Did you know? A city can publish thousands of datasets and still leave someone unable to answer a simple question: How many people use this service, has that changed, and what does the number actually mean? The information may be available to download. The answer may remain out of reach.
Public data portals fail when publication is mistaken for communication. A CSV file is not automatically a useful explanation, a chart is not automatically a fair comparison and a timestamp is not automatically evidence that the underlying facts are current.
One-minute answer: A successful public data portal lets a reader find the right information, understand where it came from, identify what the measures mean, check the time period and limitations, reuse the data appropriately and raise questions when something seems wrong. The challenge is not merely to put files online. It is to connect public evidence to questions that real people can ask.
What is a public data portal?
A public data portal is a website or service through which government bodies and other public institutions make selected datasets discoverable and accessible. Depending on the portal, people may explore statistics, download files, view charts, use an application programming interface or read supporting descriptions.
Singapore’s data.gov.sg is an example: it provides a central place to search for datasets published by participating government agencies. That is a useful public service. Yet no one portal can automatically make every dataset complete, suitable for every purpose or free from interpretation.
The interesting question is not whether the portal has a search box. It is whether somebody can move reliably from a real question to an answer that can be explained and checked.
The difference between available data and usable knowledge
Imagine a parent asking how school enrolment changed over five years. A portal returns several files with unfamiliar column labels, different classifications and missing notes on whether they describe enrolment at the beginning or end of a year.
The files are available. The comparison is not yet trustworthy. The parent needs to know which population is counted, whether the definition changed and which dates refer to the same measurement.
Usable knowledge requires a bridge between records and meaning. The bridge includes descriptions, definitions, context, quality checks and an honest account of what the evidence cannot support.
A dataset needs to answer six questions
Before using a public dataset, a careful reader should be able to answer six questions:
- What? Which real-world thing is being measured, and in what units?
- Who? Which organisation produced it and who maintains it?
- Where? Which geographic areas, institutions or groups are covered?
- When? What period does each observation represent, and when was the source updated?
- How? How was the information collected, calculated and checked?
- What is missing? Which exclusions, uncertainties or interpretation limits matter?
If a portal cannot help a user answer these questions, the obstacle may be missing metadata rather than missing data itself.
Why file names are not enough
A download called population_latest_final_2.csv may contain valuable information, but its name does not explain whether it refers to residents, citizens, households or a statistical estimate. Nor does it tell the reader whether a revision replaced earlier figures.
Good metadata gives a meaningful title, clear description, provenance, data dictionary, units, coverage, frequency, licence and an appropriate contact or feedback channel. These are not decorations. They prevent readers from making confident claims based on the wrong interpretation.
The W3C’s Data on the Web Best Practices treats metadata, data quality, provenance, versioning, licensing and access as parts of responsible publishing. The practical purpose is to help people use data as intended rather than guess at its meaning.
Why the same word can mean different things
Suppose two datasets both use the word users. One counts people who registered. Another counts people who logged in during the month. A third counts successful transactions, where one person can make several.
Dividing one measure by another as if they share a denominator may produce a persuasive but meaningless percentage.
This is the failure of definition drift. The columns look compatible, so the user assumes they describe the same thing. The repair is a data dictionary that spells out inclusion rules, measurement method and changes across releases.
Ask whether two measures truly refer to the same population, event and time period before joining them.
A worked example: the suspicious attendance trend
Consider a fictional education dataset. The published numbers say attendance rose from 91 to 96 per cent. A reader might celebrate an impressive improvement.
But the first year measured full-day attendance, while the second counted attendance in the first lesson. The percentage changed in part because the method changed. The figures cannot be treated as one continuous trend without adjustment or a prominent warning.
The lesson is not that public statistics are untrustworthy. It is that interpretation needs a stable definition. A portal should disclose a break in series clearly enough that readers do not confuse a measurement change with a real-world change.
