ACADEMIC WRITING GUIDE
How to Conduct and Report Sensitivity Analysis in a Dissertation (With Worked Examples)
A dissertation can produce a convincing-looking statistical result that depends heavily on one decision: how missing responses were handled, whether an unusual observation was retained, or which variables were included in a model.
Your supervisor may therefore ask: How do you know your conclusion would still hold under a reasonable alternative assumption?
That is the purpose of sensitivity analysis.
A sensitivity analysis examines how conclusions respond to plausible changes in assumptions, analytical methods or data-related decisions. It helps readers distinguish findings that remain reasonably stable from findings that require especially cautious interpretation.
This guide explains how to plan and report sensitivity analyses, choose meaningful alternatives and interpret differences without selectively presenting whichever result looks strongest. Two fully worked numerical examples demonstrate the logic.
Important: All datasets, participants, scores and numerical results in this guide are hypothetical teaching illustrations. They are not findings from actual research. The methods used in a dissertation must fit its design, research questions and institutional requirements.
What is sensitivity analysis in research?
Sensitivity analysis investigates the dependence of a result or inference on assumptions and analytical decisions.
Suppose a researcher concludes that students who receive more academic feedback report higher engagement. Before interpreting that association, the researcher might ask:
Does the association depend on excluding incomplete questionnaires?
Does it change substantially when influential observations are examined?
Does a plausible alternative model produce a materially different estimate?
Could measurement error, selection bias or unmeasured confounding explain the pattern?
These questions do not automatically mean that the original analysis was incorrect. They identify where uncertainty may arise beyond ordinary sampling variability.
The STROBE reporting guidance discusses sensitivity analyses for observational studies, including alternative inclusion rules, missing-data assumptions, exposure definitions and adjustment choices.
The ICH E9(R1) framework, developed for clinical trials, makes a particularly important distinction: a sensitivity analysis should investigate the assumptions behind the primary estimator while preserving the scientific target of interest.
Not every alternative calculation meets that stricter definition. Some analyses change the population, outcome or research question. Those may be valuable supplementary analyses, but they should be identified accurately.
Sensitivity analysis, robustness checks and supplementary analysis
These terms are sometimes used interchangeably, but their precise meanings depend on the research field.
| Approach | Main purpose | Example |
|---|---|---|
| Sensitivity analysis | Examine how an inference depends on uncertain assumptions | Compare plausible missing-data scenarios for the same target population mean |
| Robustness check | Examine whether a result is stable across justified analytical decisions | Compare suitably specified estimators addressing the same association |
| Supplementary analysis | Provide another perspective or answer an additional question | Analyse a different subgroup or alternative outcome |
| Diagnostic check | Investigate whether a model or data feature raises concerns | Examine residual patterns or influential observations |
| Exploratory analysis | Investigate patterns not fully specified beforehand | Explore an unexpected relationship for future research |
The categories can overlap. What matters is explaining what was varied, why it was varied, whether the target remained the same and what the comparison can establish.
A sensitivity analysis is not simply a second statistical test.
When should a dissertation include sensitivity analysis?
Sensitivity analysis can be useful when an important conclusion depends on an uncertain but consequential assumption.
Common situations include missing data, potentially influential observations, uncertain measurement choices, plausible model specifications, sampling or selection decisions and assumptions about uncontrolled bias.
For example, a student examining survey responses may need to consider whether incomplete responses could materially change the reported average. A researcher fitting a regression model may need to examine whether influential observations disproportionately affect the estimated association.
Not every dissertation needs dozens of alternative models. A focused study may benefit from one well-chosen check that addresses its most consequential uncertainty.
The best starting point is your approved analysis plan. Our guide to writing a dissertation data analysis plan explains how to connect analytical decisions to research questions before examining results.
Step 1: Define the primary result clearly
Before testing sensitivity, identify the main result and the quantity it is intended to describe.
Record:
The research question.
The target population.
The outcome and relevant predictors.
The primary analytical method.
The estimated quantity, including its units.
