8.2 A reproducible EDA workflow

8.2.1 Stage 1: Orient to the data

Record the source and version of each file, unit of analysis, dimensions, identifiers, time coverage, geographic coverage, variable types, and relationships between tables. Confirm that the number of records agrees with source documentation when a reference total exists.

dim(data)
names(data)
str(data)
summary(data)

These commands are only an opening inspection. They do not replace business definitions or targeted quality checks.

8.2.2 Stage 2: Assess data quality

Check for:

  • duplicate rows and identifiers;
  • missing counts and patterns;
  • unexpected categories;
  • impossible or conflicting values;
  • inconsistent dates, units, spelling, or formats;
  • numbers stored as text;
  • extreme and influential observations;
  • incomplete time periods;
  • totals that do not reconcile;
  • truncated exports;
  • selection and coverage limitations; and
  • discrepancies between data and dictionary.

The Quartz guide provides a useful taxonomy of data problems that may require action by the source, analyst, subject expert, or programmer (Quartz, n.d.). Classifying ownership prevents analysts from inventing corrections to questions that require domain authority.

8.2.3 Stage 3: Examine variables individually

Choose summaries that match the variable.

For categorical variables, examine counts, proportions, rare levels, ordering, and missingness. For numeric variables, examine centre, spread, range, quantiles, zeros, shape, and extremes. For dates and times, examine coverage, gaps, seasonality, incomplete periods, and operational cycles.

Do not report every available statistic. Report the features that affect interpretation or later analysis.

8.2.4 Stage 4: Examine relationships

Bivariate analysis should follow from a question. Useful comparisons include outcomes by group, counts and rates over time, numeric outcomes across categories, relationships between numeric variables, and missingness by source or period.

Always distinguish counts from rates. A facility with more visits may simply serve a larger population or operate more hours. A program with more cancellations may also have far more registrations.

8.2.5 Stage 5: Add relevant dimensions

Multivariate exploration examines whether an apparent relationship changes after another variable is considered. Faceting, stratification, grouped summaries, correlation analysis, and exploratory models can reveal that a system-wide pattern is concentrated in one location, period, or group.

Identifier codes and arbitrary category numbers should not be placed in a correlation matrix simply because software stores them as numbers.

8.2.6 Stage 6: Convert findings into decisions

End with verified facts, important limitations, patterns requiring explanation, candidate problems or opportunities, next analytical questions, and decisions about exclusions or scope.

References

Quartz. n.d. “The Quartz Guide to Bad Data.” https://github.com/Quartz/bad-data-guide.