Learn AI /Data and Infrastructure /Missing data
Missing data
- Starter
- Data and Infrastructure
- Programme & data teams
- Human rights practitioners
Also called: missingness, gaps in data.
Back to the milestone you came from
In five seconds
What it is
Some values that should be present in a dataset are absent, leaving blanks that can distort analysis and model behaviour.
In plain words
Some information that should be in the dataset is missing, and that can change what the analysis or model produces.
An everyday example
A warehouse worker in Mombasa checks a shipping manifest against the boxes that arrived. Some lines on the manifest have the item name and quantity, but the destination shelf is blank. On other lines, the quantity is blank because a label was torn off during transport. The supervisor can still process the delivery, but must decide whether to set those boxes aside, mark them as unknown, or infer the missing detail from other documents. If too many blanks affect one supplier, the stock report at the end of the day may give a distorted picture.
A manifest usually has a fixed set of boxes and fields, while real datasets can have many variables and more complex patterns of missingness. Also, in data work, filling a gap can change statistical relationships in subtle ways; a warehouse guess mainly affects one shipment record. The analogy helps with the idea of blanks and handling choices, but not with the mathematics of bias or uncertainty.
Remember thisA blank is not neutral. Check what is missing and why before deleting, flagging or estimating values, because the pattern can bias the result.
Understand it
Understand it
Why it matters
Missing data can quietly change conclusions. A chart, average, risk score, or model may look precise even when important values are absent. For non-specialists, this matters because decisions about funding, staffing, outreach, protection, or evaluation may rest on those outputs. The key question is not only how much data is missing, but who or what is missing. If gaps are concentrated in one population, language group, location, or period, the result can be systematically unfair or inaccurate. Good decisions require asking whether the blanks are harmless, informative, or a sign that the collection process itself needs attention.
How missing data changes results
Show how missing values enter a dataset, the main handling choices, and why different choices can change the final result.
01
1. Incomplete data enters
The dataset includes records with one or more empty values, such as a missing age, location, or date.
Draw a simple table with 4 rows and 4 columns. Leave several cells blank and highlight them with a dashed outline. Add a caption: 'Some values are absent'.
02
2. Check the pattern
Ask whether the gaps are scattered randomly or concentrated in one group, source, language, region, or time period.
Show the same table with coloured circles around blanks. Add small labels above groups of rows such as 'urban' and 'rural' to show blanks clustering more in one group.
03
3. Choose a handling method
Common choices are to remove incomplete records, keep them with an 'unknown' marker, or estimate the missing value from other data.
Split the flow into three side-by-side boxes labelled 'drop rows', 'mark unknown', and 'estimate value'. Use arrows from the checked table to each box.
04
4. Compare outputs
Each handling choice can change totals, averages, and model behaviour. The safest option depends on the purpose and the risks.
Merge to a final panel with three mini-results: different averages, different counts, and a note that one group may now be under-represented. Add a warning icon with text: 'Choice affects outcome'.
A left-to-right flow diagram with four boxes. It starts with a small table containing blank cells, then moves to a check for where gaps occur, then splits into handling choices, then ends with different possible outputs. Read from left to right to see that the same incomplete input can lead to different results depending on how the gaps are treated.
Hover or tap each handling method to highlight one possible path from incomplete data to output and reveal a short note on the trade-off.
Describe this without the diagram
The diagram explains that missing data is not only about blank cells but about how those blanks are handled. First, a table contains several empty cells. Second, the viewer checks whether the blanks are random or clustered in a particular group, such as one region having more gaps. Third, the process branches into three handling methods: dropping incomplete rows, marking the value as unknown, or estimating a value from other available information. Fourth, the branches lead to different outputs, showing that totals, averages, and model behaviour can change depending on the choice. The core message is that the same incomplete dataset can produce different results because handling missing values is an analytical decision.
With reduced motion: Show the same four stages as a static left-to-right diagram with the three handling choices displayed side by side and a fixed note that different choices lead to different outputs.
A dataset is a collection of information organised so people or software can analyse it. Sometimes some entries are missing. A survey may have no age for one person. A clinic record may have no follow-up date. A spreadsheet may have blank cells because the information was never collected, was lost, or could not be matched.
What goes in is incomplete data. What happens next depends on how those gaps are handled. People may remove incomplete rows, fill in a value based on other information, add a special marker such as “unknown”, or redesign the analysis so it can cope with blanks. Each choice changes the result in a different way.
What comes out can be useful, misleading, or unfair. If the missing values are random, the damage may be small. If they are concentrated in one group, place, language, or time period, the output can become biased. For example, if rural areas have more missing health records, a summary may understate their needs.
This is different from noisy data. Noisy data contains values that are present but wrong, inconsistent, or hard to interpret. Missing data means the value is absent. It is also different from class imbalance, where some categories have far fewer examples than others even if the records are complete.
How “Checking a shipping manifest against a delivery” maps onto the real thing
| In the story | In the AI system |
|---|---|
| blank fields on the manifest | missing values in a dataset |
| setting boxes aside or marking unknown | dropping incomplete records or encoding a missing marker |
| inferring details from other documents | estimating missing values from other available data |
| distorted stock report | biased analysis or model output |
A worked example
Before progress report clearance, a programme officer in a country office reviews completion rates submitted by 10 implementing partners.
