Data analysts turn uncertain questions into evidence that others can inspect. Their work is not defined by dashboards alone. It involves deciding what to measure, understanding how records were collected, testing calculations, protecting people and communicating conclusions without false certainty.
Our worked case is a fictional tutoring programme comparing two ways to arrange learning support. Route A completed 12 out of 20 eligible attempts; Route B completed 18 out of 40. The arithmetic is correct, but the management decision is not obvious. We follow this case through data cleaning, statistics, chart design, fairness and decision-making.
This job-specific article belongs to eduKate’s Gold Standard ecosystem, where excellence is defined through reliable management systems and consistent execution. It complements specialised articles on education statistics data quality and career guidance; our focus here is the analyst’s actual professional judgement.
Start with the numbers, not a verdict
| Support route | Completed | Eligible attempts | Observed rate |
|---|---|---|---|
| Route A | 12 | 20 | 60% |
| Route B | 18 | 40 | 45% |
| Both routes | 30 | 60 | 50% |
The percentages come from 12 ÷ 20 = 0.60 and 18 ÷ 40 = 0.45. Route A has a 15 percentage-point higher observed rate; Route B has six more completions in absolute count. These facts do not resolve differences in who used the routes or whether completion is an appropriate outcome for every learner.
Chapter guide
1. The analyst’s job begins with a decision, not a graph
A data analyst is often asked for ‘a dashboard’ when the real problem is that managers cannot decide what to do. A good analyst begins with a question that can be tested and an action that might follow. What decision is unclear? Who will act on the answer? What evidence would change the choice? A chart made before those questions can be visually excellent yet operationally useless.
Imagine a fictional community tutoring programme offering Route A, a self-booked help session, and Route B, a coordinator-supported booking. The manager wants to discontinue one route after seeing different completion percentages. An inexperienced analyst races to rank them. A gold-standard analyst asks who uses each route, what counts as completion, whether both serve comparable learners and what the programme is trying to improve.
Back to chapters · Continue to the next chapter
2. Define the unit of analysis before counting
A spreadsheet might have one row per student, booking, attendance session, reminder or completed activity. Those are different units. If the analyst counts bookings as learners, a person who attends several times may be counted repeatedly. If they count reminders as invitations, one learner with multiple messages can inflate the denominator. A clear unit-of-analysis statement prevents an enormous class of plausible-looking errors.
For the tutoring programme, decide whether the primary outcome is completed help sessions per unique enrolled learner within a defined period or bookings attended out of all bookings. Both measures can be useful but they do not answer the same question. Record the inclusion rules, dates, exclusions and why the measure is relevant to the decision. Definitions belong beside the result, not hidden in an analyst’s personal notebook.
Back to chapters · Continue to the next chapter
3. Follow the data through the way it was created
Operational data are not collected by invisible neutral machines. A tutor marks attendance, a booking system logs an event, a learner cancels and an administrator adjusts a record. Each step can introduce duplication, delay or inconsistent language. The analyst should meet the people who create the fields and learn what the system records under normal and exceptional conditions.
Perhaps the tutoring platform labels every moved appointment as ‘cancelled’ even if a replacement session was completed. Treating raw cancellations as non-participation would distort the result. Source-system knowledge is therefore part of analytical competence. A data dictionary should describe field meanings, formats, allowed values, extraction dates and any quirks relevant to interpretation.
Back to chapters · Continue to the next chapter
4. Build a traceable dataset rather than a one-off file
Reproducibility means a colleague can see which sources, filters, joins and transformations produced a result. Store transformation logic where permitted, document major assumptions and distinguish original source records from derived analytic tables. A file called final_version_9.xlsx is not a reliable provenance system. A concise workflow can be more valuable than elaborate technology that nobody understands.
For a route comparison, keep a dated definition of the cohort, the query or extraction method and the logic for matching learners with session completions. Where privacy and permissions allow, retain a secure audit trail so an unexpected change can be investigated. Reproducibility also means naming data revisions: the numbers in last month’s presentation may legitimately change when late records arrive, but the reason must be discoverable.
Back to chapters · Continue to the next chapter
5. Check missingness before removing records
Missing data are not automatically random. A learner with unstable internet access might leave a digital form incomplete, while someone receiving personal help may be represented in a separate record. If the analyst removes all incomplete entries without examining the pattern, the remaining sample can overrepresent people for whom the service already works.
Start by counting missing values by source, channel, time period and relevant group, when lawful and appropriate. Ask whether a blank field means ‘not collected’, ‘not applicable’, ‘not yet known’ or ‘failed to load’. Do not silently replace all missing values with zero; zero may be a real observation with a different meaning. Report the amount and consequences of missingness so the manager can see the uncertainty it creates.