The sample or observations contributing to the estimate.
Any important assumptions.
For example:
Primary question: What is the average usefulness rating for a data-visualisation clinic among all twelve eligible postgraduate attendees?
Outcome: Usefulness score from 0 to 20, with higher scores indicating greater perceived usefulness. This is an illustrative scale, not a validated instrument.
Problem: Only nine of the twelve eligible attendees provided a score.
This definition matters because the average among the nine observed respondents is not automatically the same quantity as the average among all twelve eligible attendees.
If the target changes between analyses, different results may reflect different questions rather than sensitivity to an assumption.
Step 2: Identify the assumption worth challenging
A meaningful sensitivity analysis begins with a specific vulnerability.
Consider the following questions:
| Potential vulnerability | Assumption or decision to examine | Appropriate question |
|---|---|---|
| Missing questionnaire items | Whether missing responses differ from observed responses | How much could missing values change the conclusion? |
| Unusual observations | Whether conclusions depend strongly on a small number of observations | How influential are these records, and are they valid? |
| Model specification | Whether functional-form or distributional assumptions are consequential | Does a defensible alternative estimator materially change the estimate? |
| Measurement | Whether a measure adequately represents the intended construct | What happens under another justified operational definition? |
| Confounding | Whether an unmeasured variable could affect an observational association | How strong would the bias need to be to alter the interpretation? |
Do not generate alternatives simply because the original result failed to reach a preferred significance threshold.
The uncertainty should come from the study design, research literature, data characteristics or a justified methodological concern.
Step 3: Choose plausible alternatives
A sensitivity analysis should probe realistic possibilities, not arbitrary scenarios selected to achieve a desired result.
For missing data, plausible alternatives might examine different assumptions about respondents who did not answer.
For regression, a justified check might examine a model that better accommodates a documented feature of the data while retaining the same target association.
For influential observations, the researcher may inspect how much a fitted estimate changes when a record is temporarily excluded as a diagnostic. That does not automatically justify permanently deleting the record.
Where possible, define sensitivity checks before seeing the final results. If additional checks are developed afterward, label them as post hoc or exploratory.
Worked example 1: Find a missing-response tipping point
Twelve eligible postgraduate students attend a hypothetical data-visualisation clinic. Nine rate its usefulness on a scale from 0 to 20:
10, 11, 12, 13, 14, 15, 12, 14, 16
Three attendees do not provide ratings. The target remains the mean for all twelve attendees, including those with missing ratings.
The sum of the observed scores is 117, so the observed-response mean is:
117 ÷ 9 = 13.0
That describes the nine responses. To examine the twelve-person mean, let m be the unknown average of the three missing ratings:
Twelve-person mean = (117 + 3m) ÷ 12.
Compare explicit missing-response assumptions
Define delta = m − 13, the difference between the missing-response mean and the observed-response mean. The same calculation becomes:
Twelve-person mean = 13 + 0.25 × delta.
| Assumed missing-response mean m | Delta relative to 13 | Mean across all twelve attendees |
|---|---|---|
| 13 | 0 | 13.00 |
| 10 | −3 | 12.25 |
| 7 | −6 | 11.50 |
| 4 | −9 | 10.75 |
For example, assuming m = 10 gives (117 + 3 × 10) ÷ 12 = 12.25. Assuming m = 7 gives (117 + 3 × 7) ÷ 12 = 11.50.
These are scenario calculations, not recovered observations, multiple-imputation output or evidence that nonrespondents actually had these values. The equal-mean scenario m = 13 is also an assumption, not a verified missingness mechanism. A dissertation must justify its scenario range using the study context or relevant evidence; these illustrative choices do not establish plausibility by themselves.
Identify the tipping point
Suppose a criterion of at least 12 out of 20 was chosen before examining the results. This threshold is a teaching illustration, not a validated definition of success.
