Why Science? | Institute of Data Science

Data science is useful when a defined decision or explanation guides what is collected and evaluated. Starting with available columns can create a polished answer to the wrong question.

Start with the observable system. A problem statement names user, decision, target, time horizon, cost of errors and acceptable evidence. The important question is not simply whether something can be measured, modelled or built, but which claim that work supports under the stated conditions. A useful explanation names the system boundary, the independent and dependent quantities, the comparison, and the evidence that could change the conclusion.

In a real investigation, connect this idea to the surrounding workflow. Record the starting material or participant context, instrument or model version, settings, controls, raw observation, transformation, uncertainty and decision. That record lets a classmate or future researcher see whether a surprising result came from the phenomenon, the method, the sample or the interpretation. This is how data-question framing becomes a transferable scientific capability rather than a remembered definition.

A practical way to learn it is to turn “predict health” into a bounded, measurable question with a named user. Before accepting the result, ask what was held constant, what was deliberately changed, what was actually observed, and whether an alternative explanation fits the same evidence. Add one negative or blank control and one independent cross-check wherever the question permits.

Interpretation needs a stop line. A technically solvable prediction may still have no legitimate or beneficial use. Write that boundary beside the result rather than hiding it in a footnote. Then state what additional measurement, population, comparison, duration or authority would be needed before making the next larger claim. Careful limits make an article more useful because readers can act on what is known without mistaking curiosity for certainty.

Contents · Next chapter

Every row exists because a person, sensor or system recorded something under rules. Those rules shape missingness, labels and apparent patterns.

Follow the mechanism before the headline. A process map traces who enters the dataset, when measurements occur and which events remain invisible. The important question is not simply whether something can be measured, modelled or built, but which claim that work supports under the stated conditions. A useful explanation names the system boundary, the independent and dependent quantities, the comparison, and the evidence that could change the conclusion.

In a real investigation, connect this idea to the surrounding workflow. Record the starting material or participant context, instrument or model version, settings, controls, raw observation, transformation, uncertainty and decision. That record lets a classmate or future researcher see whether a surprising result came from the phenomenon, the method, the sample or the interpretation. This is how data-generation literacy becomes a transferable scientific capability rather than a remembered definition.

A practical way to learn it is to draw how a hospital visit becomes a database row and identify who is absent. Before accepting the result, ask what was held constant, what was deliberately changed, what was actually observed, and whether an alternative explanation fits the same evidence. Add one negative or blank control and one independent cross-check wherever the question permits.

Interpretation needs a stop line. Observational data is not a neutral mirror of reality. Write that boundary beside the result rather than hiding it in a footnote. Then state what additional measurement, population, comparison, duration or authority would be needed before making the next larger claim. Careful limits make an article more useful because readers can act on what is known without mistaking curiosity for certainty.

Contents · Previous chapter · Next chapter

Source, collection time, units, transformations and ownership determine whether a value can be interpreted. Copying data without lineage breaks that chain.

Make the comparison explicit. A provenance log records immutable raw sources, versioned transformations and responsible owners. The important question is not simply whether something can be measured, modelled or built, but which claim that work supports under the stated conditions. A useful explanation names the system boundary, the independent and dependent quantities, the comparison, and the evidence that could change the conclusion.

In a real investigation, connect this idea to the surrounding workflow. Record the starting material or participant context, instrument or model version, settings, controls, raw observation, transformation, uncertainty and decision. That record lets a classmate or future researcher see whether a surprising result came from the phenomenon, the method, the sample or the interpretation. This is how data-provenance stewardship becomes a transferable scientific capability rather than a remembered definition.

A practical way to learn it is to trace one field through acquisition, cleaning, aggregation and model input. Before accepting the result, ask what was held constant, what was deliberately changed, what was actually observed, and whether an alternative explanation fits the same evidence. Add one negative or blank control and one independent cross-check wherever the question permits.

