How to Choose and Use Food Sampling Plans Correctly

“We test five samples per lot.” It’s the most common sampling statement in the food industry — and usually the least understood. What does five samples prove? What happens if one is positive? What’s the acceptance criterion? The answers are often vague, which means the lot acceptance decision — the release or rejection of product — rests on a statistical foundation nobody examined.

Sampling plans are the mathematics of lot acceptance: how many samples, what criteria, and what confidence the decision carries. Used correctly, they’re the defensible basis for release decisions. Used ritualistically, they’re the five samples that prove nothing. This guide makes them practical.

Step 1: Understand the Two-Plan and Three-Plan Basics

The attributes sampling plans come in two forms. The two-class plan (n, c): n samples are tested, and the lot is accepted if no more than c samples are positive (for pathogens, c is typically 0 — any positive rejects). The three-class plan (n, c, m, M): the lot is accepted if no more than c samples exceed m, and none exceed M — the marginal zone between m and M allowing limited tolerance for indicator organisms.

The n=5, c=0 plan for Salmonella means: test five samples, accept only if all five are negative. The n=5, c=2, m=10, M=100 plan for an indicator means: accept if no more than two samples are between 10 and 100, and none exceed 100. The plan’s string is the complete acceptance rule — quote it fully or don’t quote it at all.

Step 2: Grasp What the Plan Actually Proves

The critical insight: the sampling plan defines the producer’s and consumer’s risks — the probability of rejecting a good lot and accepting a bad one. The n=5, c=0 plan, applied to a lot with 5% contamination, accepts the lot about 77% of the time. Read that again: the standard five-sample pathogen plan misses the 5%-contaminated lot more than three-quarters of the time.

This isn’t a reason to abandon testing — it’s the reason to understand it. The sampling plan is one verification tool among many, not the safety proof. The plan’s stringency should match the decision’s importance: the higher the risk, the larger the n, the smaller the c. And the limitations should be honestly acknowledged in the program’s rationale.

Step 3: Match the Plan to the Decision

Different decisions need different plans. The lot acceptance for the high-risk pathogen uses the stringent two-class plan (n=60, c=0 for the statistically meaningful pathogen detection — the n=5 is the compromise, not the ideal). The process verification for indicators uses the three-class plan with the m/M reflecting the process capability. The investigation sampling uses the targeted approach — the suspect areas, not the random.

The plan selection considers: the hazard severity (the severe hazard justifies the larger n), the lot homogeneity (the well-mixed lot needs fewer samples than the stratified), and the consequence of the wrong decision. Document the rationale — the auditor asks “why this plan?” and the answer should be the risk reasoning, not “that’s our procedure.”

Step 4: Sample Representatively

The plan’s mathematics assume the samples represent the lot — the random or systematic sampling across the lot’s extent. The five samples all taken from the first pallet aren’t five samples from the lot; they’re five samples from one pallet. The sampling procedure specifies how representativeness is achieved: the stratified sampling across pallets, the systematic intervals, the random selection.

For the heterogeneous lots — the agricultural products, the non-uniform processes — the stratification matters more: the samples drawn from each identifiable sub-unit. The sampling record documents what was actually sampled — the traceability that lets the result be interpreted.

Step 5: Handle Compositing Correctly

Compositing — combining multiple samples into one analytical sample — saves money but costs sensitivity. The composite of five samples, tested once, detects the contamination only if it’s present at five times the method’s detection limit in the composite. For pathogen testing, compositing is generally inappropriate where the detection of low levels matters.

Where compositing is used (typically for indicators, not pathogens), the procedure specifies the compositing scheme, and the specification accounts for the dilution effect. The unrecorded compositing — the lab combining samples without the plan allowing it — invalidates the result’s interpretation. Agree the compositing rules with the lab in advance.

Step 6: Apply the Decision Rules Consistently

The plan’s acceptance rule is applied as written — no retesting to get a better outcome, no “engineering” the decision. The n=5, c=2, m=10, M=100 plan with three samples above m rejects the lot; the retest of new samples to replace the failures is the manipulation, not the procedure.

