AssayIndex
The analysis this index exists to be able to do

Inter-laboratory agreement

Where two laboratories assayed the same physical lot, the difference between their results is one laboratory's bias minus the other's, plus measurement noise. That is a matched pair, and the index holds 4,343 of them. Aggregating the pairs recovers each laboratory's offset with a confidence interval attached. Nothing on this page is asserted — it is computed from the pairs, and where the pairs are too few to compute anything the page says so.

4,343 matched pairs6 laboratoriesreference Janoshikmedian |difference| 0.34 pp
Section 1

Method

The unit of analysis is a lot, not a report. For every batch in the index assayed by two or more laboratories, each laboratory's results on that lot are reduced to their median — so a lot assayed five times by one laboratory and once by another contributes one pair, not five. That prevents a heavily re-tested lot from dominating the estimate.

For a pair of laboratories A and B, the difference on lot i is d_i = median(purity_A,i) − median(purity_B,i). Under the model that both laboratories measure the same latent lot purity with a fixed offset and independent noise, E[d] = bias_A − bias_B, and the standard deviation of d estimates the combined repeatability of the two.

The index reports, for every pair: the mean difference, its 95 % confidence interval (±1.96 · sd/√n), the 95 % limits of agreement (±1.96 · sd, a Bland–Altman construction describing where a single future difference would fall), and the share of pairs agreeing within 0.5 pp.

Assumptions, stated because they are load-bearing

  • The lot is homogeneous. If vials within a lot differ, that variance is attributed here to inter-laboratory disagreement, and the offsets are inflated. The index's median within-batch spread across lots assayed by a single laboratory bounds this: it is 0.40 pp.
  • The material did not change between assays. Pairs are not filtered by the interval between assay dates. Oxidative degradation over months is real, and it would show up here as a laboratory offset if one laboratory systematically tested later than another.
  • The offset is constant across compounds. It is not. Short-gradient laboratories carry a much larger offset on the acylated incretins than on anything else, because that is where the mass-identical diastereomers are. The all-compound figure below understates the effect where it matters and overstates it where it does not.
  • The reference is a convention. Janoshik Analytical is the zero point because it has the most pairs. Absolute accuracy is not established by any of this and would need a certified reference material assayed by all 6.
Section 2

Measured offsets against the reference

The "published" column is the offset each laboratory's profile page carries. The "measured" column is what this pairing recovers from the data. They should agree, and where they do not the discrepancy is printed rather than reconciled.

LaboratoryPublished biasMeasured from pairsPairs95 % CI on the mean95 % LoA (±)
Janoshik Analytical+0.000.000 (reference)————
PeptideMeter-0.05-0.024569-0.052 to 0.004±0.67AGREES
Medutest+0.08+0.0655170.030 to 0.101±0.81AGREES
VendorInvestigate-0.11-0.493380-0.533 to -0.453±0.78DIVERGES
Nordanalyt Laboratories+0.02-0.005357-0.043 to 0.033±0.72AGREES
Kestrel Bioanalytical+0.24+0.3822960.324 to 0.439±0.99CLOSE
-2.0-1.3-0.60.10.81.52.2PeptideMeter − JanoshikPeptideMeter − Janoshik: Q1 -0.26, median -0.02, Q3 0.19, n=569outlier -1.14outlier -1.10outlier -1.00outlier -0.94outlier 0.93outlier 0.99outlier 1.02outlier 1.16n=569Medutest − JanoshikMedutest − Janoshik: Q1 -0.19, median 0.05, Q3 0.33, n=517outlier -1.20outlier -1.14outlier -1.08outlier -1.04outlier -1.03outlier -0.97outlier 1.18outlier 1.41n=517VendorInvestigate − JanoshikVendorInvestigate − Janoshik: Q1 -0.75, median -0.50, Q3 -0.24, n=380outlier -1.75outlier -1.58outlier -1.56outlier -1.52outlier 0.94n=380Nordanalyt − JanoshikNordanalyt − Janoshik: Q1 -0.23, median 0.00, Q3 0.23, n=357outlier -1.44outlier -1.24outlier -1.09outlier -1.00outlier -0.97n=357Kestrel − JanoshikKestrel − Janoshik: Q1 0.00, median 0.33, Q3 0.73, n=296outlier 1.90n=296 Difference against the reference laboratory (pp)
Figure 1. Distribution of paired same-lot differences against Janoshik. The vertical line is zero. A box centred on zero is a laboratory with no detectable offset; a box shifted right is one that reads high. Box width is combined repeatability and matters as much as position — a laboratory with no offset and a wide box will still disagree with you on any given vial.
Section 3

Full pairwise matrix

Row minus column, in percentage points, with the number of matched lots beneath. Cells with fewer than two matched lots are not computed.

Row − columnJanoshikPeptideMeterMedutestVendorInvestigateNordanalytKestrel
Janoshik—+0.02
n=569		-0.07
n=517		+0.49
n=380		+0.00
n=357		-0.38
n=296
PeptideMeter-0.02
n=569		—-0.11
n=362		+0.45
n=272		-0.02
n=244		-0.41
n=230
Medutest+0.07
n=517		+0.11
n=362		—+0.56
n=249		+0.11
n=224		-0.37
n=188
VendorInvestigate-0.49
n=380		-0.45
n=272		-0.56
n=249		—-0.50
n=174		-0.91
n=143
Nordanalyt-0.00
n=357		+0.02
n=244		-0.11
n=224		+0.50
n=174		—-0.33
n=138
Kestrel+0.38
n=296		+0.41
n=230		+0.37
n=188		+0.91
n=143		+0.33
n=138		—
Section 4

How large is the disagreement, in practical terms

Median |difference| 0.34pp n=4343
90th percentile 0.93pp n=4343
Pairs agreeing within 0.5 pp 66.6% n=4343
02254506750.00.30.60.91.31.61.92.22.50.00–0.10 %: n=8020.10–0.20 %: n=6020.20–0.30 %: n=5710.30–0.40 %: n=5030.40–0.50 %: n=3830.50–0.60 %: n=3590.60–0.70 %: n=2940.70–0.80 %: n=1840.80–0.90 %: n=1810.90–1.00 %: n=1131.00–1.10 %: n=911.10–1.20 %: n=831.20–1.30 %: n=601.30–1.40 %: n=391.40–1.50 %: n=241.50–1.60 %: n=241.60–1.70 %: n=71.70–1.80 %: n=101.80–1.90 %: n=41.90–2.00 %: n=32.00–2.10 %: n=32.10–2.20 %: n=32.20–2.30 %: n=02.30–2.40 %: n=02.40–2.50 %: n=0median 0.34 Absolute difference on the same lot (pp) Reports
Figure 2. Absolute differences across all 4,343 matched pairs, regardless of which laboratories they involve. This is the practical answer to "how much should I trust one purity figure": half of all paired measurements of the same material differ by more than 0.34 pp, and 8 % differ by more than a full percentage point. Any comparison between two suppliers on a difference smaller than this is reading noise.