Solve for the missing-response mean at which the twelve-person mean equals 12:
(117 + 3m) ÷ 12 = 12
117 + 3m = 144
m = (144 − 117) ÷ 3 = 9.0
The conclusion that the group mean meets the criterion holds when the three missing ratings average at least 9.0. It fails when their mean is below 9.0. Equivalently, the tipping point is delta = -4.0.
For the stronger claim that the group mean exceeds 12, the missing-response mean must be greater than 9.0; equality is not enough. Report the criterion precisely.
This gives the reader a concrete question: is an average below 9 among the missing ratings sufficiently credible to qualify the conclusion? The observed data alone cannot answer that question.
Separate scenarios, bounds and sampling uncertainty
The scale permits values from 0 to 20. Its mathematical extremes give:
Lower bound: (117 + 3 × 0) ÷ 12 = 9.75.
Upper bound: (117 + 3 × 20) ÷ 12 = 14.75.
The range from 9.75 to 14.75 is a set of deterministic bounds for this particular group, not a confidence interval. Extreme bounds need not be substantively plausible scenarios. Neither the bounds nor the tipping-point calculation accounts for sampling uncertainty when generalising beyond these attendees.
To reproduce the scenarios in a spreadsheet, enter an assumed missing-response mean in B2 and use =(117+3*B2)/12. Enter each justified assumption separately and retain the formulas. This calculation does not supply a statistical missing-data model or valid inferential standard errors.
For the wider choice among complete-case, imputation and model-based approaches, see our guide to handling missing survey data.
Worked example 2: Examining the influence of an unusual observation
Suppose a hypothetical study records the number of hours nine participants spent completing a training activity:
3, 4, 4, 5, 5, 6, 6, 7, 23
The total is 63 hours.
The mean is:
63 ÷ 9 = 7.0 hours
The median is the fifth ordered value: 5.0 hours.
One participant reported 23 hours, substantially more than the others.
Should that observation be deleted?
Not without investigation.
The recorded value may be a genuine experience, an entry error, a different interpretation of the question or evidence that participants faced very different circumstances.
Examine the influence of the unusual record
As a diagnostic, temporarily calculate the average of the remaining eight observations:
3 + 4 + 4 + 5 + 5 + 6 + 6 + 7 = 40
40 ÷ 8 = 5.00 hours.
| Summary | Value |
|---|---|
| Mean using all nine observations | 7.00 hours |
| Median using all nine observations | 5.00 hours |
| Mean using eight observations, excluding 23 | 5.00 hours |
| Difference between the two means | 2.00 hours |
The full-sample mean is substantially higher than the mean calculated after omitting the unusual observation.
This indicates that the observation is influential for this descriptive summary.
But it does not prove the value is wrong.
More importantly, excluding a genuine participant changes whose data contribute to the summary. The eight-person mean is therefore a diagnostic comparison, not automatically an equally valid estimate of the original nine-person mean.
If the recorded 23 hours can be confirmed as correct, a defensible response may be to retain it, report the mean and median, explain the variation, and discuss why the mean is sensitive to the highest observation.
If independent source evidence establishes a recording error, the researcher should correct it using a documented procedure rather than choosing whichever value creates a preferred outcome.
Our data-cleaning guide explains how to distinguish verified errors from legitimate unusual records.
Model dissertation paragraph
Completion times ranged from 3 to 23 hours, with a mean of 7.00 hours and a median of 5.00 hours across nine hypothetical observations. An influence check showed that the mean of the other eight observations was 5.00 hours. This comparison indicates that the full-sample mean is sensitive to the highest recorded value. The analysis alone does not justify exclusion: the validity of the observation must be assessed against the original record and study protocol. If retained, the variation should be reflected in the descriptive interpretation rather than concealed.
This example demonstrates an influence diagnostic. More formal regression models require diagnostics appropriate to their estimator, data structure and research question.
Worked example 3: Planning a sensitivity check for regression
Consider a hypothetical dissertation examining whether practice hours are associated with chart-interpretation scores.
The primary regression model includes practice hours as the predictor and chart-interpretation score as the outcome.
Before fitting the model, the researcher identifies two concerns:
Chart-interpretation scores may have a nonlinear relationship with practice hours.
