The assumption inside every unweighted fit.

The ICAL looks clean at the top. The 500 ppb standard back-calculates at +0.1%. The 0.5 ppb standard back-calculates at −91.8%. Same data, same detector. The only problem is the regression.

Ordinary least squares places the line to minimize the sum of squared residuals. It treats every residual as equally trustworthy. That holds only when the scatter of the response is the same at every concentration, a condition statisticians call homoscedasticity.

Chromatographic data rarely meet it. The absolute scatter of a peak area at 500 ppb is far larger than at 0.5 ppb, even when relative precision is the same. An unweighted fit spends its effort on the high standards, where residuals are largest. The low standards absorb the error.

Method 8000D describes exactly this tendency and names weighting as the compensation [1]. If your lowest standard keeps failing, check the weighting as well as the instrument.

Match the weighting factor to the variance.

A weighting factor tells the regression how much to trust each standard. It is how you tell the regression the truth about your data. Gu and colleagues give the rule that connects the factor to the data [2]:

The last row describes constant relative precision: a fixed percent RSD at every level. That is the common case for chromatographic detectors. It is why 1/x² is so often the right answer.

Burrows and Parr note that 1/x² weighting gives every calibration point the same weight in relative terms. They also show that an average response factor calibration is equivalent to a 1/x²-weighted regression through the origin [3].

So read the weight from the scatter you measure. Habit is not evidence.

Weighted regression on a ten-level curve.

The data below are a ten-level calibration from 0.5 to 500 ppb, with each level injected in triplicate. Four levels are shown with the scatter of their replicates.

The standard deviation grows by roughly three orders of magnitude across the range. The relative standard deviation stays in single digits. That is the third row of the rule: scatter proportional to concentration, which points to 1/x².

The weighting factor is how you tell the regression the truth about your data.

To test that prediction, the mean responses were fit three ways and each standard was back-calculated.

Illustrative data set constructed for this article. Every derived value is computed from the mean responses shown.

The unweighted fit misses the lowest standard by 91.8%. That fails Method 8000D’s 30% limit at the bottom of the range. 1/x brings every level inside 10%, though its largest errors still sit at the low end.

1/x² spreads the error evenly and holds every level within 2.9%. That is what the replicate data predicted. When the replicates and the fit agree, the weight is easy to defend.

Choosing the weight with the relative error test.

Replicates are the best evidence, but most routine ICALs have a single injection per level. For those, the practical test comes from the bioanalytical literature.

Almeida, Castel-Branco and Falcão fit each candidate weight. They chose the one with the smallest sum of absolute percent relative error (%RE) across the range [4]. Dolan recommends the same comparison and adds a constraint: use the least amount of weighting that minimizes the error [5].

In this example the test agrees with the replicates. The sum of absolute errors falls from 169 unweighted to 30 with 1/x and 17 with 1/x².

Gu and colleagues add a caution worth keeping in the record. In their LC-MS/MS assays, an incorrect weight changed reported concentrations only modestly, yet it significantly affected assay performance [2]. A wrong weight may move the numbers only a little. It still leaves a method that is harder to defend and less reliable near its limits.

What EPA, FDA and ICH say about weighting.

None of these documents mandates one factor for every method. All of them expect the choice to be made deliberately and recorded [1][6][7]. If an auditor asks why your method uses its weight, the answer should already be in the validation file.

A weighted regression checklist for the ICAL record.

1. Characterize the scatter. Use replicate injections from method development to see whether SD or RSD is constant.
2. Match the factor to the variance. Constant SD: unweighted. Variance proportional to x: 1/x. Constant RSD: 1/x².
3. Compare candidates on relative error. Back-calculate every standard under each weight. Record the sum of absolute %RE.
4. Use the least weighting that minimizes error. Do not pick a heavier weight than the data support.
5. Fix the weight in the method. Set it during validation, never batch by batch to rescue a failing curve.
6. Check the low standard. Weighting exists to protect accuracy near the reporting limit. Confirm it does.

Where WLTR fits.

Few labs compare three weightings for every analyte, in every ICAL, by hand. Most keep the weighting set when the method was first built, whether or not current data still support it.

WLTR is calibration data integrity infrastructure for regulated analytical laboratories. For each analyte, it derives nine calibration regression models: average response factor, plus linear and quadratic fits across four weightings each. It scores every model against fifteen ICAL parameters, including RSE. All of it is computed orthogonally to the CDS, under U.S. Patent 12,038,448 B2 [P].