nav[aria-label="Primary Navigation"] { padding: 0; & ul { list-style: none; width: 100%; display: flex; flex-direction: row; justify-content: start; align-items: start; gap: 30px; padding: 0; & li { margin: 0; } & ul li { list-style: none; } } }

DOE full factorial design with partial results shortage

stevenfoong1

Dear all.

Good day. I am beginner for DOE and recently I've created a full factorial table to run my electroplating test. However during sample preparation I found some parameters have produce bad visual quality which is not justified to measure the responses. Eventually some legs created from JMP tables have no results.

Is it fine to leave it blank if no result available? Can I still able to use "fit model" to analyze this incomplete data set? What is the consequence by doing so?

Or should I remove rows which has no results, and run "fit model"?

Thank you very much.

* Kindly refer to attachment for JMP file

JMP discussion file.jmp

Accepted answers

Mark_Bailey

I noticed some things with your study.

You have many runs, so you should have very how power to begin with. The omission of some runs should not dramatically affect your power.
The model for Metal% exhibits strong lack of fit. Predictions from this model will be biased and likely very inaccurate.
I tried adding quadratic terms to account for the lack of fit, likely non-linearity in the response. The resulting singularity report indicates that the data do not support estimating these parameters. This result goes to my earlier point about the lack of some observation might prohibit the estimation of some parameters.
The omitted rows cover conditions with a low concentration and upper half of Current range. Were all of the runs with these conditions omitted? Can this data be added back or re-run?

All comments

Mark_Bailey

First of all, an bad outcome is not necessarily a bad response. A DOE should explore as much of the factor space as possible. It is likely, even desirable, that some conditions produce bad outcomes so that the resulting model (the sole purpose of an experiment) is then able to predict the full range of the response including both good and bad outcomes. I would encourage you to complete the runs with the bad visual quality and include this data in the analysis of your experiment.

Second, in any case of missing data, you do not need to delete these rows. You can if you want to but it is not necessary. JMP will behave accordingly.

Third, the loss of data will compromise the analysis to some extent. It is difficult to be specific because the loss depends on both the number of missing values and the conditions that were not observed. Some compromises include loss of power in the hypothesis tests about the model and model parameters or the ability to estimate some parameters at all.

Mark_Bailey

I noticed some things with your study.

You have many runs, so you should have very how power to begin with. The omission of some runs should not dramatically affect your power.
The model for Metal% exhibits strong lack of fit. Predictions from this model will be biased and likely very inaccurate.
I tried adding quadratic terms to account for the lack of fit, likely non-linearity in the response. The resulting singularity report indicates that the data do not support estimating these parameters. This result goes to my earlier point about the lack of some observation might prohibit the estimation of some parameters.
The omitted rows cover conditions with a low concentration and upper half of Current range. Were all of the runs with these conditions omitted? Can this data be added back or re-run?

stevenfoong1

Thanks Mark. I will re-measured those bad samples and try run it again.