A small number of observations may be highly influential.
A justified analytical plan might include:
| Analysis | Purpose | Interpretation boundary |
|---|---|---|
| Primary prespecified model | Estimate the planned association | Depends on the model's assumptions |
| Diagnostic inspection | Examine functional form, residual patterns and influence | Identifies potential concerns; does not itself provide an alternative conclusion |
| Appropriately specified nonlinear model | Examine whether the assumed linear relationship is consequential | The same substantive association may require different numerical summaries |
| Influence analysis | Examine dependence on highly influential observations | Removing valid cases changes the analysis sample and may change the target |
The researcher should identify which comparisons truly address the same target quantity and which are supplementary descriptions.
For example, if a nonlinear model is used, a single linear-regression slope is not automatically comparable with every coefficient in the nonlinear model. The researcher may instead need to compare appropriately defined predictions or contrasts at specified values of the predictor.
It would be misleading to present fabricated coefficients and p-values merely to make the example look complete. An actual dissertation must calculate estimates from the real dataset, including appropriate uncertainty measures.
Our guide to interpreting effect sizes explains why the magnitude of an association deserves attention alongside its statistical significance.
Step 4: Keep the research question consistent
A common analytical mistake is to treat two different estimates as though they answer the same question.
For example:
The mean among all eligible participants is different from the mean among complete responders.
A model for final assessment score is different from a model for whether students passed.
An adjusted association is not necessarily equivalent to an unadjusted association.
A result for undergraduate students alone may not describe the same population as a result for all university students.
Such comparisons can still reveal useful information, but their interpretation must reflect the changed target.
In stricter sensitivity-analysis frameworks, including ICH E9(R1), changing the quantity being estimated may make an analysis supplementary rather than a sensitivity analysis of the original estimator.
Kahan and colleagues' estimands primer illustrates how to align a question, its target quantity and the analytical method in clinical trials. This alignment principle is useful here without treating trial-specific guidance as a universal dissertation rule.
This distinction is particularly important in studies making causal claims. A different model does not automatically reveal the same causal effect, especially when assumptions about selection, confounding or post-treatment variables change.
Step 5: Compare magnitude, direction and uncertainty
A sensitivity analysis should not be reduced to asking whether the p-value remains below .05.
Instead, examine:
Direction: Does an association remain positive or negative, or does it change direction?
Magnitude: Does the estimate remain similar enough to support the same substantive interpretation?
Precision: Do confidence intervals or other relevant uncertainty measures change materially?
Assumptions: Which scenario is more credible, and why?
Decision relevance: Would the result support the same practical or theoretical conclusion?
A p-value moving from .049 to .051 does not by itself establish a dramatic scientific reversal. Equally, two estimates may both be statistically significant yet differ enough in magnitude to change their practical implications.
When confidence intervals are appropriate, use a method matched to the data and estimator. Our confidence-interval guide explains how to report uncertainty without treating an interval as a guarantee.
Step 6: Decide whether the conclusion is robust
The conclusion may be reasonably robust when several justified checks produce substantively similar estimates and interpretations.
However, agreement across checks does not prove that all assumptions are valid. Two models may share the same unmeasured confounding, measurement error or selection bias.
A result is sensitive when its substantive interpretation changes under a meaningful alternative assumption. That is valuable information rather than a failed research project.
Consider three possible outcomes:
| Pattern | Responsible interpretation |
|---|---|
| Estimates remain similar across credible alternatives | The result appears reasonably stable under the assumptions examined |
| Estimates differ modestly but support a similar substantive interpretation | The conclusion may be cautiously maintained, with the variation disclosed |
| Estimates or interpretations change materially | The conclusion depends on consequential assumptions and should be qualified |
Robustness is always conditional on what was tested. No finite set of checks can guarantee that every possible source of bias has been eliminated.
Step 7: Report the sensitivity analysis in your dissertation
A well-reported sensitivity analysis explains what was planned, why the check was needed, what changed, how the results compared and what the differences mean.