The marginal results — the samples in the m-to-M zone — trigger the defined response: the investigation, the increased scrutiny of the next lots, the trend review. The plan’s c isn’t just the acceptance number; it’s the early warning threshold. The lots consistently running at c are the process drifting toward the failure.

Step 7: Use Variables Plans Where Appropriate

Not all sampling is attributes (pass/fail). The variables plans — based on the measured values (the mean and standard deviation) — are more powerful for the characteristics like net weight, composition, or counts where the actual numbers matter. They need fewer samples for the same confidence, but they require the measurement data and the statistical competence to apply.

The program uses variables approaches in the trending and process capability work even where the acceptance uses attributes: the mean and standard deviation of the indicator counts, tracked over time, reveal the process shifts that the attributes plans only catch at the failure. The SPC mindset complements the acceptance sampling.

Step 8: Review Plans With the Data

The sampling plans are reviewed against the accumulated data: is the n adequate for the observed variability? Are the m/M values still appropriate for the process capability? Do the rejection rates suggest the plan is too tight (rejecting good lots) or too loose (passing bad ones)?

The review also examines the cost-effectiveness: the expensive large-n plans applied where the risk doesn’t justify them are the candidates for rationalization — with the documented rationale. The sampling program, like the testing program, evolves with the evidence.

Practical tips

Quote the full string. n, c, m, M — the complete acceptance rule, every time. The partial quote is the misunderstood plan.

Know the math. The 5%-contaminated lot passes n=5, c=0 most of the time — the honest understanding that keeps testing in its proper role.

Sample the whole lot. Stratified, systematic, representative — the samples that actually represent what you’re deciding about.

Apply rules as written. No retest-to-pass, no engineering — the discipline that keeps the decision honest.

Let data refine plans. Review against accumulated results — the program that gets smarter with evidence.

Common mistakes

The ritual five samples. n=5 with no understanding of what it proves is theater. Know the statistics, or don’t claim the assurance.

Unrepresentative sampling. All samples from one pallet means the lot decision is based on a corner. Stratify and systematize.

Hidden compositing. Samples combined without the plan allowing it silently loses sensitivity. Agree the compositing rules with the lab and document them.

Retesting to pass. New samples taken to erase a failure is manipulation. The plan’s rule is final.

Frozen plans. A plan set decades ago and never reviewed is irrelevant. Review with data.

Case snapshots

The auditor’s statistics. A “five samples proves safe” claim met the auditor’s probability lesson — the awkward silence that followed rewrote the rationale honestly, with the verification role properly defined.

The pallet sampling. All five samples had come from the first pallet — the finding that rewrote the procedure into genuinely stratified, representative sampling.

The compositing discovery. The lab was compositing pathogen samples the plan didn’t allow. The agreement was rewritten — compositing explicit, and only where appropriate.

The drifting c. Lots consistently landing at c=2 revealed a trend — and the investigation caught process drift early. The c value became the early warning it was meant to be.

Takeaways

Statistics with honesty. Plans understood, limitations acknowledged, roles defined — the sampling that’s defensible.

Representativeness is everything. The math assumes it; the procedure ensures it; the record proves it.

Rules are rules. Applied as written, reviewed with data — the discipline that makes sampling meaningful.

Checklist

• [ ] Two-class (n, c) and three-class (n, c, m, M) plans understood; full strings quoted
• [ ] Plan stringency matched to decision risk; statistical limitations honestly acknowledged
• [ ] Plan rationale documented per application (hazard severity, lot characteristics, decision consequence)
• [ ] Representative sampling procedures defined (stratified/systematic); records maintained
• [ ] Compositing rules agreed with lab; only where appropriate (not for low-level pathogen detection)
• [ ] Decision rules applied consistently; no retest-to-pass; marginal results trigger defined responses
• [ ] Variables/SPC approaches used for trending and process capability alongside attributes acceptance
• [ ] Plans reviewed against accumulated data for adequacy and cost-effectiveness