Moisture Analyzer vs Oven Method: How to Correlate Them Properly

A note on the numbers in this article

Every dataset below — the twelve paired results, the after-adjustment comparison, and the replicate tables — is an illustrative worked example. The figures are constructed to demonstrate the calculation and are internally consistent and reproducible; they are not measurements from any specific instrument, laboratory or customer, and they should not be used as expected performance for your material. Manufacturer specifications quoted alongside them are A&D published figures and are identified as such.

This is the deep version of a question introduced in how moisture analyzers work — which explains the loss-on-drying principle and why methods legitimately disagree — and in the moisture analyzer buying guide, which covers resolution tiers and sample preparation. Read those for the "why"; this one is the "how". The instruments themselves are in our moisture analyzer range.

Why does a moisture analyzer disagree with the oven at all?

Three separate causes, and telling them apart is most of the work.

The methods measure different things at different rates. Loss-on-drying reports mass lost on heating, whatever leaves the sample — water plus any other volatile. An oven at a lower temperature over four hours and a halogen analyzer at 200 °C for six minutes will not remove the same set of substances from a spice, a high-fat food, or a solvent-damp resin. A&D's halogen analyzers use a 400 W straight halogen lamp with the SRA (Secondary Radiation Assist) filter, and the pan reaches 200 °C from ambient in about two minutes — a heating profile no oven reproduces.

The end-points are defined differently. An oven method usually stops at a fixed time, or when consecutive weighings agree within a stated amount. A moisture analyzer can stop on a fixed timer (1–480 minutes on the A&D range) or on an automatic end-point that triggers when the rate of moisture change falls below a set rate. Two different stopping rules produce two different final masses from the same sample. This is the single most adjustable cause of bias, and the one you should attack first.

The two sub-samples are not the same sample. You cannot put the same 5 g into both instruments. Whatever difference exists between the two halves you split shows up in your data as method disagreement when it is actually sampling variance. A study that does not separate this out will over-state the bias.

What does "correlating" the two methods actually mean?

It means quantifying agreement, not computing a correlation coefficient. This distinction is the most common technical error in these studies, and an auditor who knows the subject will look for it.

Take the illustrative twelve-pair dataset in the next section. The Pearson correlation coefficient between the oven and analyzer results is r = 0.9998 — as close to a perfect straight line as data gets. And yet the analyzer reads high on every single one of the twelve pairs, by +0.133 percentage points on average. A correlation coefficient measures whether two methods move together; it is completely blind to a constant offset between them. A study that reports r = 0.9998 and concludes "excellent agreement" has demonstrated nothing about agreement at all.

What you need instead is the distribution of the differences between paired results — the mean bias, how much that bias scatters, and how confident you are in the mean. This is the Bland–Altman approach to method comparison, and it is the standard tool for exactly this problem.

How do you design the paired-sample study?

Design decisions, in the order you make them.

Cover the range you actually work in. Pick samples spanning your specification range and a little beyond it — if you release product between 8% and 12%, run pairs from roughly 6% to 13%. A study run entirely at one moisture level tells you nothing about whether the bias changes with level.

Use real production material, across batches. Draw from at least three different production batches or incoming lots. Material from a single batch understates the variability the method will meet in service.

Choose n and justify it. No standard fixes the number of pairs. Twelve to twenty pairs is a common and workable starting point: at n = 12 the 95% confidence interval on the mean bias is about ±0.64 × s_d (t ÷ √n = 2.201 ÷ 3.464), which is usually tight enough to decide. Set the number your quality system can defend, and record the reasoning.

Split each sample properly. Homogenise and quarter each sample, then take one sub-sample for the oven and one for the analyzer at the same moment. Sub-sampling error is the largest avoidable source of noise, and in Gulf conditions an exposed sample exchanges moisture with the room within minutes. Seal, or test immediately.

Randomise and blind where you can. Run the pairs in a scrambled order, and have the oven analyst work from a sample code rather than the analyzer result. Human expectation is a real source of drift in a method study.

Fix everything else before you start. Sample mass, layer thickness, pan type, drying temperature and end-point criterion must be identical for every analyzer run in the study. If you change any of them mid-study, you have two studies and neither of them is finished.

How do you calculate mean bias and limits of agreement?

Work in percentage points (pp) throughout — the difference between 9.31% and 9.16% is 0.15 pp, not 1.6%.

For each pair, the difference is d = analyzer − oven. From the set of differences: mean bias d̄ = Σd ÷ n · SD of the differences s_d = √( Σ(d − d̄)² ÷ (n − 1) ) · standard error of the mean bias SE = s_d ÷ √n · 95% confidence interval on the bias d̄ ± t × SE, using t for (n − 1) degrees of freedom · 95% limits of agreement d̄ ± 1.96 × s_d.

