Ask Mallard
Analysis method

Delta-adjusted tipping-point analysis

5 min read

Delta-adjusted tipping-point analysis asks how far missing outcomes must differ from MAR predictions before a prespecified conclusion changes.

Fit

When to use it, and when not to

Use it when

  • Missing outcomes could depend on unseen values even after recorded information is considered.
  • You can specify clinically plausible shifts and a conclusion to challenge before seeing the sensitivity results.

Not when

  • The outcome is undefined after death. Define an estimand for death rather than imputing an ordinary outcome. Read: The estimand
  • Your concern is unmeasured confounding rather than missing outcomes. Choose a sensitivity analysis for that bias. Read: Sensitivity analyses

Before you start

What you need

Decide these first
  1. Outcome scale and directionoutcomeTypeDefine when it is measured and whether higher values are better.
  2. A clinically justified delta gridState which missing outcomes shift, in which groups, and by how much.
  3. A conclusion fixed in advanceDistinguish crossing no mean difference, losing statistical significance, and losing a clinically useful effect.
It rests on these
  1. Start with MCAR, MAR and MNAR as distinct assumptions.How to check: Missing completely at random (MCAR) means gaps depend on no observed or unseen values. Under missing at random (MAR), recorded data explain gaps. Suitable multiple imputation under MAR is a starting analysis, not inherently superior. Missing not at random (MNAR) means unseen values still predict gaps. Delta specifies a departure from MAR.
  2. The shift is appropriate for the outcome.How to check: An additive shift suits some continuous outcomes. Binary or bounded outcomes need suitable models and shifts on a valid scale.
  3. Delta is a sensitivity choice, not a fitted fact.How to check: Use clinical judgment or external information to choose plausible ranges. Vary both arms when missingness in both could matter.
  4. Uncertainty is retained within each scenario.How to check: Apply the specified shifts within each imputed dataset, refit the analysis and pool estimates and uncertainty. Preserve observed outcomes.

Plain language

How it works

Fit the planned imputation model under MAR, then specify departures under MNAR. In a simple continuous-outcome pattern-mixture model, add delta to each missing outcome’s imputed value. A negative delta means missing outcomes are lower than their MAR predictions. Delta zero recovers the starting MAR analysis.

Find and interpret the tipping boundary

Repeat across a grid of deltas and refit the planned analysis within each imputation. Pool results at each grid point. Find where the chosen conclusion changes, and show the surrounding estimates and intervals. A boundary far outside a credible range suggests robustness only to those examined scenarios.

Three different assumptions about gaps
Three small diagrams of what missingness depends on. Missing completely at random: whether a value is missing depends on nothing in the data. Missing at random: it depends only on observed data. Missing not at random: it depends on the missing value itself.Completely at random: dependson nothingObservedValueMissing?At random: depends on observeddataObservedValueMissing?Not at random: depends on thevalue itselfObservedValueMissing?Observed: data that were recorded. Value: the measurement that may be missing. Missing?: whether itis. Each label is an assumption about what predicts missingness, chosen for a reason, not a testresult.
Conceptual diagram. It shows the structure of the idea, not data from any study.

Worked example

A worked example

Invented for illustration
Setting
A trial measures a symptom score where higher is better.
Question
What shift makes the unadjusted mean contrast reach zero?
Approach
Each arm has 100 people. Treatment has 80 observed scores with mean 54 and 20 missing scores imputed at mean 54. Control is complete, with mean 50.
Result
At delta -20, the missing-score mean becomes 34. Treatment’s mean becomes 50, so the mean difference is 0.
What it means
The point-estimate tipping point is -20 score units. This arithmetic does not locate a significance tipping point or compute an imputation interval.

In practice

Recent studies PubMed search · not vetted

Searching PubMed…

Watch out

Common mistakes

The one to remember

The data cannot identify delta or prove that a missing-at-random analysis is correct. A tipping point matters only relative to credible MNAR scenarios and a conclusion defined in advance.

  • Shifting observed outcomes along with missing ones.Instead: Leave observed values intact; shift only the missing outcomes defined by the scenario.
  • Presenting the tipping point as the true missing-data mechanism.Instead: Report the whole scenario grid and explain which shifts clinicians consider plausible.
  • Treating a point-estimate crossing as a significance crossing.Instead: State the criterion and examine its own estimates and intervals.

Inside Mallard

How Mallard uses this

Plans do not link to this method yet. You can find it in Learn and the Atlas.

Code example in Mallard: Delta-adjusted tipping-point analysis for missing outcomes (delta-adjusted-tipping-point).
Sample size: Mallard sizes the main comparison with two means. It does not size adjustment or any secondary analysis.
Atlas node analysis.delta_tipping_point, the id a plan cue records.
Code for this method
Loading…

Related Learn pages

Related designs and methods

Evidence

Sources & further reading

  1. Analysis of longitudinal trials with protocol deviation: a framework for relevant, accessible assumptions, and inference via multiple imputation (opens in a new tab) · J Biopharm Stat · doi 10.1080/10543406.2013.834911
  2. Fast tipping point sensitivity analyses in clinical trials with missing continuous outcomes under multiple imputation (opens in a new tab) · J Biopharm Stat · doi 10.1080/10543406.2022.2058525
  3. Sensitivity analysis for clinical trials with missing continuous outcome data using controlled multiple imputation: A practical guide (opens in a new tab) · Stat Med · doi 10.1002/sim.8569
Prepared with AI assistance for the owner-approved one-method wave 1. Indexed sources were checked by PMID, title, authors, year and DOI against PubMed on 2026-11-06. Record verification is not methodological approval. Carpenter and colleagues develop controlled imputation assumptions. Gorst-Rasmussen and Tarp-Johansen develop tipping-point calculations, including two-way shifts. Cro and colleagues explain delta shifts and uncertainty in practice.
Last source review: 2026-11-06
AI-generated educational content
Mallard uses reputable methodological and reporting sources, but this material may contain errors or omit important context. Verify important methodological decisions with the cited sources and, when appropriate, a qualified research methodologist or statistician.