Interpretation needs a stop line. A clean filename does not prove origin, permission or integrity. Write that boundary beside the result rather than hiding it in a footnote. Then state what additional measurement, population, comparison, duration or authority would be needed before making the next larger claim. Careful limits make an article more useful because readers can act on what is known without mistaking curiosity for certainty.

Contents · Previous chapter · Next chapter

Duplicates, impossible values and missing entries require rules tied to domain meaning. Automatic deletion can remove real edge cases.

Keep the evidence chain visible. Cleaning code produces counts and reasons for every change, while sensitivity tests compare plausible treatments. The important question is not simply whether something can be measured, modelled or built, but which claim that work supports under the stated conditions. A useful explanation names the system boundary, the independent and dependent quantities, the comparison, and the evidence that could change the conclusion.

In a real investigation, connect this idea to the surrounding workflow. Record the starting material or participant context, instrument or model version, settings, controls, raw observation, transformation, uncertainty and decision. That record lets a classmate or future researcher see whether a surprising result came from the phenomenon, the method, the sample or the interpretation. This is how data-cleaning judgement becomes a transferable scientific capability rather than a remembered definition.

A practical way to learn it is to write three alternative rules for a missing value and predict their impact. Before accepting the result, ask what was held constant, what was deliberately changed, what was actually observed, and whether an alternative explanation fits the same evidence. Add one negative or blank control and one independent cross-check wherever the question permits.

Interpretation needs a stop line. Tidy data can be systematically distorted by an unjustified rule. Write that boundary beside the result rather than hiding it in a footnote. Then state what additional measurement, population, comparison, duration or authority would be needed before making the next larger claim. Careful limits make an article more useful because readers can act on what is known without mistaking curiosity for certainty.

Contents · Previous chapter · Next chapter

A sample can be large and still exclude groups, places or times. Selection affects which population the result can describe.

Test the boundary as carefully as the main effect. Sampling frames, response patterns and weighting are documented before generalising beyond observed cases. The important question is not simply whether something can be measured, modelled or built, but which claim that work supports under the stated conditions. A useful explanation names the system boundary, the independent and dependent quantities, the comparison, and the evidence that could change the conclusion.

In a real investigation, connect this idea to the surrounding workflow. Record the starting material or participant context, instrument or model version, settings, controls, raw observation, transformation, uncertainty and decision. That record lets a classmate or future researcher see whether a surprising result came from the phenomenon, the method, the sample or the interpretation. This is how sampling-bias literacy becomes a transferable scientific capability rather than a remembered definition.

A practical way to learn it is to compare a convenience sample with a stratified sample for one city question. Before accepting the result, ask what was held constant, what was deliberately changed, what was actually observed, and whether an alternative explanation fits the same evidence. Add one negative or blank control and one independent cross-check wherever the question permits.

Interpretation needs a stop line. More rows reduce random error without removing selection bias. Write that boundary beside the result rather than hiding it in a footnote. Then state what additional measurement, population, comparison, duration or authority would be needed before making the next larger claim. Careful limits make an article more useful because readers can act on what is known without mistaking curiosity for certainty.

Contents · Previous chapter · Next chapter

Plots and summaries show distributions, relationships and anomalies. Repeated searching also increases the chance of finding patterns by accident.

Start with the observable system. Exploratory work is labelled, with later confirmation on held-out or newly collected evidence. The important question is not simply whether something can be measured, modelled or built, but which claim that work supports under the stated conditions. A useful explanation names the system boundary, the independent and dependent quantities, the comparison, and the evidence that could change the conclusion.

In a real investigation, connect this idea to the surrounding workflow. Record the starting material or participant context, instrument or model version, settings, controls, raw observation, transformation, uncertainty and decision. That record lets a classmate or future researcher see whether a surprising result came from the phenomenon, the method, the sample or the interpretation. This is how exploratory-data reasoning becomes a transferable scientific capability rather than a remembered definition.

A practical way to learn it is to explore five variables, then write a separate confirmation plan for one pattern. Before accepting the result, ask what was held constant, what was deliberately changed, what was actually observed, and whether an alternative explanation fits the same evidence. Add one negative or blank control and one independent cross-check wherever the question permits.