Worked example (illustrative figures): study A, before method adjustment

The twelve pairs below are constructed to demonstrate the calculation. They are not measurements from a specific instrument, laboratory or customer. The scenario: a food powder, oven reference method against a halogen analyzer running at 105 °C with an automatic end-point, 5 g sample, twelve pairs.

Results: mean bias +0.133 pp; s_d = 0.0385 pp; SE = 0.0111 pp; t(0.975, df 11) = 2.201, giving a 95% confidence interval of +0.109 to +0.158 pp. The interval excludes zero, so the bias is real, not sampling noise. The 95% limits of agreement are +0.058 to +0.209 pp — any single future pair should fall in that band.

Is the bias proportional? Regress each difference against the mean of its pair. Here the slope is −0.003 pp per pp with a t-statistic of −0.46 against a critical value of 2.228 (df 10) — nowhere near significant. The bias is constant across the 6.6% to 12.4% span of pair means, so a single offset describes it. Had the slope been significant, no single offset would be valid and you would need a level-dependent relationship or a narrower claimed working range.

Now the decision. Is +0.133 pp acceptable? That is not a statistical question — it is a question about your product. If your specification is 10.0% ± 1.0%, a 0.133 pp bias consumes 13% of your tolerance and is arguably tolerable. If the specification is 10.0% ± 0.2%, it is not.

Worked example (illustrative figures): study B, after adjusting the drying programme

Continuing the same constructed example — again, not measurements from a specific instrument, laboratory or customer. Rather than accept the offset, the drying temperature was lowered and the study repeated on split samples from the same twelve lots. The twelve differences (analyzer − oven, in pp) come out as +0.03, −0.02, +0.04, 0.00, +0.05, −0.01, +0.02, +0.01, −0.03, +0.04, +0.01, +0.02 — so you can check every figure in the right-hand column below.

Study B's confidence interval includes zero, so at n = 12 there is no demonstrable bias between the methods. That is a far stronger position than a documented correction factor, and it is why method development beats arithmetic correction: a correction factor is a permanent obligation to explain, while a tuned drying programme is just the method.

This is exactly what A&D's WinCT-Moisture software is for. Supplied as standard with the MS-70 and MX-50, its RsTemp function automatically determines the optimum heating temperature in a single measurement across the 30–200 °C range, and RsFig displays the real-time moisture-rate-change graph so you can see whether the sample has genuinely plateaued or is still slowly decomposing. Data saves as CSV with statistics, which is the raw evidence a correlation study needs.

How do you check repeatability — and against what?

Bias tells you where the methods sit relative to each other. Repeatability tells you whether either of them is stable enough for the answer to mean anything. Run two different checks; they answer different questions.

Check 1 — instrument repeatability, on a homogeneous reference. A&D supplies sodium tartrate dihydrate with the range as a standard accessory, a stable reference with a theoretical moisture content of 15.66%. Run replicates on one instrument and take the standard deviation.

Check 2 — in-service repeatability, on your product. This will always be worse than the published specification, because it contains sub-sampling variance as well as instrument variance.

The table below shows what such a check set looks like. In that illustration, 0.030 pp on real product against 0.017 pp on tartrate is not an instrument fault; it is the sample being a sample.

Two things follow that most QC managers find genuinely surprising. First, the reference oven is usually the noisier method — 0.077 pp against 0.030 pp in the illustration — because it accumulates desiccator handling, cooling time and ambient exposure at every step. Second, if the oven's own repeatability is worse than the bias you are chasing, you cannot resolve that bias no matter how many pairs you run. Measure the reference method's repeatability before you decide what bias is worth pursuing.

What should your instrument repeatability be compared against?

A&D publishes moisture-content repeatability as a standard deviation at two sample masses. This is the acceptance criterion your tartrate check should be read against — and it shows plainly why sample mass matters more than most operators expect.

Every model is roughly five times more repeatable on a 5 g sample than on a 1 g one. If your method specifies a 1 g sample because that is all you have, size the instrument accordingly — or find a way to run 5 g, which is usually cheaper than buying a tier up.

Two cautions on using these figures. They are instrument figures measured on a homogeneous reference: never quote them as the expected scatter on a real product, which carries sub-sampling variance on top. And they are repeatability, not accuracy — an instrument can be perfectly repeatable and still biased against your oven, which is the entire reason for the correlation study above.

What if my specification is written against Karl Fischer, not an oven?

The same procedure applies unchanged — paired samples, mean bias, confidence interval, limits of agreement — but expect a larger and more sample-dependent bias, because Karl Fischer titration measures water chemically while loss-on-drying measures total mass loss.