The STROBE statement recommends reporting applicable sensitivity analyses and other additional analyses in observational research. Its explanation and elaboration document also discusses how to describe these decisions in the methods and results.
Example methodology paragraph
The primary analysis was specified to address the study's main research question. Sensitivity checks were selected to examine assumptions judged likely to affect the interpretation, including incomplete responses and potentially influential observations. Each check was documented with its rationale, analytical sample, method and relevant assumptions. Alternative results were interpreted in relation to the original research question rather than selected according to statistical significance. Analyses addressing a different population or outcome were distinguished from checks of the primary estimator.
This is a model of methodological writing, not a claim that these procedures have been conducted.
Example results paragraph
The observed mean clinic-usefulness rating was 13.0 among nine respondents. Three of twelve eligible attendees did not provide ratings. Under illustrative missing-response means of 13, 10, 7 and 4, the twelve-person mean was 13.00, 12.25, 11.50 and 10.75, respectively. The criterion of at least 12 was met only when the missing-response mean was at least 9.0. The scale's deterministic bounds were 9.75 to 14.75; these were not confidence intervals or estimates of the missing responses.
Example discussion paragraph
The sensitivity illustration shows that the group-level conclusion depends on assumptions about item nonrespondents. The observed responses alone do not establish that all twelve attendees met the mean-rating criterion. The dissertation should explain whether missing-response means below the tipping point are credible in context and qualify its conclusion accordingly. Improved response procedures could reduce missingness, but additional responses would not by themselves resolve every source of selection or measurement bias.
A strong dissertation does not copy these paragraphs as evidence that checks occurred. The language must describe the actual study and verified analyses.
Where should sensitivity analysis appear in a dissertation?
The placement depends on the department's required structure.
In the methodology chapter, explain the planned checks, their assumptions and how each addresses a specific risk.
In the results chapter, present the calculated comparisons, analytical sample sizes, estimates and appropriate uncertainty measures.
In the discussion chapter, explain how the checks strengthen, qualify or challenge the interpretation of the primary findings.
Additional tables, code and technical details may belong in an appendix where permitted.
Avoid repeating every numerical result in three chapters. Give each section a distinct purpose.
How many sensitivity analyses are enough?
There is no universal required number.
A single carefully designed sensitivity analysis can be more informative than numerous poorly justified alternatives.
Prioritise the assumptions most capable of changing the conclusion. In a small descriptive survey, nonresponse may be a central concern. In observational regression, uncontrolled confounding or measurement error may be more consequential.
Some assumptions require specialised approaches, including quantitative bias analysis. For example, Greenland's foundational discussion of sensitivity analysis explains how unmeasured confounding, misclassification and selection bias can affect observational findings.
These methods should not be improvised without understanding their assumptions. A straightforward, transparent limitation may be preferable to an incorrectly implemented complex procedure.
Common mistakes in sensitivity analysis
Trying alternative models until one becomes significant
This converts analysis into result selection. Explain exploratory changes transparently and do not retrospectively present them as prespecified tests.
Removing every outlier
An unusual observation may be correct. Investigate its source, influence and relevance before deciding how to handle it.
Treating a different population as the same target
Complete-case and full-sample estimates may answer different questions. State precisely who contributed data.
Calling every alternative model a robustness check
A different model may estimate a different effect. Compare like with like, or describe the analysis as supplementary.
Reporting only the most favourable result
Readers need the primary analysis and relevant alternative findings, including those that weaken the proposed interpretation.
Comparing p-values without comparing estimates
Scientific interpretation depends on magnitude, uncertainty, assumptions and substantive meaning—not merely threshold crossings.
Claiming that consistent results prove validity
Several models can share the same underlying bias. State which uncertainties remain untested.
Inventing tests or software outputs
Never claim that a sensitivity analysis, regression diagnostic, confidence interval or statistical procedure was performed unless it actually was and the result was verified.
Frequently asked questions
Is sensitivity analysis mandatory in every dissertation?