Missing, zero and not applicable are different
A blank cell can mean that the agency did not collect a value, the result was suppressed, the record was invalid or the data genuinely did not apply. A zero means the measured quantity was zero, provided that is how the dataset defines it.
Replacing every blank with zero can transform uncertainty into a false claim. It may also distort averages, trends and the apparent differences between neighbourhoods.
Public datasets should use documented missing-value conventions. Readers should preserve those distinctions rather than quietly making the spreadsheet look complete.
Updated yesterday does not always mean current yesterday
A portal may show that a dataset was updated yesterday. That might mean the file was republished, corrected or refreshed through an automated process. It does not necessarily mean the observations describe yesterday’s conditions.
Look for the reference period, date collected, release date and revision date. These answer different questions.
A monthly series can legitimately arrive several weeks after a month ends. A yearly figure may be published much later. The important task is to make those lags visible and not present old observations as live measurements.
The importance of version history
Suppose a local-services table is revised after an agency discovers duplicate entries. A student downloads the original file for an assignment; a researcher downloads the corrected version two weeks later. Their totals no longer match.
Without revision information, both readers may believe the other made a mistake. A useful portal should show what version was used, when it changed and, where feasible, which corrections were made.
Singapore’s data.gov.sg documentation describes dataset-history features for supported datasets. Readers should check which history and change records are actually available for the resource in question, rather than assume every dataset has identical version tracking.
The problem of a graph without its denominator
A chart says complaints rose from 200 to 240. Is that evidence that service quality worsened? Perhaps. But suppose the number of transactions rose from 10,000 to 20,000 over the same period. The complaint rate would then have fallen from two per cent to 1.2 per cent.
Those are fictional numbers to illustrate why context matters. Neither the raw total nor the rate answers every question, but they tell different stories.
A trustworthy portal offers enough surrounding information to let readers distinguish growth in the underlying population from growth in the rate of an event.
Why averages can hide a neighbourhood
A citywide average can improve while conditions worsen in some neighbourhoods. A national statistic can be accurate and still be a poor guide to one school, family or household.
Where disclosure is safe and justified, geographic or demographic breakdowns may help. However, very small groups can introduce privacy risks and unstable estimates. More detailed publication is not always better.
A good portal explains the level at which the measure is reliable. It also avoids presenting one summary statistic as the complete experience of every resident.
What a map can conceal
Maps offer a quick way to understand locations, but visual conventions can change the apparent story. Colour ranges, geographic boundaries, missing areas and population differences all affect interpretation.
Suppose a map shades districts according to total complaints. A large district may appear worse simply because more people live there. Showing rates alongside totals may offer a fairer view, if both measures are appropriate.
When using mapped data, ask what the colours mean, whether areas are comparable and whether the geographic unit suits the question.
A portal that only experts can use
A technical download page may be sufficient for a researcher who knows how to clean data and write code. It may be inaccessible to a student, a parent, a small-business owner or someone using assistive technology.
Public access should not be confused with expert-only access. Helpful features may include a plain-language summary, data dictionary, accessible tables, keyboard-friendly navigation, examples and explanations of commonly misunderstood fields.
Accessibility is not a separate luxury feature. It determines who can use information paid for or maintained for public benefit.
The language barrier inside technical English
Even when a portal is presented in English, specialist terms such as seasonally adjusted, index base, provisional, geospatial geometry or administrative boundary can become obstacles.
A glossary should explain what these terms mean in the context of the dataset. The best description does more than translate jargon: it tells the reader when a measure is appropriate, when it is not and which question it can reasonably answer.
For multilingual communities, accessible explanations and translation may improve reach, but any translated statistical definition must preserve the original meaning and caveats.
Downloading a CSV is only the beginning
A CSV is useful because many tools can read it. Yet a file may have inconsistent dates, ambiguous character encodings, undocumented categories or changes in column names between versions.
A downloadable spreadsheet is not a substitute for explaining the schema. A well-designed portal documents the format, field types, allowed values and common cleaning steps.
