6. Data Analysis, Statistics, and Visualisation

AI can assist with data cleaning plans, spreadsheet formulas, code generation, chart selection, interpretation of outputs, and documentation. It does not remove the need for sound research design, correct statistical assumptions, data protection, and reproducible analysis.

6.1 Common tools

Excel and Google Sheets include increasingly capable AI-assisted functions and remain useful for tabular analysis. Power BI and Tableau support business intelligence and dashboards. Python libraries such as pandas, NumPy, SciPy, statsmodels, and scikit-learn support reproducible analysis; R provides a strong statistical and visualisation ecosystem. SPSS, Stata, SAS, and other specialist packages remain relevant where required by a discipline or institution. ChatGPT and other assistants may help explain code, draft syntax, or interpret output when the user supplies adequate context.

6.2 Safe analytical workflow

1. Preserve an untouched copy of the original data and maintain a data dictionary.
2. Remove or mask direct identifiers before uploading data to any external tool, unless authorised safeguards and agreements permit it.
3. Ask AI to propose a cleaning plan; inspect each proposed transformation before applying it.
4. Use reproducible scripts or documented formulas rather than undocumented manual edits.
5. Check missing values, duplicates, outliers, coding errors, and variable definitions.
6. Confirm assumptions and select analyses based on the research question, design, measurement level, and sampling strategy.
7. Reproduce key calculations independently and compare results with the chosen statistical package.
8. Report effect sizes, uncertainty, limitations, and appropriate interpretation—not only p-values.

6.3 Avoiding overinterpretation

A cross-sectional survey typically cannot establish temporal order or causal effects by itself. Self-report measures can introduce common-method and social-desirability biases. Small or convenience samples limit generalisability. AI-generated explanations may overstate a correlation as a cause or describe an association as an impact. Require the tool to state the design limitations and use wording consistent with the evidence.

6.4 Visualisation

Ask for a chart that answers a defined question. Bar charts compare categories; line charts show change over ordered time; scatterplots explore relationships between numeric variables; boxplots display distributions; and maps require appropriate geographic data. Check labels, units, scale, sample size, missing data, uncertainty, and whether the chart could mislead. Never use a visually polished graph as a substitute for checking the underlying data.

6.5 Reproducibility

Record software versions, prompts that generated code, data transformations, package versions, random seeds where relevant, and analytical decisions. Review code line by line and test it on a small sample. Do not run unfamiliar code against sensitive systems or production data without review.