One published data point is worth knowing, with an important qualification. A&D publishes a direct comparison on PET plastic pellets and, unusually for a manufacturer, publishes the conditions alongside it. The two methods agreed to within 0.009 pp, the halogen method was the more repeatable of the two, and it took roughly a third of the time. A&D also states that on PET and similar materials the MS-70 resolves a moisture change of less than 1%.

Read it for what it is. This is one material, one temperature, five replicates, published by the manufacturer — and note the sample masses differ by more than thirty-fold, which alone can shift a result. It is evidence that low-moisture correlation with Karl Fischer is achievable; it is not a substitute for running the study on your material. Any correlation study is material-specific by definition. The same figures appear in our guide to how moisture analyzers work — one canonical set of numbers across both pages.

What goes into the in-house method document?

The study is only worth what the document says. At minimum, an in-house loss-on-drying method should record seven things.

Scope — the exact material, grade and supplier the method covers, and the moisture range validated. Reference method — the oven procedure this was correlated against, named precisely enough that someone can repeat it, including the specification or monograph it derives from. Instrument and settings — model and serial number, drying temperature, measurement programme (the A&D range offers Standard with HI/MID/LO accuracy, Automatic, Quick, Timer and Manual), end-point criterion, sample mass, pan and preparation.

The correlation data itself — the raw paired results, n, mean bias, s_d, confidence interval, limits of agreement and the proportional-bias check. Attach the table, not just the conclusion. Acceptance criteria and the decision — the bias you accepted and the reasoning, or the offset applied and its justification. Verification routine — the periodic check that confirms the relationship still holds, and what triggers a re-study. Approval and revision history — who authorised it, when, and what changed.

Store the raw drying curves alongside it. WinCT-Moisture's CSV export makes this straightforward, and a curve that plateaus cleanly is much better evidence than a single final number.

Storing the method on the instrument matters too. The MS-70 and MX-50 hold 20 sets of measurement conditions (MF-50: 10; ML-50: 5), so the validated programme is recalled rather than re-entered. An operator typing a temperature by hand is a documented method waiting to be broken.

How often should the correlation be re-verified?

No standard fixes an interval, and any source that gives you a number without knowing your process is guessing. Set a schedule and justify it on risk, exactly as you would for calibration intervals.

What should trigger a re-check is more concrete than how often: a change of raw material supplier, grade, particle size or formulation · a change to the reference oven method itself · replacement of the halogen lamp (user-replaceable, 5000-hour rated life) or any heater service · a drift in the routine sodium tartrate check · a run of results near a specification limit that the laboratory disputes.

Between full re-studies, a light periodic check — a handful of paired samples confirming the bias still sits inside the limits of agreement you established — is far cheaper than repeating the whole exercise, and is usually what an auditor is looking for.

Two verification points that are frequently conflated: calibrating the weighing element and verifying the drying temperature are separate exercises. A weight calibration says the balance section weighs correctly; it says nothing about whether the heater reaches the temperature the programme claims. A&D offers an optional certified temperature calibrator for the MS-70 and MX-50 where the drying temperature itself needs documented verification, with calibration output in GLP, GMP and ISO formats.

Where does UAE regulation come into this?

For most users, an in-house moisture method is a quality-system document rather than a legal-metrology one — but confirm that, rather than assume it.

In the UAE, the Ministry of Industry and Advanced Technology (MOIAT) administers legal metrology under Federal Decree-Law No. 20 of 2020, and in Dubai verification is carried out by Dubai Municipality's Dubai Central Laboratory under MOIAT authorisation. Whether a particular instrument in your facility falls within the scope of legal metrology verification is a question for those authorities — confirm it with them for your specific use rather than relying on a general rule.

Separately, Dubai Municipality mandates HACCP-based food safety management, and where a moisture specification is a control point in your HACCP plan, the method behind that number needs to be documented and defensible. That is precisely what a correlation study and an in-house method give you. Where that moisture figure derives from a water activity target, our guide to water activity vs moisture content covers the translation.

Which A&D model supports this kind of method work?

The correlation procedure is model-independent, but the instrument determines how easily you can do it. Two features matter for method development specifically: the automatic end-point mode, and drying-curve software.

The practical rule: if your correlation study needs to resolve a bias of a few hundredths of a percentage point — low-moisture plastics, pharmaceutical intermediates — you need the MS-70's 0.001% readability, because you cannot demonstrate agreement finer than your instrument's own resolution. For food and general chemical QC in the 5–20% range, the MX-50's 0.01% readability and WinCT-Moisture support are the sensible combination for method development. The MF-50 and ML-50 run validated methods perfectly well but give you less to work with while developing one.

A&D Gulf is the authorised A&D Japan distributor in the UAE and can support the method-development side of a correlation study. Availability is confirmed per model with your quotation.