Additional Capabilities of Parallel Line Method

· 5 min read

On this page
  1. Bioassay — Parallel Line Method
  2. What Was Added
  3. Regulatory and Statistical Background
  4. The Mathematics
  5. Reasoning: Why Two Conventions, and When to Use Each
  6. What Changes in the Output, and What Does Not
    1. Affected:
    2. Not affected:
  7. Where to Find This Capability
  8. Verification Against Published Reference Data
  9. References

Bioassay — Parallel Line Method

Capability added: Optional Potency Confidence Interval Error Term ("Regression Residual" vs. "Pure Error")

 

What Was Added

The Parallel Line Method now offers a second, optional way of choosing which measure of assay variability is used to build the confidence interval around the estimated relative potency. Previously, only one convention was available. The estimated potency itself is never affected by this setting — only the width of its confidence interval, and (indirectly) how conservative or how tight the reported precision is.

This addition does not change any default behavior. It is an opt-in setting, added specifically so that results can be reconciled with the convention used by other EP 5.3-compliant biological assay software, and with historical assay records that were analyzed under that convention.

Regulatory and Statistical Background

A parallel-line assay works by fitting a straight line to each preparation's response against log(dose), constrained so that all the fitted lines run parallel. Before a potency estimate can be trusted, the assay must pass three validity checks: that the dose-response relationship is statistically significant, that the preparations' lines are truly parallel, and that each preparation's relationship is linear over the tested dose range (no lack of fit).

Both of the recognized pharmacopoeial frameworks — European Pharmacopoeia (Ph. Eur.), Chapter 5.3 — "Statistical Analysis of Results of Biological Assays and Tests" and United States Pharmacopeia, General Chapter <111> — "Design and Analysis of Biological Assays" — build the potency confidence interval using Fieller's theorem, which requires an estimate of the underlying error variance of the response. Where the two conventions in practice diverge is in exactly which variance estimate is fed into that theorem, and this is what the new option controls.

The Mathematics

Convention A — "Regression Residual" (this tool's existing default). After fitting the common-slope, parallel-lines model, whatever variability remains unexplained by that model is pooled into a single residual mean square:

MS(regression residual) = SS(residual after common-slope fit) / df(residual after common-slope fit)

This residual technically contains three things blended together: pure replicate-to-replicate noise, any (hopefully negligible) departure from parallelism, and any (hopefully negligible) departure from linearity. In a genuinely valid assay — one that has already passed its three validity checks — the latter two contributions should be small, so this quantity is very close to pure noise in practice.

Convention B — "Pure Error" (the new option; the convention EP 5.3-compliant commercial bioassay software defaults to). Instead of the model residual, this uses only the variability between replicate readings taken at the exact same preparation and dose — the same "Residual (pure) error" row already reported at the bottom of the assay's own ANOVA table, and already used as the denominator for the three validity F-tests themselves:

MS(pure error) = SS(pure error) / df(pure error), where SS(pure error) is the sum, over every preparation x dose combination, of each replicate's squared deviation from its own cell mean

Whichever mean square is selected feeds the same, unchanged Fieller confidence interval calculation (Fieller, 1944) — only the variance estimate and its degrees of freedom change, not the underlying formula for combining slope and intercept uncertainty into a ratio confidence interval.

Reasoning: Why Two Conventions, and When to Use Each

Both conventions are legitimate and both appear in accepted regulatory guidance and commercial software; neither is "more correct" in an absolute sense — they simply answer slightly different questions. The regression-residual convention gives a confidence interval that reflects the total uncertainty actually observed in the fitted model, including any small, non-significant amount of non-parallelism or non-linearity that happened to occur in that particular run. The pure-error convention isolates only the irreducible measurement noise between replicates, deliberately excluding any model lack-of-fit from the precision estimate, on the reasoning that a valid assay should not be penalized for lack-of-fit variation the validity tests already confirmed was not statistically significant.

In practice, for an assay that comfortably passes its validity tests, the two conventions produce very similar confidence intervals, because there is little non-parallelism or non-linearity variation for the pure-error convention to exclude. They diverge more noticeably when non-parallelism or non-linearity, while not significant enough to fail the assay, is still large enough to visibly inflate the pooled residual.

The practical reason to select "Pure Error" is reconciliation: laboratories that have historically used EP 5.3-compliant commercial software, or that need a result directly comparable to a validated legacy method or a regulatory submission built on that convention, can select it to match those numbers. Where no such external comparison is needed, the tool's existing default remains a valid, defensible choice.

What Changes in the Output, and What Does Not

Affected:

  • Lower CL and Upper CL (potency confidence limits) in the potency results table
  • Fieller's g statistic (a measure of confidence-interval reliability, since it also depends on the chosen variance)
  • A new "CI Error Term" column identifying which convention and how many degrees of freedom were used

Not affected:

  • The estimated relative potency itself
  • The regression coefficients (slope, intercepts) and the full ANOVA table
  • The three validity test results (Significance of Regression, Parallelism, Linearity) — these already used the pure-error mean square as their denominator regardless of this setting

Where to Find This Capability

On the Parallel Line Method tool's input form, under the confidence-interval related settings, a new field labelled "Potency CI error term" offers a choice between "Regression residual (this tool's default)" and "Pure error — within-cell replicate variance only (EP 5.3 default)". The pure-error option requires at least one preparation/dose combination with more than one replicate observation.

pl-option


Verification Against Published Reference Data

The pure-error option was checked against a fully worked, published example rather than being implemented from the formula alone. The example is a five-dose, completely randomised parallel-line assay comparing a standard and three test preparations of hepatitis B vaccine, with three replicate readings at each of five doses per preparation (natural-log response transform).

Source: European Pharmacopoeia, 5th Edition (2005), Chapter 5.3, worked example 5.1.4.a — as reproduced in Stegmann Systems' PLA 3.0 published example report (Document-3536, "PLA_3.0.7_Example_Parallel-Line_Quantitative_Response_Assay_Standard_Report.pdf")


Preparation

Metric

Published (PLA/EP 5.3)

Computed (Pure Error option)

T

Estimated potency

43.41962

43.4196

T

95% CI

40.54479 – 46.53966

40.5448 – 46.5397

U

Estimated potency

35.16298

35.1630

U

95% CI

32.86981 – 37.64049

32.8698 – 37.6405

V

Estimated potency

39.40168

39.4017

V

95% CI

36.81254 – 42.20575

36.8125 – 42.2057

 

The full ANOVA decomposition (Treatments, Regression, Non-Parallelism, Non-Linearity — including its per-preparation breakdown — Residual (pure) error, and Total, with every sum of squares, mean square, F-value and p-value) also matched the published report exactly. Selecting "Pure Error" reproduced the published confidence intervals to the precision shown above; the tool's existing default convention is intentionally slightly different, as explained above, and was not expected to match.

References