Interpretation needs a stop line. An attractive correlation found after many tries is a hypothesis, not a settled result. Write that boundary beside the result rather than hiding it in a footnote. Then state what additional measurement, population, comparison, duration or authority would be needed before making the next larger claim. Careful limits make an article more useful because readers can act on what is known without mistaking curiosity for certainty.

Contents · Previous chapter · Next chapter

Estimates vary because samples and measurements vary. Confidence intervals and models express uncertainty under assumptions.

Follow the mechanism before the headline. Analysts state estimand, model, assumptions and diagnostics instead of reporting a p-value alone. The important question is not simply whether something can be measured, modelled or built, but which claim that work supports under the stated conditions. A useful explanation names the system boundary, the independent and dependent quantities, the comparison, and the evidence that could change the conclusion.

In a real investigation, connect this idea to the surrounding workflow. Record the starting material or participant context, instrument or model version, settings, controls, raw observation, transformation, uncertainty and decision. That record lets a classmate or future researcher see whether a surprising result came from the phenomenon, the method, the sample or the interpretation. This is how statistical-inference literacy becomes a transferable scientific capability rather than a remembered definition.

A practical way to learn it is to compare two estimates with overlapping uncertainty and explain what cannot be concluded. Before accepting the result, ask what was held constant, what was deliberately changed, what was actually observed, and whether an alternative explanation fits the same evidence. Add one negative or blank control and one independent cross-check wherever the question permits.

Interpretation needs a stop line. Statistical significance does not establish practical importance or correct design. Write that boundary beside the result rather than hiding it in a footnote. Then state what additional measurement, population, comparison, duration or authority would be needed before making the next larger claim. Careful limits make an article more useful because readers can act on what is known without mistaking curiosity for certainty.

Contents · Previous chapter · Next chapter

Machine-learning models estimate outcomes for new cases. Improvement matters relative to a simple, relevant baseline and the cost of different errors.

Make the comparison explicit. Metrics are selected around class balance, calibration and decision thresholds, with results shown across groups and conditions. The important question is not simply whether something can be measured, modelled or built, but which claim that work supports under the stated conditions. A useful explanation names the system boundary, the independent and dependent quantities, the comparison, and the evidence that could change the conclusion.

In a real investigation, connect this idea to the surrounding workflow. Record the starting material or participant context, instrument or model version, settings, controls, raw observation, transformation, uncertainty and decision. That record lets a classmate or future researcher see whether a surprising result came from the phenomenon, the method, the sample or the interpretation. This is how predictive-model evaluation becomes a transferable scientific capability rather than a remembered definition.

A practical way to learn it is to compare accuracy with a majority baseline and add precision, recall and calibration. Before accepting the result, ask what was held constant, what was deliberately changed, what was actually observed, and whether an alternative explanation fits the same evidence. Add one negative or blank control and one independent cross-check wherever the question permits.

Interpretation needs a stop line. A single high metric can hide useless or harmful operating behaviour. Write that boundary beside the result rather than hiding it in a footnote. Then state what additional measurement, population, comparison, duration or authority would be needed before making the next larger claim. Careful limits make an article more useful because readers can act on what is known without mistaking curiosity for certainty.

Contents · Previous chapter · Next chapter

Random splits can leak repeated people, locations or later information. Time, group and site splits test the kind of generalisation deployment requires.

Keep the evidence chain visible. A validation plan separates development, tuning and final evaluation with no hidden reuse. The important question is not simply whether something can be measured, modelled or built, but which claim that work supports under the stated conditions. A useful explanation names the system boundary, the independent and dependent quantities, the comparison, and the evidence that could change the conclusion.

In a real investigation, connect this idea to the surrounding workflow. Record the starting material or participant context, instrument or model version, settings, controls, raw observation, transformation, uncertainty and decision. That record lets a classmate or future researcher see whether a surprising result came from the phenomenon, the method, the sample or the interpretation. This is how model-validation judgement becomes a transferable scientific capability rather than a remembered definition.

