topic Re: What technique do I use to justify eliminating measurements at a specific timepoint in a series? in Discussions

What technique do I use to justify eliminating measurements at a specific timepoint in a series?

TCM — Fri, 09 Jun 2023 00:48:03 GMT

I have measurements of an attribute at initial, 2hr, 4hr, 6hr. I want to drop the 4hr measurements because it does not add much value to the business application. The measurements exhibit a negative slope from initial to final... What technique or rationale might I use to justify eliminating the 4hr measurement? [I was thinking along the lines of Information Gain, but I admit I don't know much about it beyond the context of Decision Trees]. I appreciate any guidance.

Re: What technique do I use to justify eliminating measurements at a specific timepoint in a series?

P_Bartell — Wed, 13 Apr 2022 16:49:00 GMT

IMO the technique or rationale is not a statistical methods question but one driven by knowledge of the process, data, and problem at hand. If the 4th data point adds no value, and eliminating it is not unethical or against some application accepted standard or process rationale...then just use the JMP exclude/hide capability to eliminate those observations from any analysis or reports.

Re: What technique do I use to justify eliminating measurements at a specific timepoint in a series?

TCM — Mon, 18 Apr 2022 12:47:24 GMT

Thank you for your response. To be clear, it is not the 4hr result that I am eliminating, but the analysis altogether so that going forward, we measure only at initial, at 2hrs and at 6hrs.

I can convince myself to not do the 4hr measurement, but as the data scientist, I think I would need to advance a technical rationale for the action with the domain subject experts.

Teresa

Re: What technique do I use to justify eliminating measurements at a specific timepoint in a series?

ih — Tue, 19 Apr 2022 11:45:13 GMT

I agree with @P_Bartell that this is more a business question than a statistical question. I am guessing this data is used to make a decision. If so you could do this:

Using all three points, work out what decision you should have made in each instance in history (this might or maybe should match what actually happened).
Remove the second point, pretend it was never taken. Adjust your procedure to only use the other two points. Now retroactively figure out what decision they would have made for each instance in history.

Were any different decisions made? If so what was the impact of those different decisions? This could be an actual cost from customer complaints, rework, reduced capacity, or it could be a change in risk. What would be saved by removing the additional test? Now someone in the business needs to decide whether that test is worth the cost.