No. Its relevance depends on the design, analytical assumptions and institutional requirements. It is particularly useful when important conclusions depend on uncertain choices or incomplete information.
Is sensitivity analysis the same as statistical significance testing?
No. It examines the dependence of results on assumptions or decisions. Some checks involve statistical tests, but sensitivity should be evaluated through estimates, uncertainty and substantive interpretation.
Can I conduct sensitivity analysis in Excel?
Simple scenarios, bounds and descriptive recalculations can be examined in a spreadsheet. More complex regression, missing-data or bias analyses may require appropriate statistical software and methodological expertise. The software does not determine whether the comparison is scientifically meaningful.
Can I conduct sensitivity analysis after the main analysis?
Yes, but distinguish prespecified checks from those developed after observing the data. Report the reason for any post hoc analysis and avoid portraying exploratory findings as prior predictions.
What happens if sensitivity analysis changes my conclusion?
Report the change. It may show that the original conclusion relied on a fragile assumption. A carefully qualified finding is more credible than an apparently decisive result that conceals consequential uncertainty.
Do qualitative dissertations use sensitivity analysis?
Some qualitative approaches examine alternative interpretations, discrepant cases and the influence of analytic decisions. These procedures are not necessarily equivalent to numerical sensitivity analysis. Use terms consistent with the qualitative methodology rather than importing statistical labels without justification.
Does a sensitivity analysis remove bias?
No. It can show how an inference changes under specified assumptions, but it does not automatically identify the correct assumption or eliminate unmeasured bias.
Sensitivity-analysis checklist before submission
Is the primary research question and target estimate clearly defined?
Have I identified the assumption or decision being challenged?
Is each alternative methodologically justified?
Have I distinguished sensitivity checks from analyses that change the target question?
Are sample sizes, exclusions and denominators disclosed?
Are all numerical results actually calculated and verified?
Have I compared magnitude, direction and uncertainty rather than p-values alone?
Have I reported results that contradict or weaken the preferred conclusion?
Are prespecified and post hoc checks distinguished?
Have I described which uncertainties remain unaddressed?
Does the final interpretation remain consistent with the study design?
Have I followed my department's methodological and reporting requirements?
Final takeaway
Sensitivity analysis does not exist to rescue a disappointing result. Its purpose is to show how much confidence a reader can place in a conclusion when important assumptions are examined.
Begin with a clearly defined research question. Identify a genuine vulnerability. Choose plausible alternatives. Compare the results carefully, keeping the target quantity consistent wherever possible. Then explain what remained stable, what changed and what cannot be established.
A dissertation becomes more credible when it makes uncertainty visible rather than pretending that one calculation settles every question.
References and further reading
Greenland, S. (1996). Basic methods for sensitivity analysis of biases. International Journal of Epidemiology, 25(6), 1107–1116. DOI: 10.1093/ije/25.6.1107-a.
International Council for Harmonisation. (2019). Addendum on estimands and sensitivity analysis in clinical trials to the guideline on statistical principles for clinical trials E9(R1). Final Step 4 version, adopted 20 November 2019.
Kahan, B. C., Hindley, J., Edwards, M., Cro, S., & Morris, T. P. (2024). The estimands framework: A primer on the ICH E9(R1) addendum. BMJ, 384, e076316.
von Elm, E., et al. (2007). Strengthening the reporting of observational studies in epidemiology (STROBE) statement: Guidelines for reporting observational studies. BMJ, 335(7624), 806–808.
Vandenbroucke, J. P., et al. (2007). Strengthening the Reporting of Observational Studies in Epidemiology (STROBE): Explanation and elaboration. PLoS Medicine, 4(10), e297.
Editorial note: All worked examples are hypothetical and intended for learning. Students should use their actual data, approved methodology, verified analytical outputs and institutional guidance.
Need help strengthening your dissertation? Explore Acadelyra's academic support services for assistance with research planning, methodological clarity, statistical-result presentation, editing and academic writing support that respects institutional integrity requirements.