A practical way to learn it is to design a time-aware split for an urban forecasting model. Before accepting the result, ask what was held constant, what was deliberately changed, what was actually observed, and whether an alternative explanation fits the same evidence. Add one negative or blank control and one independent cross-check wherever the question permits.

Interpretation needs a stop line. Performance on a familiar test set does not guarantee transport to a new setting. Write that boundary beside the result rather than hiding it in a footnote. Then state what additional measurement, population, comparison, duration or authority would be needed before making the next larger claim. Careful limits make an article more useful because readers can act on what is known without mistaking curiosity for certainty.

Contents · Previous chapter · Next chapter

Prediction asks what is likely; causality asks what would change under an intervention. Confounding and selection can make association misleading.

Test the boundary as carefully as the main effect. A causal diagram and identification strategy state assumptions before estimation, supported by design or justified natural variation. The important question is not simply whether something can be measured, modelled or built, but which claim that work supports under the stated conditions. A useful explanation names the system boundary, the independent and dependent quantities, the comparison, and the evidence that could change the conclusion.

In a real investigation, connect this idea to the surrounding workflow. Record the starting material or participant context, instrument or model version, settings, controls, raw observation, transformation, uncertainty and decision. That record lets a classmate or future researcher see whether a surprising result came from the phenomenon, the method, the sample or the interpretation. This is how causal-data literacy becomes a transferable scientific capability rather than a remembered definition.

A practical way to learn it is to draw a three-variable confounding example and compare predictive with causal questions. Before accepting the result, ask what was held constant, what was deliberately changed, what was actually observed, and whether an alternative explanation fits the same evidence. Add one negative or blank control and one independent cross-check wherever the question permits.

Interpretation needs a stop line. A model feature importance score is not an intervention effect. Write that boundary beside the result rather than hiding it in a footnote. Then state what additional measurement, population, comparison, duration or authority would be needed before making the next larger claim. Careful limits make an article more useful because readers can act on what is known without mistaking curiosity for certainty.

Contents · Previous chapter · Next chapter

Measurements close in time can share trends, seasons and shocks. Treating them as independent inflates confidence and leaks future information.

Start with the observable system. Temporal baselines, rolling evaluation and residual checks test whether a model learns stable signal. The important question is not simply whether something can be measured, modelled or built, but which claim that work supports under the stated conditions. A useful explanation names the system boundary, the independent and dependent quantities, the comparison, and the evidence that could change the conclusion.

In a real investigation, connect this idea to the surrounding workflow. Record the starting material or participant context, instrument or model version, settings, controls, raw observation, transformation, uncertainty and decision. That record lets a classmate or future researcher see whether a surprising result came from the phenomenon, the method, the sample or the interpretation. This is how time-series reasoning becomes a transferable scientific capability rather than a remembered definition.

A practical way to learn it is to build a rolling forecast evaluation across four periods. Before accepting the result, ask what was held constant, what was deliberately changed, what was actually observed, and whether an alternative explanation fits the same evidence. Add one negative or blank control and one independent cross-check wherever the question permits.

Interpretation needs a stop line. A smooth historical fit may fail after a regime change. Write that boundary beside the result rather than hiding it in a footnote. Then state what additional measurement, population, comparison, duration or authority would be needed before making the next larger claim. Careful limits make an article more useful because readers can act on what is known without mistaking curiosity for certainty.

Contents · Previous chapter · Next chapter

Locations interact through distance, boundaries and movement. Aggregation can reverse patterns and administrative units may not match the process.

Follow the mechanism before the headline. Maps state projection, resolution and denominator, while spatial validation holds out coherent areas. The important question is not simply whether something can be measured, modelled or built, but which claim that work supports under the stated conditions. A useful explanation names the system boundary, the independent and dependent quantities, the comparison, and the evidence that could change the conclusion.