Where multiple formats exist, the published files should represent the same underlying definitions. Otherwise people may receive different answers depending on which download button they selected.
APIs can fail even when the website looks fine
An application programming interface allows software to access datasets programmatically. This helps researchers, developers and public services use recurring information without manually downloading every file.
But APIs create their own responsibilities: clear documentation, stable identifiers, pagination, error responses, change notices and appropriate access limits. If an endpoint silently changes the meaning of a field, dependent applications may continue running while producing misleading results.
The data.gov.sg developer guide explains its API approach. Developers should follow the current documented behaviour rather than assume older examples remain valid forever.
Why a stable link matters
A teacher builds a lesson using a dataset URL. A community group shares a reference in a report. Months later, the address changes and both references break. The information may still exist, but its public trail has been damaged.
Stable dataset identifiers, redirects where appropriate and meaningful revision notes make evidence easier to revisit. They also enable citations that another reader can independently check.
Public data is most valuable when it survives beyond a single news cycle or website redesign.
A fictional transport-data mistake
A student wants to compare bus-stop accessibility in two areas. The first file contains bus-stop locations; the second counts registered routes. The student draws a conclusion about the number of buses that arrive in each area.
The conclusion does not follow. A stop location is not a service frequency, and a route count does not directly establish actual arrival times, reliability or passenger accessibility.
A responsible analysis starts with the question and identifies the exact measure needed. If the relevant data is absent, the honest result is that the available files do not answer the question.
Evidence is not the same as an explanation
A chart can reveal a pattern, but it rarely identifies the cause by itself. If energy consumption declines, that could reflect efficiency, weather, economic changes, population differences or a mix of factors.
Causal claims require additional evidence or a suitable research design. A public portal should help users observe accurately without encouraging them to treat every pattern as a discovered cause.
The most useful reader habit is to separate description, comparison and explanation. Each needs a different strength of evidence.
Why public datasets may reasonably omit information
Not every record should be published in detail. Personal information, confidential operations, security concerns and other legal or ethical restrictions can require aggregation, suppression or exclusion.
A privacy-preserving portal should make relevant exclusions understandable without disclosing the information it is protecting. Users should not infer that an omitted group has a value of zero.
The goal is responsible openness: publish useful information while respecting people whose lives appear in the data.
Licences are part of knowing what can be done
Finding a dataset online does not automatically answer whether it may be reused in a commercial product, modified or republished. The relevant licence and any dataset-specific conditions matter.
Singapore’s Open Data Licence outlines permitted use and important requirements, including attribution and restrictions against suggesting official endorsement. Always consult the applicable licence itself and current terms.
A good portal makes the licence easy to find before a user builds an application or distributes a chart.
When feedback falls into a black hole
A researcher notices an inconsistent date. A parent sees an unclear column. A developer finds that an API result differs from the documentation. If the portal offers no meaningful feedback route, the problem may be repeated by hundreds of later users.
A useful feedback process lets readers flag unclear definitions, suspected errors or broken downloads and provides some indication of ownership. Not every suggestion can be accepted, but unanswered problems should not become permanent parts of the public record.
Trust improves when correction is possible and visible.
How a public data portal can be audited
Begin with one realistic question, not a checklist of visual features. Can a secondary-school student find the correct dataset for that question? Can they identify the producer, period, measurement and limitations? Can they explain the result to another person?
Then test whether the figures can be independently reproduced from the download or API, whether the cited link remains accessible and whether a correction route exists.
Audits should include people with different levels of technical experience and access needs. A system tested only by expert analysts may miss the frustrations of its widest audience.
A simple five-stage data-to-answer routine
Use the following process when working with a public dataset.
- Find: State the question in one sentence and locate a dataset that actually measures the relevant subject.
- Define: Read the description, units, coverage, population and methodology before calculating.
- Check: Inspect dates, missing values, revisions and any change in definition.