Back to chapters · Continue to the next chapter
6. Detect duplicated events without deleting genuine activity
Duplicate data often appear when systems retry a submission, import the same file twice or create new identifiers after a change. Yet two rows with similar names can represent two valid activities, and people can have repeated sessions for legitimate reasons. The analyst needs a documented matching rule that recognises the business process rather than a casual ‘remove duplicates’ command.
Use unique identifiers only as authorised and appropriate, compare timestamps and event types and inspect a sample of ambiguous records. Keep a record of rows changed and the reason. If a learner attended two different sessions, their second attendance should not disappear simply because the date of birth and name match. Accuracy depends on preserving true events while eliminating fabricated repetition.
Back to chapters · Continue to the next chapter
7. Read numerator and denominator together
In the fictional data, Route A has 12 completed sessions out of 20 eligible attempts, or 60%. Route B has 18 out of 40, or 45%. Route A has the higher observed completion proportion; Route B has more completed sessions in absolute number. Neither fact alone proves that one route is better for a particular learner. The denominator defines which group is being described.
Ask whether the attempts occurred over the same interval, whether one channel serves more complex needs, and whether the ‘completed’ definition is consistent. A dashboard that shows a green bar for 60% and a red bar for 45% without the counts is incomplete. eduKate’s bird survey case study provides another lesson: a reported trend can change because observation conditions changed, not because the world necessarily did.
Back to chapters · Continue to the next chapter
8. Distinguish association from causal explanation
A route’s higher completion rate does not mean it caused those completions. Learners may select Route A when they have predictable schedules, while Route B may deliberately support people facing greater barriers. If so, the channels are serving different populations. The analyst should describe observed differences, identify competing explanations and avoid claiming that an intervention worked without an appropriate evaluation design.
Where decision consequences justify it, consider carefully designed comparisons, experiments or quasi-experimental methods with suitable oversight. Not every dataset supports causal inference. Sometimes the right answer is that the available records tell us where to investigate, not what programme to cancel. Evidence-based management includes recognising when a question cannot yet be answered responsibly.
Back to chapters · Continue to the next chapter
9. Learn from distributions rather than averages alone
An average waiting time can hide a small number of learners who wait much longer. Consider six illustrative waits of 1, 2, 2, 3, 4 and 24 days: the mean is six days, while the median is 2.5 days. Reporting only the average leaves a misleading picture of what most users experienced; reporting only the median can hide the serious outlier. The right summary depends on the decision at hand.
Look at distributions, extreme cases, seasonality and the likely reasons behind them. An analyst should not automatically remove the 24-day wait as an ‘outlier’ when it may represent the very service failure the manager must solve. Check data validity first, then preserve or explain important extremes. Quantitative accuracy is meaningful only when readers understand what the statistic summarises and what it conceals.
Back to chapters · Continue to the next chapter
10. Choose the right comparison period
School holidays, exam schedules, public campaigns and system outages affect how data should be compared. A week with two public holidays is not equivalent to a full ordinary week for staffing or demand analysis. A year-on-year comparison may reduce one kind of seasonality but introduce changes in policy, user population or collection systems.
The tutoring programme’s April completion rate should not be interpreted in isolation if an examination began during that month. A robust analyst marks relevant events and checks whether the comparison groups and definitions remained stable. Include the chosen date range and any excluded periods directly in the report. This saves decision-makers from building action plans on an accidental calendar difference.
Back to chapters · Continue to the next chapter
11. Make charts faithful to the underlying numbers
A chart is an argument, even when it contains no words. Bars should not use distorted axes to exaggerate small changes; scatter plots should not imply a relationship is linear when it is not; missing observations should not be connected by an invented line. Label units, time periods, denominators and source dates. Choose a chart because it reveals a pattern readers need, not because it looks sophisticated.
Accessible visualisation also matters. Colour alone should not convey meaning, legends should be clear and a table should be available when exact values drive decisions. A useful example is showing both Route A’s 12/20 and Route B’s 18/40 next to their percentages. Viewers can then see why an apparently simple ranking is an oversimplification. Clarity is more important than visual density.
Back to chapters · Continue to the next chapter
12. Protect personal data throughout the analysis
Analysts may work with student, employee, customer or patient information. Permission to access a source system is not necessarily permission to reuse its data for a new purpose or share an export widely. Define the lawful and organisational basis for the work, restrict access, minimise identifiable fields and consult the relevant data-protection team for sensitive use. In Singapore, the PDPC’s anonymisation guidance explains methods and re-identification risks.