In a real investigation, connect this idea to the surrounding workflow. Record the starting material or participant context, instrument or model version, settings, controls, raw observation, transformation, uncertainty and decision. That record lets a classmate or future researcher see whether a surprising result came from the phenomenon, the method, the sample or the interpretation. This is how spatial-data literacy becomes a transferable scientific capability rather than a remembered definition.

A practical way to learn it is to aggregate the same fictional points into two boundary systems and compare rates. Before accepting the result, ask what was held constant, what was deliberately changed, what was actually observed, and whether an alternative explanation fits the same evidence. Add one negative or blank control and one independent cross-check wherever the question permits.

Interpretation needs a stop line. A map hotspot can arise from population density or reporting access rather than the target process. Write that boundary beside the result rather than hiding it in a footnote. Then state what additional measurement, population, comparison, duration or authority would be needed before making the next larger claim. Careful limits make an article more useful because readers can act on what is known without mistaking curiosity for certainty.

Contents · Previous chapter · Next chapter

Text, images, sensors and tables capture different views. Joining them requires consistent entities, time and consent.

Make the comparison explicit. Fusion models are tested against single-modality baselines and missing-modality cases. The important question is not simply whether something can be measured, modelled or built, but which claim that work supports under the stated conditions. A useful explanation names the system boundary, the independent and dependent quantities, the comparison, and the evidence that could change the conclusion.

In a real investigation, connect this idea to the surrounding workflow. Record the starting material or participant context, instrument or model version, settings, controls, raw observation, transformation, uncertainty and decision. That record lets a classmate or future researcher see whether a surprising result came from the phenomenon, the method, the sample or the interpretation. This is how multimodal-data reasoning becomes a transferable scientific capability rather than a remembered definition.

A practical way to learn it is to align a clinic note, image and sensor stream around one fictional event. Before accepting the result, ask what was held constant, what was deliberately changed, what was actually observed, and whether an alternative explanation fits the same evidence. Add one negative or blank control and one independent cross-check wherever the question permits.

Interpretation needs a stop line. More modalities can add noise, leakage and privacy risk as well as information. Write that boundary beside the result rather than hiding it in a footnote. Then state what additional measurement, population, comparison, duration or authority would be needed before making the next larger claim. Careful limits make an article more useful because readers can act on what is known without mistaking curiosity for certainty.

Contents · Previous chapter · Next chapter

Large pretrained models can support many tasks, but their outputs reflect training coverage, objectives and prompting. Evaluation must match the actual domain.

Keep the evidence chain visible. Teams use task-specific tests, retrieval checks, refusal criteria and human review for high-impact uses. The important question is not simply whether something can be measured, modelled or built, but which claim that work supports under the stated conditions. A useful explanation names the system boundary, the independent and dependent quantities, the comparison, and the evidence that could change the conclusion.

In a real investigation, connect this idea to the surrounding workflow. Record the starting material or participant context, instrument or model version, settings, controls, raw observation, transformation, uncertainty and decision. That record lets a classmate or future researcher see whether a surprising result came from the phenomenon, the method, the sample or the interpretation. This is how foundation-model judgement becomes a transferable scientific capability rather than a remembered definition.

A practical way to learn it is to design an evaluation set with common, rare and adversarial cases. Before accepting the result, ask what was held constant, what was deliberately changed, what was actually observed, and whether an alternative explanation fits the same evidence. Add one negative or blank control and one independent cross-check wherever the question permits.

Interpretation needs a stop line. Fluent output is not proof of source, truth or suitability. Write that boundary beside the result rather than hiding it in a footnote. Then state what additional measurement, population, comparison, duration or authority would be needed before making the next larger claim. Careful limits make an article more useful because readers can act on what is known without mistaking curiosity for certainty.

Contents · Previous chapter · Next chapter

Data minimisation, access control, de-identification and federated methods can reduce exposure. Each protects against specific threats, not every threat.

