Outliers, missing data, negative results, and exploratory analysis
Handle inconvenient observations with a documented rule, preserving what happened and separating exploration from planned decisions.
Field diagramPlot evidence with units
Estimated reading
8 min
Estimated practice
12 min
Equipment
Supplied dataset containing a missing entry, a negative result, and a potential outlier; analysis worksheet.
Progress stays in this page session. Allow lesson storage in Cookie settings to remember it; browser settings may also prevent saving.
Objective
By the end, you can:
Distinguish an unusual value from an error and an outlier decision from a visual impression.
Represent missing and negative results without silently converting them.
Label exploratory analyses and preserve the preregistered or planned decision rule.
Unusual is not automatically wrong
An outlier is an observation that is unusually distant from others under a chosen model or display. It may reveal a transcription error, instrument problem, sample difference, real rare event, or simple chance. The word does not identify the cause. First check identity, units, transcription, instrument status, method deviations, and sample condition. Only then apply an exclusion rule that was specified in advance or justify a clearly labeled exploratory sensitivity analysis.
Missing is a state, not zero. Record why a value is missing when known: not collected, instrument failure, below reporting threshold, or lost record. A negative result can be a meaningful observation, a signed change, a background-corrected value, or a signal below zero caused by noise. Do not replace it with zero or delete it because it conflicts with a desired direction.
Planned and exploratory decisions
A planned analysis uses rules selected before seeing the outcomes, such as an instrument-failure criterion and repeat policy. An exploratory analysis asks what patterns might be worth testing next; it can be useful, but its findings need confirmation. Show results with and without a proposed exclusion when that helps readers assess sensitivity, and explain which version supports the stated conclusion.
Preserve raw value and status.
Record the reason, evidence, and person making an exclusion decision.
Do not call a post hoc removal ‘predefined.’
Use a supplied dataset for practice; this is not a license to manipulate research data.
Worked example
A transparent sensitivity statement
A group has readings 9.9, 10.0, 10.1, and 12.0 mL. The 12.0 reading has no recorded deviation, but its source label is verified.
01Initial status: unusual but not proven erroneous; retain in raw data.
02Exploratory view: summarize with all four and with 12.0 excluded, clearly labeled.
03Conclusion: if the decision changes, report that sensitivity and propose a predefined rule for the next study.
Materials
Set out what you need.
Representative dataset
Data dictionary
Deviation and exclusion log
Plotting tool
Safety & stop conditions
Pause if the work no longer fits the plan.
Do not upload sensitive data to an unapproved tool.
Keep the raw file and an immutable or versioned copy.
Ask for qualified statistical review when exclusions affect a high-stakes conclusion.
Method
Work through the steps.
01
Flag, do not erase
Mark unusual, missing, or negative entries with a status and preserve the original field.
02
Audit the record
Check identity, units, transcription, instrument status, timing, and deviations before assigning a cause.
03
Apply the rule
Use the preplanned criterion if one exists. If not, label any rule as exploratory and explain why it was considered.
04
Run sensitivity views
When appropriate, compare the summary including and excluding a justified suspect value, keeping the analysis transparent.
05
Report limitations
State what the missingness, negative value, or exclusion does to precision, bias, and generalization.
Checkpoints
Observe, record, investigate.
Missing values are not coded as zero without justification.
A negative result retains its sign and unit.
Any exclusion has a criterion, evidence, and record.
Exploratory and planned conclusions are labeled separately.
Common mistakes
Correct the process, preserve the record.
The largest value is called an outlier because it is inconvenient.
Investigate its record and use a stated statistical or method rule, not its effect on the conclusion.
A missing result is filled with the group mean.
Preserve the missing state and use an appropriate missing-data plan; mean filling can distort variation and relationships.
A negative background-corrected value is changed to zero.
Keep the signed value and explain the measurement scale and reporting rule.
A post hoc exclusion is reported as planned.
Label it exploratory and show how the conclusion changes.
Practice
Audit three awkward values
Use this supplied practice set: 9.9, 10.0, missing because the instrument stopped, −0.2 after background correction, and 12.0 mL with no recorded deviation. Write a status and next action for each.
What evidence would make the high reading suspect?
Why is missing not zero?
What does the negative sign mean in context?
What must be reported if the high reading is excluded?