- Calculate: Use appropriate denominators and record enough steps to reproduce the result.
- Explain: State what the evidence shows, what it cannot prove and where another reader can verify the source.
This is a reading method, not advanced software. It works for a classroom exercise, community inquiry or research project.
An example of transparent reporting
Instead of writing, ‘Our neighbourhood has twice the transport problem,’ try a defensible form: ‘In the published dataset, this area recorded X complaints during period Y under definition Z. The other area recorded W. We have not adjusted for differences in population or service volume, so the comparison does not establish the relative risk faced by an individual commuter.’
That sentence is longer, but it is more accurate. It distinguishes a measured count from a claim about people and identifies what remains unknown.
Good explanation does not weaken evidence by admitting limits. It makes the evidence stronger by showing where its boundaries lie.
What public agencies can improve first
A portal improvement programme does not need to start with an expensive redesign. Fix the fields that most often confuse readers: meaningful descriptions, current ownership, clear dates, usable column definitions, stable access and obvious licence terms.
Prioritise datasets that receive many questions or support important public decisions. Give every dataset a way to report problems and record significant changes. Make examples available for common tasks without presenting them as the only valid interpretation.
An agency that reduces avoidable misunderstandings may create more public value than one that simply adds another hundred poorly explained files.
What schools and tutors can teach with open data
Public datasets provide opportunities for genuine mathematics, science, geography and civic reasoning. Students can practise reading axes, comparing percentages, asking how evidence was collected and distinguishing correlation from causation.
The most valuable classroom exercise sometimes ends with ‘We cannot answer that from these data.’ That is not a failure to complete the question. It is a success in understanding what a source does and does not establish.
A tutor can scaffold the process: identify the population, decode the units, reconstruct a simple calculation and ask the learner to explain the limitations in ordinary language. The goal is independent reasoning rather than copying an attractive chart.
What happens when AI summarises a dataset badly?
AI tools can help explain unfamiliar terminology or propose chart ideas, but generated descriptions may invent definitions, overlook missing values or treat an outdated file as current.
Do not paste confidential information into an unapproved system. When using an approved AI assistant for public data, give it verified field definitions and check all numerical claims against the source.
A fluent paragraph is not provenance. The reader must still be able to locate the actual dataset, its reference period and the calculation behind the answer.
Frequently asked questions
Is a public dataset trustworthy because it comes from government? Government provenance is important but does not mean every dataset fits every question. Check definitions, methods, timing and limitations.
Does the latest upload date prove the values are current? No. A file can be republished without new underlying observations. Inspect the reference period.
Are blanks the same as zero? Not unless the dataset explicitly defines them that way. Missing or suppressed values must not be silently converted into zero.
Can I reuse everything I find? Review the applicable licence and conditions for each dataset. Public availability and permission to reuse are related but different questions.
Does a chart prove what caused a trend? No. Most descriptive charts show relationships or changes, not causal mechanisms by themselves.
Must every public dataset include a live API? No. Different uses need different methods of access. The key is that the published format and documentation make the information usable for its intended audience.
What success actually looks like
A successful public data portal is a bridge from evidence to understanding. The user can find the right source, see what was counted, identify when the count applies, recognise what is missing and explain a conclusion with appropriate caution.
A portal with fewer but well-described datasets may answer more real questions than one with thousands of unexplained files. The point is not simply that data exists online. It is that people can use it without being misled.
Sources and further reading
data.gov.sg — Singapore’s Open Data Portal. data.gov.sg — Developer API Overview. data.gov.sg — September 2025 Release Notes. Singapore Open Data Licence. W3C — Data on the Web Best Practices. These resources explain the roles of access, documentation, versioning, licensing and data reuse. Examples in this article are fictional unless expressly identified otherwise.
Continue learning with eduKateSG
Read How Data Systems Fail, How Statistics Mislead, How Digital Accessibility Fails and the How Things Fail series.