Test the boundary as carefully as the main effect. A privacy threat model names attacker, data, linkage and allowed outputs before choosing safeguards. The important question is not simply whether something can be measured, modelled or built, but which claim that work supports under the stated conditions. A useful explanation names the system boundary, the independent and dependent quantities, the comparison, and the evidence that could change the conclusion.

In a real investigation, connect this idea to the surrounding workflow. Record the starting material or participant context, instrument or model version, settings, controls, raw observation, transformation, uncertainty and decision. That record lets a classmate or future researcher see whether a surprising result came from the phenomenon, the method, the sample or the interpretation. This is how privacy-preserving data science becomes a transferable scientific capability rather than a remembered definition.

A practical way to learn it is to compare centralised and federated analysis for one multi-site study. Before accepting the result, ask what was held constant, what was deliberately changed, what was actually observed, and whether an alternative explanation fits the same evidence. Add one negative or blank control and one independent cross-check wherever the question permits.

Interpretation needs a stop line. Keeping data local does not automatically prevent inference, leakage or misuse. Write that boundary beside the result rather than hiding it in a footnote. Then state what additional measurement, population, comparison, duration or authority would be needed before making the next larger claim. Careful limits make an article more useful because readers can act on what is known without mistaking curiosity for certainty.

Contents · Previous chapter · Next chapter

A model joins software, people, thresholds and workflows. Drift, outages and feedback can change performance after launch.

Start with the observable system. Owners monitor inputs, outcomes, subgroup metrics, incidents and override patterns with rollback and recourse. The important question is not simply whether something can be measured, modelled or built, but which claim that work supports under the stated conditions. A useful explanation names the system boundary, the independent and dependent quantities, the comparison, and the evidence that could change the conclusion.

In a real investigation, connect this idea to the surrounding workflow. Record the starting material or participant context, instrument or model version, settings, controls, raw observation, transformation, uncertainty and decision. That record lets a classmate or future researcher see whether a surprising result came from the phenomenon, the method, the sample or the interpretation. This is how data-system operations becomes a transferable scientific capability rather than a remembered definition.

A practical way to learn it is to draw a deployment loop from input to action, outcome, audit and update. Before accepting the result, ask what was held constant, what was deliberately changed, what was actually observed, and whether an alternative explanation fits the same evidence. Add one negative or blank control and one independent cross-check wherever the question permits.

Interpretation needs a stop line. A validated model can become unsafe when context or incentives change. Write that boundary beside the result rather than hiding it in a footnote. Then state what additional measurement, population, comparison, duration or authority would be needed before making the next larger claim. Careful limits make an article more useful because readers can act on what is known without mistaking curiosity for certainty.

Contents · Previous chapter · Next chapter

A credible portfolio records question, population, source, provenance, cleaning, code, split, baseline, model, uncertainty, privacy, owner, monitoring and claim.

Follow the mechanism before the headline. The receipt lets another reader reproduce the analysis and see where judgement entered. The important question is not simply whether something can be measured, modelled or built, but which claim that work supports under the stated conditions. A useful explanation names the system boundary, the independent and dependent quantities, the comparison, and the evidence that could change the conclusion.

In a real investigation, connect this idea to the surrounding workflow. Record the starting material or participant context, instrument or model version, settings, controls, raw observation, transformation, uncertainty and decision. That record lets a classmate or future researcher see whether a surprising result came from the phenomenon, the method, the sample or the interpretation. This is how data-science career preparation becomes a transferable scientific capability rather than a remembered definition.

A practical way to learn it is to build one page for a fictional project and ask a peer to locate every transformation. Before accepting the result, ask what was held constant, what was deliberately changed, what was actually observed, and whether an alternative explanation fits the same evidence. Add one negative or blank control and one independent cross-check wherever the question permits.

Interpretation needs a stop line. Real data use requires permission, governance and domain responsibility as well as technical skill. Write that boundary beside the result rather than hiding it in a footnote. Then state what additional measurement, population, comparison, duration or authority would be needed before making the next larger claim. Careful limits make an article more useful because readers can act on what is known without mistaking curiosity for certainty.

Contents · Previous chapter