Anonymisation is not accomplished merely by removing a name column. Small groups, rare combinations and linked external data may allow someone to infer who a row represents. Suppress or aggregate appropriately, apply access and disclosure controls, and assess the risk of harm in context. A gold-standard analyst protects the people represented by the data, not just the storage drive holding the file.
Back to chapters · Continue to the next chapter
13. Do not turn a model into an unreviewable decision-maker
Predictive models can estimate likely demand, identify possible delays or help allocate outreach. They also reflect historical data quality, changing conditions and design choices. Test the model on relevant unseen data where appropriate, document accuracy and failure modes and monitor changes after deployment. Avoid using a high model score as proof that the recommended action is fair or suitable.
Singapore’s PDPC advisory guidelines for AI recommendation and decision systems address the use of personal data in such systems. An analyst must still understand the organisation’s legal and ethical obligations. Decisions affecting individuals deserve accountable review, meaningful explanations where applicable and an escalation path when a recommendation is disputed.
Back to chapters · Continue to the next chapter
14. Work with data engineers and subject experts
A data analyst is not expected to personally build every database, understand every clinical code or administer every cloud service. Yet effective work requires communication with people who own source systems, data infrastructure and the real-world process. Agree on field definitions, extraction schedules, incident reporting and quality checks. A report built on an unstable pipeline can mislead even when its statistical calculations are impeccable.
If an attendance feed stops receiving updates for two days, the analyst should mark the dashboard as potentially incomplete, inform its users and work with the relevant engineer to investigate. Publishing another refresh timestamp does not establish that fresh records were actually ingested. The gold standard is explicit responsibility from source to interpretation.
Back to chapters · Continue to the next chapter
15. Use public data responsibly, with its metadata attached
Public datasets can add useful context, but they often differ in coverage, vintage, classification and methodology. A data analyst should read the dataset description, update history, geographic scope and usage conditions before combining sources. Revisions are normal in official statistics, especially where later information improves an initial estimate.
Singapore’s data.gov.sg metadata guide describes a technical path to dataset metadata. The Singapore Department of Statistics statistical best-practice handbook provides wider quality principles. The lesson is not that official data are infallible, but that responsible analysis carries the source’s context and limitations into the output.
Back to chapters · Continue to the next chapter
16. Write conclusions at the strength the evidence warrants
Readers often want a crisp answer: ‘Which route wins?’ The analyst can be concise without pretending to know more than the data support. A credible short conclusion might say, ‘The observed completion proportion was higher for Route A (12/20) than Route B (18/40), but the groups may differ in important ways. Check eligibility definitions and participant characteristics before changing access.’
Support conclusions with enough context that an executive can act sensibly. Distinguish a descriptive result, an uncertain hypothesis and a recommendation requiring further validation. Avoid burying the principal limitation on slide 29. When a decision is urgent, propose a reversible step and an evidence-collection plan rather than creating artificial certainty from a small sample.
Back to chapters · Continue to the next chapter
17. Build useful dashboards around thresholds and action
An effective operational dashboard says what the numbers mean, when they were last checked and what action should follow. It distinguishes normal variation from conditions requiring attention. A metric without an owner or response pathway becomes decoration. Analysts should interview the people expected to use the dashboard and test whether they can interpret it correctly.
The tutoring programme might monitor learners waiting more than a defined number of days for help, but a threshold should be approved by the service owner and justified by the programme’s context. A sudden missing-data spike should show a data-quality warning, not a reassuring zero. Visual alerts are valuable when they send the right person toward an important, verifiable exception.
Back to chapters · Continue to the next chapter
18. Make analytical work reproducible across staff changes
Organisations sometimes depend on one analyst who remembers an undocumented correction in a spreadsheet. When the person leaves, reports cannot be reproduced. Store versioned query or transformation logic, definitions, approved access routes and quality checks in maintainable locations. Separate exploratory experiments from the approved production reporting process.
An analyst should be able to rerun a monthly report, explain why a late-arriving record changed the prior result and identify which stakeholder was told. If the source system changes a column meaning, the team should update both code and reporting definitions. Documentation is not an accessory to the clever part of the job; it preserves the value of all past effort.
Back to chapters · Continue to the next chapter
19. Build a career in analysis by practising disciplined questions
Skills include statistics, spreadsheets, SQL, data cleaning, visualisation, domain knowledge, writing and ethical judgement. Coding is useful, but programming syntax does not compensate for a mistaken definition of ‘completed’. Begin with small datasets where the source and correct arithmetic are known. Then progress to multi-table joins, missing-data analysis, stakeholder presentations and controlled automation.
