How to Handle qPCR Technical Replicates with One Outlier
If you run triplicates and one Ct value is more than 0.5 Ct away from the other two, you have an outlier — and you need to decide whether to drop it or keep it. The short answer: if two replicates agree tightly (within ~0.3 Ct) and the third is off by more than 0.5 Ct, it's reasonable to exclude it and average the remaining two. But you need a rule you apply consistently, decided before you look at your data, and you need to document it.
The longer answer depends on how far off the outlier is, what your sample looks like, and whether the problem is a pipetting error or something more systemic. A triplicate of 22.1, 22.3, and 23.8 is a different situation than 22.1, 22.3, and 22.6. Let's walk through how to think about this so you can defend your choice in a paper, a committee meeting, or just to yourself at 11 PM when you're staring at a spreadsheet.
What counts as an outlier in technical replicates?
Technical replicates exist to catch pipetting errors — they're the same cDNA loaded into multiple wells to confirm you can reproducibly measure the same template. They are not biological replicates and they don't increase your sample size for statistics. Their only job is quality control.
For most qPCR workflows, a standard deviation of less than 0.167 Ct across triplicates (which corresponds to a coefficient of variation of about 0.5 Ct range) is considered acceptable. In practice, if all three values fall within a 0.5 Ct window, you're fine — average them and move on. Where it gets messy is when two cluster together and one drifts.
Here's a practical threshold that works well and is easy to explain:
- Keep all three if the range (max − min) is ≤ 0.5 Ct.
- Flag the outlier if one replicate deviates by > 0.5 Ct from the median of the three.
- Exclude the outlier if it deviates by > 1.0 Ct from the other two, which almost certainly indicates a pipetting error, a bubble, or a seal issue.
- Discard the entire sample if no two replicates agree within 0.5 Ct — at that point, nothing is salvageable and you likely have a sample integrity problem.
Some labs use a stricter cutoff of 0.3 Ct, which is defensible for high-copy targets like 18S or GAPDH where Ct values should be extremely tight. For low-abundance targets (Ct > 30), you'll naturally see more scatter, and 0.5 Ct is more realistic.
The case for dropping one replicate (and when it's fine)
Imagine your triplicates for HPRT1 come back as 24.2, 24.4, and 26.1. The first two agree within 0.2 Ct. The third is 1.8 Ct higher — that's roughly a 3.5-fold difference in apparent template, which almost certainly reflects a well that got shorted on volume or had a bubble interfere with fluorescence detection.
Dropping that third replicate and averaging 24.2 and 24.4 to get 24.3 is completely reasonable. You're left with a duplicate rather than a triplicate, which is less ideal but still gives you a usable measurement. This is the most common scenario, and every working qPCR lab handles it this way.
A few important ground rules:
Set your exclusion criteria before analysis. Write it in your methods or lab notebook: "Technical replicate outliers were defined as values deviating > 0.5 Ct from the median and were excluded." This is not optional — deciding post hoc which points to drop is a fast lane to unconscious cherry-picking.
Apply the rule uniformly. If you drop an outlier from your treated sample, you must apply the same rule to your controls. You cannot selectively clean up one group.
Never drop a replicate to "fix" a result. If excluding the outlier changes your biological conclusion (gene goes from non-significant to significant), that's a red flag about the robustness of your data, not a justification for exclusion.
Track how many you're dropping. If you're excluding outliers from more than 10-15% of your wells across a plate, the problem isn't individual wells — it's your pipetting technique, your plate seal, or your cDNA quality. Fix the upstream issue.
Why statistical tests on triplicates are usually overkill
You'll occasionally see recommendations to run Grubbs' test or Dixon's Q test on your technical replicates. With three data points, these tests have essentially no statistical power. Grubbs' test on a triplicate can only flag the most extreme outliers (roughly > 1.15× the standard deviation from the mean for n=3 at p < 0.05), and the result is not meaningfully different from just eyeballing whether one point is far from the other two.
For triplicates, a simple distance-from-median rule is more practical and more transparent than a formal statistical test that gives a false sense of rigor. If you're running quadruplicates or more (rare, but some high-throughput setups do this), then Grubbs' test becomes more useful.
Where you absolutely should use proper statistics is on your biological replicates — comparing ΔCt values across your n=3 or n=6 biological samples using a t-test or ANOVA. That's where your actual statistical power lives. Technical replicate averaging is just a data-cleaning step that feeds into that analysis.
A worked example
Let's say you're measuring ACTB (reference) and IL6 (target) in treated vs. control samples, three biological replicates each, with technical triplicates.
One of your control samples gives these IL6 Ct values: 29.4, 29.6, 31.8.
Step 1: The median is 29.6. The third replicate is 2.2 Ct away. That's well beyond the 0.5 Ct threshold — exclude it.
Step 2: Average the remaining two: (29.4 + 29.6) / 2 = 29.5.
Step 3: Your ACTB triplicates for the same sample are 18.1, 18.2, 18.3. All within 0.2 Ct — no outlier, average to 18.2.
Step 4: ΔCt = 29.5 − 18.2 = 11.3.
Now carry that ΔCt forward into your biological replicate analysis alongside the other two control samples.
What if the IL6 values had been 29.4, 29.6, and 30.0? The range is 0.6 Ct, and the third point is only 0.4 Ct from the median. This is borderline. Personally, I'd keep all three and average to 29.67. A 0.6 Ct range for a target at Ct ~30 isn't unusual — you're dealing with lower copy numbers where stochastic variation is real.
What the outlier might be telling you
Before you drop the value and forget about it, take 10 seconds to check a few things:
Look at the amplification curve. On a QuantStudio or CFX96, pull up that individual well. Does it have a normal sigmoid shape, just shifted late? Or is it a weird partial curve, a flat line that barely crossed threshold, or a curve with an early bump? Abnormal curve shapes suggest a reagent or template problem in that well specifically.
Check neighboring wells. If the outlier is on the edge of the plate, thermal edge effects on older instruments can shift Ct values. This is less of an issue on modern Peltier-based cyclers but still shows up occasionally on 384-well runs.
Check the melt curve (SYBR assays). If the outlier well shows a different melt peak — a shoulder, a double peak, or a shifted Tm — that replicate likely amplified a different product (primer dimer, off-target). Exclude it with confidence, and note the reason.
Check for a pattern across the plate. If well H12 is always the outlier across multiple targets, you might have a seal issue in that corner, or your multichannel isn't delivering evenly to the last column.
Setting a lab-wide policy
The best time to establish your outlier-handling rule is before your first experiment in a project. Write it into your analysis pipeline and your methods section. Here's language you can adapt:
Technical replicates (n=3) were assessed for concordance. Replicates deviating > 0.5 Ct from the median of the triplicate were flagged; replicates deviating > 1.0 Ct were excluded. Samples with fewer than two concordant replicates were repeated. Exclusion criteria were applied uniformly across all samples and targets.
This is specific, defensible, and tells reviewers exactly what you did. It also makes your analysis reproducible — someone else can apply the same rule to your raw Ct exports and get the same results.
If you're running your analysis through VoilaPCR, it flags technical replicate outliers automatically using a threshold you set, marks excluded wells in the output, and logs the decision so you have a clean record without manually scanning every triplicate on a 96-well plate.
The one thing that actually matters
Consistency. A 0.5 Ct cutoff and a 1.0 Ct cutoff are both defensible. What's not defensible is dropping outliers when they make your treated group look better and keeping them when they don't. Pick a rule, write it down, apply it blindly, and move on. Your data will be cleaner and your scientific conscience will be clear.