Measuring for Trends in Numeric Ordinal Scores Measured on Lots over Time

[#1]

What is the best statistical measurement technique to monitor ordinal data so that we can detect trends in the data over time?

I have individual odor panel data (y-axis) for each lot of product (x-axis) produced over time and would like to detect a trend change. Details: (1) 2 odor panelists score each lot of product produced, (2) two panelists are random from a pool of trained panelists, (3) odor scores are ordinal with scores of 1, 2, 3, 4, or 5, (3) score of 3 is our low specification limit for release, (4) a lot might be measured twice if the filling lot order occurs inter-shift. 

The data is attached for one product SKU produced in 2023-2024 (to date).

[#2]

I'm not sure I can answer your question, but here are my thoughts:

1. There isn't enough discrimination in the measurement system to be of much use to detect trends especially when you aren't using the entire scale.  You essentially have nominal data (e.g., 4 or 5).

2. Change the definitions of the categories.  For example, use the current 3 as 1 and the current 5 as 5 and add 3 more categories between those.

3. You might try increasing the number of panelists and using the average of their scores (after you assess the between panelist variation).  

4. Find another continuous measure that correlates with the odor measurement.

[#3]

statman:

Thank you for that perspective.  While our quality systems are really good at producing product that is almost always a 5 (best), the whole scale is in fact used.  However it is not often that we produce a 3 (lower specification limit for release) and as you can see we do on occasion produce product that's a 4.  That said I'm sure we can improve detection of malodors, but given the current system we'd like to be able to detect when our existing panelists evaluate something that's maybe not quite right that might turn into a failure (1,2, or a combination of 1,2,3 that averages less than 3).

[#4]

I've included a different product LAEU with a longer history of production that also has a wider range of scores that includes lots that failed for odor.  Does this larger dataset aid in determining a best measure to spot trends when the odor data is numeric / ordinal?

Again the odor scores were evaluated as a 1, 2, 3, 4, or 5; 5 is best, 3 is the lower limit for specification (to pass).  Odor scores are evaluated each time by two, trained, random panelists in a trained pool of panelists, and a lot if produced over two shifts is evaluated on each shift.