Singapore’s IMDA Skills Framework for ICT includes data analytics roles, technical capabilities and career pathways. A student can build a portfolio showing the question, input, cleaning decisions, workings, uncertainties and sensible conclusion. A traceable, modest project often demonstrates professional readiness better than a dazzling dashboard whose data source cannot be explained.
Back to chapters · Continue to the next chapter
20. Measure the analyst by decisions improved, not charts produced
An analyst can complete one hundred reports without changing an organisation’s understanding. Strong work prevents errors, identifies meaningful opportunities, makes uncertainty visible and helps managers avoid unsuitable actions. Feedback should examine whether definitions remained correct, users understood the results, privacy was protected and decisions improved as evidence accumulated.
Return to Routes A and B. The analyst’s job is not to choose a favourite bar colour or declare that one group won. It is to establish that 12 out of 20 and 18 out of 40 are comparable observations, explain why they may not represent causal effects, and help the programme test an improvement without harming those who rely on it. That is analytical excellence as accountable work.
A second worked check: a misleading average
Suppose a service recorded six waiting times: 1, 2, 2, 3, 4 and 24 days. Their total is 36, so the mean wait is 36 ÷ 6 = 6 days. With six ordered observations, the median is the average of the third and fourth values: (2 + 3) ÷ 2 = 2.5 days. Both results are correct. They answer different questions. Most people in the example waited far less than the mean, while one person waited much longer.
An analyst who shows only the mean may overstate the typical wait; one who shows only the median may conceal a very serious service failure. The more useful report includes counts, central tendency, the long tail, relevant segments and context explaining why the 24-day wait occurred. Never delete a valid difficult case merely because it makes the average look unattractive.
A repeatable analyst checklist
- Write the decision question and who is accountable for acting on it.
- Confirm the unit of analysis, dates, population, numerator and denominator.
- Document source fields, data ownership, collection changes and transformation rules.
- Review missing, duplicated and inconsistent records with the people who create them.
- Choose calculations and charts that reveal rather than obscure variation.
- Check privacy, legitimate use, anonymisation limits and access rights.
- Explain alternative interpretations, uncertainty and what evidence would change the conclusion.
- Ensure a colleague can reproduce the result and explain later revisions.
Frequently asked questions about data analysts
What is the difference between a data analyst and data scientist?
Titles overlap. Analysts often focus on defining metrics, examining operational data and explaining decisions; data scientists may use more advanced modelling and experimentation. The actual duties depend on the employer.
Does a data analyst need to code?
SQL and scripting can make workflows more capable and repeatable, but a sound understanding of data definitions, statistics, communication and privacy is equally important. Some roles rely substantially on spreadsheet and visualisation tools.
Why isn’t a higher percentage always the better outcome?
Percentages depend on what enters the denominator and who belongs to each group. Selection bias, different difficulty levels, small samples and changing definitions can produce misleading rankings.
Can an analyst publish anonymised records?
Not automatically. Anonymisation must be assessed against the realistic risk of re-identification and applicable rules. Removing names is not sufficient on its own, especially for small or linkable datasets.
What does reproducible analysis mean?
A suitably authorised colleague can identify the source, reproduce filters and transformations, recover the reported numbers and understand exceptions and revisions.
What should a beginner include in an analytics portfolio?
One real question, clearly permitted or synthetic data, a data dictionary, cleaning steps, checked calculations, a readable visual, limitations and a decision-ready conclusion.
Should an analyst always use AI or machine learning?
No. Choose the simplest suitable method. A complex model can obscure a decision that would be better served by a checked proportion, distribution or direct process review.
How do employers measure an analyst’s quality?
Check data correctness, reproducibility, responsible use, clarity about uncertainty and whether insights improve the decisions for which the analyst was engaged.
Official sources and complementary eduKate articles
- IMDA: Skills Framework for ICT — data analytics careers and skills.
- PDPC: Basic anonymisation — privacy safeguards and re-identification risk.
- Singapore Department of Statistics: Statistical best practice — quality in statistical production.
- data.gov.sg: Dataset metadata — understanding dataset source and definitions.
- eduKate: Education equity statistics and safe reporting — policy-sector context.
- eduKate: Case study on changing observation sites — a misleading trend’s underlying cause.
The gold standard of data analysts is a conclusion that survives someone asking, ‘What exactly did you count, where did it come from, and what might make your explanation wrong?’