DOE Options¶
These settings change how designs are built and what a regression reports. They are remembered between sessions and apply to every workbook, so a setting changed here changes the next design and the next regression you run.
QXL DOE New > Options
The dialog has four pages, listed down the left as Create Design, D-Optimal, Regression and Charts.
The four buttons¶
| Button | What it does |
|---|---|
| Save settings | writes every page's settings. It starts greyed out and only becomes available once you change something. Nothing you change takes effect until you press it |
| Default settings | restores the defaults on all four pages at once, not just the page you are looking at |
| Cancel | closes the dialog. With unsaved changes it asks first: Settings are not saved. Do you want to close the form without saving changes? |
| Help | opens this topic |
If a setting cannot be written, the dialog reports Some settings were not saved.
Create Design Settings¶
Variations when creating new design. Sets the order the runs come out in for a newly created design.
| Setting | Pattern |
|---|---|
| Last factor changes first. | Second line: Example: (A = -1, -1, 1, 1; B = -1, 1, -1, 1) |
| First factor changes first. | Second line: Example: (A = -1, 1, -1, 1; B = -1, -1, 1, 1) |
Both give the same set of runs and the same design. Only the order of the rows differs.
Run order for new design, with Standard and Randomized.
Standard writes the runs in the systematic order above. Randomized shuffles them. Both produce the same set of runs.
This setting is what the Create Design builder uses. The Design Wizard asks per design instead, on its own Do you want to randomize runs? page, and writes the answer here only if Use this setting always is ticked.
You can also randomize an existing design later, from Modify Design.
D-Optimal settings¶
Both check boxes on this page ship ticked, and the seconds box ships at 1.
Run the modified Federov at end to get the last bit of improvement. This one works. It leaves the faster search to do the bulk of the work and then makes one thorough pass over every row at the end.
Two of the three controls on this page do nothing. Their values are stored and never used, so setting them changes no result:
- Force the Federov algorithm. Federov gives a better answer but takes way longer. The search behaves the same whether this is ticked or not.
- Number of seconds to look for initial design before giving up (default is 1 sec). The number is saved and never read.
Both are recorded in code-flags-for-engineering.md, since a stored setting with no effect is a product
issue rather than a documentation one.
See Create D-Optimal for what these settings are searching for.
Regression settings¶
Display DOE Advisor on regression sheet. The DOE Advisor writes plain-language findings onto the regression sheet: things it noticed about the residuals, the model or the design. It also marks runs red on the design sheet, either for a large standardized residual or for high leverage, which is the one Advisor action that changes a sheet other than the regression sheet. See The DOE Advisor for everything it writes, and Understanding Least Square Residuals for the thresholds it uses.
Display comments in the regression sheet. Attaches a cell comment to each statistic on the regression sheet explaining what it is and how it was calculated. Useful while learning the report, and worth turning off once the report is familiar or before sending the workbook to someone who will find the comment markers distracting.
Show P-Value for the uncoded regression. Off by default, so an uncoded regression leaves its P column empty until you turn this on. The report says why in its heading block, in red, on the line under Regression in uncoded units, not in the empty cells themselves.
Custom designs are exempt. An uncoded regression on a custom design prints its p values whatever this setting says, and shows no notice. See Uncoded Coefficients for why the default is off.
Use weighted S-Hat regression. The setting is saved and changes no result. With it on, Quantum XL prepares one weight per row for the standard deviation model, by \(n - 1\) where \(n\) is that row's replicate count, so that a standard deviation from more replicates would count for more. The regression that fits that model is then always run unweighted, so every coefficient and standard error on it is the same whichever way the setting is left.
Two conditions decide whether the setting does anything visible at all:
- The weights are prepared only when the replicate counts differ between design points. On a design where every point has the same number of replicates, turning it on or off does nothing.
- An output with its own weights column bypasses it entirely. That column is used, and this setting is not read.
When the counts differ and the setting is on, the standard deviation half of the regression sheet's heading block carries the line S-hat regression weighted by df, unless that model failed, in which case its error message takes the same spot. That line is written from the setting rather than from the fit, so it does not tell you the fit was weighted.
Recorded in code-flags-for-engineering.md, since a stored setting with no effect is a product question
rather than a documentation one.
Weights on the response model are a different thing and do work. An output with a weights column has its response model fitted by weighted least squares. See Ordinary Least Squares Regression.
Full details: Weighted Least Squares (QR).
Hierarchy warnings. Controls whether Quantum XL objects when a model contains an interaction without the terms below it:
| Setting | What it checks |
|---|---|
| Strict. Check for all interactions. | Every term below the one you included. With ABC in the model it wants A, B, C, AB, AC and BC as well. On a blocked design Strict is downgraded to Normal, so only the main effects are required |
| Normal. Check for main effects only. | Only the main effects. With ABC in the model it wants A, B and C, and says nothing about AB, AC or BC |
| None | No warnings, except nesting, which is checked whatever this is set to |
The warning is not always advisory. Which of the two it is depends on the cause, not on this setting:
- A missing term that involves a nested factor is an error you cannot pass. The message has an OK button only and the operation stops. This happens at None as well, because nesting requirements are built whatever the setting.
- Any other missing term is a Yes/No question, defaulting to No, and answering Yes proceeds.
The check runs when the regression runs, not while you edit the model. See Regression.
Charts settings¶
Line type between two points of categorical factor, with Display dotted line and Do not display line.
This setting currently has no effect
On the QXL DOE New tab the interaction, main effects and thumbnail charts always draw a dotted line between the levels of a categorical factor, whichever option is selected here. The control is left over from an earlier version of the charts.
The dotted line itself is explained on the Interaction Plots page: it marks that no setting exists between two levels, so the line is a guide for the eye rather than a set of reachable predictions.