Data Thinking

Computational thinking traditionally starts with the algorithm. Data thinking insists you start one step earlier — with the evidence the algorithm will consume. A perfectly correct program fed skewed, stale, or unrepresentative data produces confident nonsense. In an era where most decisions are data-mediated and most AI systems are only as good as their training sets, reasoning carefully about data is no longer a specialist skill. It is a core part of thinking computationally.

Data as evidence, not as truth

The single most useful mental move is to treat every dataset as evidence produced by a process, not as a neutral window onto reality. A table of customer complaints is not a picture of customer sentiment — it is a record of who was angry enough, and able, to file a complaint through a specific channel. The gap between what the data measures and what you wish it measured is where most analytical errors live.

Three questions expose that gap quickly:

• What generated this? Every value was created by some instrument, form, logging call, or human judgment. That generating process has quirks: a sensor that saturates, a dropdown that forces a choice, a default that most users never change.
• What is missing, and why? Absent rows are rarely random. Patients who never returned, transactions that failed before logging, users who churned before the survey — the silence is itself a signal.
• Who is not in here? Coverage gaps become discrimination the moment a model trained on the data makes decisions about the people it never saw.

Bias and representativeness

Bias in data is not a moral accusation; it is a structural fact about how the sample relates to the population you care about. The classic failure is selection bias — the data that was easiest to collect systematically differs from the data you needed. A hiring model trained on past “successful” employees learns the demographics of past hiring decisions, not the traits of good work.

Watch for these recurring patterns:

• Survivorship bias — you only see the cases that made it far enough to be recorded. Reinforcing aircraft where returning planes were hit is the famous inversion.
• Measurement bias — the instrument systematically over- or under-reports (a scale that reads 2 kg heavy, a survey question that leads the respondent).
• Temporal drift — the world moved but the data did not. A fraud model trained on 2023 patterns degrades as fraudsters adapt.
• Aggregation traps — a trend that holds in every subgroup can reverse when the groups are combined (Simpson’s paradox). Always check whether a headline number hides opposing sub-stories.

The data lifecycle

Thinking computationally about data means thinking about its whole lifecycle, because errors introduced at any stage propagate downstream and are expensive to detect later.

    flowchart LR
    A[Generation] --> B[Collection]
    B --> C[Cleaning and Transformation]
    C --> D[Analysis and Modeling]
    D --> E[Interpretation]
    E --> F[Action and Feedback]
    F --> A
  

Each stage adds assumptions. Collection decides what is representable at all. Cleaning is where subjective calls — how to treat outliers, how to impute missing values, how to bucket categories — quietly become “the data.” The loop matters most: actions taken on the data change the world, which changes future data. A predictive-policing model that sends patrols to a neighborhood then records more arrests there, confirming its own prior. Recognizing these feedback-through-action loops is where data thinking meets systems thinking.

A practical data-audit checklist

Before trusting any dataset for a decision, walk this list. It takes minutes and prevents the most common and expensive mistakes.

Why this belongs in computational thinking

Abstraction — a core CT pillar — is exactly the act of deciding what to keep and what to discard. Data thinking is abstraction applied to evidence: choosing which measurements represent the phenomenon faithfully, and being honest about what the abstraction throws away. Get this wrong and every downstream algorithm, however elegant, computes a precise answer to the wrong question.

References

• AlphaFold Protein Structure Database in 2024: over 214 million sequences
• The Vera C. Rubin Observatory Data Preview 1
• Revised CSTA Standards: Data & Analysis concept