Do QC Exclusions Drive the Result? A qPCR Sensitivity Analysis
# Do QC Exclusions Drive the Result? A qPCR Sensitivity Analysis
Quality control protects an analysis from unreliable measurements. It can also change the sample size and, potentially, the conclusion.
That creates a reasonable question:
**Would we obtain the same result under another plausible replicate rule?**
This is what a sensitivity analysis is for. We do not change the primary rule after seeing the result. We keep the primary analysis intact and compare it with deliberately more permissive alternatives.
In [Part 4](/qpcr-normalization-changes-mirna-results/), reference choice changed several estimates substantially. For this post, the normalizer is held constant: every result uses **ref-RNA-2**, the more stable reference from Part 3. Only the treatment of technical-replicate problems changes.
## Three QC scenarios
The comparison uses three increasingly permissive rules.
### 1. Strict QC
This is the primary analysis.
– Both technical replicates must be detected.
– Their absolute Cq difference must be no greater than 0.75 cycles.
– The target and ref-RNA-2 must both pass.
Pairs that are discordant or have only one detected replicate do not receive a primary-analysis mean.
### 2. Keep discordant complete pairs
This scenario requires two detected measurements but retains pairs whose difference exceeds 0.75 cycles.
The two values are averaged even when they disagree. Single-detected pairs remain excluded.
This answers whether the discordance rule itself materially changes the estimate.
### 3. Use every available mean
This is the most permissive scenario.
– Complete pairs are averaged, including discordant pairs.
– If only one replicate is detected, that single value becomes the available mean.
This is not presented as the preferred method. It is a stress test: if the result changes only when weak measurements are allowed back into the analysis, that dependence should be visible.
## Compare the estimates, not only the p values
The figure shows inflammatory-minus-reference mean ΔCq differences. Negative values indicate greater relative expression in the inflammatory group. All intervals are 95% bootstrap confidence intervals.
The main pattern is reassuring. Candidate-miR-A remains negative under all three rules, while candidate-miR-B and candidate-miR-C remain compatible with little or no group difference.
## Candidate-miR-A: the result is stable
The estimates are:
– **Strict QC:** −0.40 ΔCq (95% CI −0.69 to −0.11), p = 0.010;
– **Keep discordant complete pairs:** −0.42 (−0.71 to −0.12), p = 0.007;
– **Use every available mean:** −0.42 (−0.71 to −0.13), p = 0.007.
The point estimate changes by less than 0.02 cycles. The confidence intervals are similar, and the direction is unchanged.
This suggests that candidate-miR-A’s group difference is not being created by the exclusion of a few discordant measurements.
The permissive analyses contain slightly more samples, but the additional observations do not materially increase or reduce the estimated effect.
## Candidate-miR-B: no hidden signal appears
The estimates are:
– **Strict QC:** +0.13 ΔCq (95% CI −0.18 to +0.44), p = 0.412;
– **Keep discordant complete pairs:** +0.08 (−0.23 to +0.40), p = 0.616;
– **Use every available mean:** +0.08 (−0.23 to +0.40), p = 0.616.
All intervals cross zero by a comfortable margin. Allowing discordant pairs increases the sample count but does not reveal a meaningful group difference.
This also reinforces the finding from Part 4: the nominal candidate-miR-B result produced by ref-RNA-1 was driven by normalization, not by the strict QC rule.
## Candidate-miR-C: the direction changes, but the conclusion does not
Candidate-miR-C contains all nine single-detected replicate pairs, so it is the assay most affected by the permissive scenario.
– **Strict QC:** −0.08 ΔCq (95% CI −0.41 to +0.23), p = 0.630;
– **Keep discordant complete pairs:** −0.04 (−0.36 to +0.27), p = 0.790;
– **Use every available mean:** +0.03 (−0.28 to +0.34), p = 0.836.
The point estimate crosses from slightly negative to slightly positive, but every estimate remains close to zero and every confidence interval is wide enough to include both directions.
This is an important distinction. A sign change can look dramatic when only the plus or minus sign is reported. Here it reflects movement among several small, uncertain estimates—not reversal of a clear biological effect.
## The complete results
The table includes the available sample count for each scenario. Strict QC has the smallest numbers because the target and reference must both pass. Keeping complete discordant pairs restores several samples. The most permissive scenario restores the single-detected candidate-miR-C pairs as well.
Sample size alone does not determine data quality. A larger analysis based on weaker measurements is not automatically preferable to a smaller analysis with an explicit precision requirement.
## Why non-detects deserve special treatment
A non-detect is not necessarily an ordinary missing value.
In a real qPCR experiment, the probability of non-detection may increase when the target concentration is low. Missingness can therefore be related to the unobserved measurement itself. Research on [qPCR non-detects](https://pmc.ncbi.nlm.nih.gov/articles/PMC4133581/) describes them as a potentially non-random missing-data problem rather than values that can always be discarded or replaced by a fixed maximum cycle.
In this synthetic dataset, the nine non-detect measurements were deliberately assigned to candidate-miR-C during generation. We know exactly how they arose. In real data, additional questions would be necessary:
– Was amplification truly absent?
– Did the amplification curve fail a quality criterion?
– Was the measurement below a validated limit of detection or quantification?
– Was there a plate, reagent, or pipetting problem?
– Is non-detection more frequent in one study group?
Replacing every non-detect with Cq 35 or Cq 40 can create an artificial pile-up and overstate certainty. A model for censored or non-random missing measurements may be more appropriate when non-detects are common. A recent methodological study also discusses [multiple imputation and direct estimation](https://pubmed.ncbi.nlm.nih.gov/33243147/) for this setting.
This tutorial does not need such a model because non-detects are few and the aim is to demonstrate transparent sensitivity analysis. The permissive scenario should not be mistaken for a general recommendation to accept single measurements.
## Why we do not choose the rule with the smallest p value
The QC rule was defined before testing group differences. That protects the analysis from a subtle form of result selection.
If we tried several thresholds and reported only the one producing the most attractive p value, the QC process would become another unreported hypothesis search.
The correct sequence is:
1. define a primary rule using assay knowledge;
2. calculate the primary result;
3. identify plausible alternative rules;
4. report how much the estimate changes;
5. explain any material disagreement.
In this dataset, the strict rule remains the primary analysis. The other two scenarios demonstrate that its central conclusions are not fragile.
## A compact interpretation
The sensitivity analysis supports three statements:
– candidate-miR-A retains a negative ΔCq group difference under every replicate rule;
– candidate-miR-B remains close to no difference;
– candidate-miR-C remains uncertain, even when all available measurements are used.
The analysis does not prove that every excluded measurement was erroneous. It shows that excluding them according to the predefined rule is not secretly generating the headline result.
## The practical takeaway
QC exclusions should be traceable, counted, and challenged with a reasonable alternative.
Do not hide the discarded measurements. Do not change the threshold until the desired result appears. Keep normalization fixed when evaluating a QC rule, and compare effect estimates with confidence intervals rather than watching only whether p crosses 0.05.
That turns a preprocessing decision into an auditable part of the analysis.
**Next:** *From Raw Measurements to Final Result: A Reproducible qPCR Checklist*
Leave a Reply