weight or frequency gives different results - weight provides a correct standard error for the coefficient on distance, but an incorrect root mean square error, while frequency provides a correct root mean square error and an incorrect standard error for the coefficient.  I've attached an example data set with 3 embedded regression models (no role for passengers, passengers as weight, and passengers as frequency).

[#2]

The analysis roles for numerical procedures do not all translate well into the visual procedures. JMP translates the Weight role into the Color role in Graph Builder. I assigned the Y role to :weight, the X role to :height, and the Weight role to :age. I launch Graph Builder and it is empty, waiting for me to build my graph interactively. I click the Dialog button at the top and see three things.

First, there is no Weight or Freq analysis role for this platform. Second, JMP translated the pre-assigned Weight role into the Color role. Third, I now see all the roles that are relevant to Graph Builder.

So the pre-assigned roles do not help much with this platform. They pre-date Graph Builder, and again, they do not always translate, so JMP gracefully ignores them or re-assigns them.

[#3]

I thought it might be the case that Graph Builder did not recognize the pre-assigned roles.  However, putting passengers (in my example file) as color does not treat it as a weight.  If you look at the least squares fit in Graph Builder, then it is the unweighted regression.  So, it appears that Graph Builder does not have the capacity to automatically deal with weights.

Can you address my other question about the inconsistency between using a variable as a weight or a frequency?

[#4]

The confusion between a weight and a frequency is common, and in some instances, it does not make a difference. But to be sure that it is what you want, let me explain.

Frequency is intended to help when you have a data table that is summarized. JMP expects each observation to be placed on a separate row. But some times you have each kind of observation in a row, and then a count of each kind of observation. That is the intended purpose of the frequency role.

On the other hand, weights are just that. A continuous measure of importance or uncertainty. An example is regression. An important assumption of regression is constant variance. If the variance is not constant, then one approach is to weight each observation by the reciprocal variance. It does not serve as a count of observations. It is a more granular purpose.

I hope this explanation helps.

[#5]

I also have a question about your screenshot.  I don't get a window like you show when I launch Graph Builder.  In fact, yours shows a box for Freq, and option I do not get at all.  And mine has the drop zones which your screenshot does not show.  Where do I get the screen you are showing?

[#6 · Accepted solution]

Click the Dialog button at the top left of Graph Builder.

[#7]

Thank you.  I was looking at the red arrow and completely missed the button labeled "Dialog."  Indeed it gives the option of a Freq variable and works fine.  I understand your explanation of the difference between Freq and Weight.  To paraphrase, I'd say the former is more a function of the data while the latter is a function of the analysis you are doing.  I work a lot with survey data where the sampling procedure gives weights that show how many households of a particular type each responding household represents.  Thus, the data is itself like a summary table, and frequency would seem to be appropriate.  When doing a weighted regression, the weights are not really a property of the data but a property of the analysis technique, and weight is appropriate.

The thing that still seems confusing is the difference in the standard errors.  Using frequency gives the right root mean square error, but the wrong (order of magnitude too small) standard error for the regression slope.  I imagine this is due to JMP thinking the sample size is much larger than it really is.  But, if I use weight, then the standard error of the regression slope is correct, but the root mean square error is much too large.  I don't really understand this difference.