Starting from: Reported completion rates: 100%, 90%, 85%, blank, 75%, 70%, blank, 60%, 50%, 40%.
- The officer identifies 2 missing values among the 10 expected returns.
- Dropping those records gives an available-case average of 71.25% across the 8 reported values.
- The tracking log shows that both blanks concern partners that requested workplan extensions. The officer therefore cannot safely assume the values are missing completely at random.
- The officer requests the missing returns and checks sensitivity rather than filling the blanks without evidence. If each missing value could range from 0% to 100%, the overall average could range from 57% to 77%.
Result: The progress report labels 71.25% as an available-case estimate, not the completion rate for all partners. It also records the follow-up and notes that the estimate may be too high if delayed partners had lower completion.
In a United Nations workplace
At the annual workplan review, a legal officer in a regional bureau checks the results framework before the progress report goes to the front office. Several country-office entries show a review deadline but no completion status.
The blank statuses cluster in files returned for revision, so the officer cannot safely assume they are random or count them as completed. Checking the pattern, consulting the tracking log and following up with focal points prevents the progress report from overstating delivery.
Before you try thisDo not treat blanks as zero, completed or unchanged without evidence, and do not ask an AI tool to invent the missing status. Confidential or case-specific material must not be uploaded to a tool that has not been approved for it. Use approved systems and record any imputation or exclusion clearly.
Where you may hear it
- At an annual workplan review, when indicator fields are still blank
- During progress report clearance, when totals do not reconcile with partner returns
- In an evaluation office meeting on whether to impute unanswered survey items
- At a vendor demo for a dashboard that quietly excludes incomplete records
- During an internal audit review of gaps in administrative records
Avoid confusion
Avoid confusion
Often confused with
Noisy data
Data that contains errors, irrelevant variation or unclear signals, making patterns harder for people or models to detect reliably.
Class imbalance
A dataset has this problem when one class appears far more often than another, which can make overall accuracy look strong while a model misses rare cases.
Sampling (of data)
Choosing a smaller set of cases from a larger group so you can measure, inspect or train on it without using every case.
Common misconceptions
Common misconception: Blank cells are just a minor cleaning issue and can usually be ignored.
Correction: Sometimes they can be minor, but not always. If the missing values are concentrated in a particular group or type of case, ignoring them can shift averages, hide needs, or make a model perform worse for that group. The impact depends on the pattern, not only the number of blanks.
Common misconception: The best fix is always to delete any row with a missing value.
Correction: Deleting incomplete rows is simple, but it can remove a large share of data and can bias the result if the missingness is not random. In some cases, keeping an 'unknown' category or using a suitable estimation method is safer. The right choice depends on purpose, risk, and how the gaps arose.
Common misconception: Missing data and noisy data mean the same thing.
Correction: They are different problems. Missing data means a value is absent. Noisy data means a value is present but may be wrong, inconsistent, or imprecise. A blank age and an age typed as 250 both create problems, but they need different checks and remedies.
Quick check
Why can missing data be more serious than it first appears?
Missing data becomes especially risky when the gaps are systematic rather than random. If one group has more missing records, summaries and models can undercount or misrepresent that group. A dataset with blanks can still be usable, but only after checking the pattern and choosing a suitable handling method.
Go deeper
Go deeper
The simple version
Some information that should be in the dataset is missing, and that can change what the analysis or model produces.
The technical version
Missing data refers to absent values for variables intended to be observed. Handling it requires examining the pattern and mechanism of missingness. Data are missing completely at random when missingness is unrelated to observed or missing values. They are missing at random when, given observed data, missingness does not depend on the missing values. Otherwise, missingness may be not at random. Treatments include complete-case analysis, missing indicators, imputation, and models that accommodate missingness. The choice affects variance, bias, representativeness, and model performance.
Reviewed external resources
Primary sourceAI Risk Management Framework (opens in a new tab)NIST
The NIST AI Risk Management Framework helps readers connect data quality issues such as missing data to practical risk management, and suits professionals responsible for governance or assurance.
Intermediatelink checked 2026-07-31
Primary sourceOECD AI Principles (opens in a new tab)OECD
The OECD AI Principles provide high-level policy guidance on trustworthy AI, useful for readers who want the governance context rather than technical detail.
Intermediatelink checked 2026-07-31
EncyclopaediaMissing data (opens in a new tab)Wikipedia
Wikipedia gives a broad introductory overview of missing data concepts and common handling approaches, which suits beginners who want a quick starting point.
Intermediatelink checked 2026-07-31
VideoVideo explanations: Missing data (opens in a new tab)YouTube search
This YouTube search results page helps readers find video explanations of missing data in different teaching styles, and suits people who learn better by watching; it is a search page, not one definitive source.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Missing data (opens in a new tab)Reddit search
This Reddit search results page helps readers see common questions and informal discussions about missing data, and suits readers who want practical language and examples; it is a search page, not one definitive source.
search results, not one pagelink checked 2026-07-31
Where this came from
Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.
The one thing to remember
A blank is not neutral. Check what is missing and why before deleting, flagging or estimating values, because the pattern can bias the result.