3

I wonder whether I am behaving correctly with my Student t-test exercise.

I am continuously testing a small data sample of metal-detector antenna readings against a calibration sample distribution.

I am experiencing slow drifts in the distribution of my data (due to several factors like temperature raise, or soil composition changes, etc.), and I want to be able to automatically follow the drift in order to keep the t-test reliable in the long run.

My current idea is to adjust the mean of the calibration sample at each iteration, given some constraints:

WHILE(true) {
    ... t-test calculations...
    IF (p-value < P_LIMIT) {
        E(X_cal) += (DRIFT_WEIGHT * (E(X_test)-E(X_cal))
    }
}

Hence, only when I have a certain confidence on the null-hypothesis, then I adjust the calibration in the direction of the currently computed mean, weighting properly.

In this situation, in case of a slow drift, the calibration mean would be pushed towards the current mean iteration after iteration until it calibrates.

QUESTION: I wonder if this is making some sense, or whether I am violating important corner-stone rules in statistical distribution comparisons.

kjetil b halvorsen
  • 77,844
Campa
  • 163
  • What does your actual raw data look like? What are you testing? – oneloop Mar 29 '18 at 13:52
  • Metal-detector antenna values, imagine a normal distribution (verified) with a mean value around 500.000 and a variance at rest of around a few hundreds. – Campa Mar 29 '18 at 14:20
  • And then something counts as a detection? So you want to detect outliers somehow? – oneloop Mar 29 '18 at 15:16
  • I want to detect differences between the calibration and test sample means: H0 = abs(mu1-mu2)>threshold. That would mean that there is a relevant presence of metal and hence a positive shall be triggered. – Campa Mar 29 '18 at 15:19
  • Maybe you should make a control chart and look into methods from quality control? One goal of such methods is to detect drifts in processes. – kjetil b halvorsen Nov 05 '18 at 10:49
  • 1
    Thank you @kjetilbhalvorsen. I'm afraid I might have not clearly understood what you mean: any useful link so that I can dig in a little more? – Campa Nov 05 '18 at 15:52
  • 1
    You could start with this: https://stats.stackexchange.com/questions/15364/understanding-statistical-control-charts And here is a longer list – kjetil b halvorsen Nov 05 '18 at 15:58

0 Answers0