Data Cleaning, Coding and Visualization for Research Datasets

Survey platforms, lab instruments, administrative records and spreadsheets all export data in their own ways: text instead of numbers, several codes for 'missing', duplicated respondents, dates in mixed formats and variable names like Q14_3_TEXT. Cleaning is not cosmetic. How you define missing values, treat extreme scores, score scales and join files affects every later result, so these decisions need to be deliberate and reported.

We clean data with code (in SPSS syntax, Stata, R or Python, matching your analysis software), never by editing cells by hand. The raw file stays untouched, a script produces the clean file, and a cleaning log explains each decision in plain language. You also get a codebook that describes every variable, which your supervisor, coauthors and future self will thank you for.

Visualization is part of the same work. Exploratory plots expose problems — impossible values, ceiling effects, odd distributions — before they reach a model, and well-designed final figures make results easier to understand. We apply established design principles to both.

A good coding scheme is predictable. Likert items run in the same direction with the same numeric range, yes/no items are coded 0/1 so means become proportions, and each missing reason has its own code (for example −8 'not applicable', −9 'refused') so you can tell them apart before setting them to missing. Open-text 'other' answers are read and either back-coded into existing categories or given new ones, with the rule written down.

Reverse-worded items are recoded into new variables (never overwriting the original) using the scale's range, so on a 1–5 scale a response x becomes 6 − x. We then check that recoded items correlate positively with the rest of the scale before computing scores, which catches the most common scale-scoring mistake.

There is rarely one correct answer for missing data or extreme values, but there are indefensible ones: deleting cases without saying so, or replacing missing values with the mean. We describe how much is missing and whether it looks related to other variables, then agree an approach suited to your analysis — often full information maximum likelihood or multiple imputation for model-based work, or complete-case analysis with a clear justification when missingness is minimal.

For outliers, we first ask whether a value is possible. A reaction time of 3 ms or an age of 212 is an error to correct or remove. A genuinely extreme but valid value is kept unless there is a pre-stated rule, and results are checked with and without it. The cleaning log records the number of cases affected at each step, which lets you write a clear participant-flow description.

A codebook is the dictionary of your dataset. Ours list every variable's name, label, type, valid values and labels, missing codes, the source question, and — for derived variables — the exact formula. It is useful immediately, when you write your methods section, and later, when you deposit data or return to the project after months away.

Where you plan to share data, we align the codebook with repository expectations and your data management plan, and flag variables that could identify participants in combination, such as detailed dates, postcodes or rare job titles.

We choose charts by the question: distributions with histograms, density or box-and-dot plots; comparisons with dot plots or bars that start at zero; change over time with lines; relationships with scatterplots and a fitted line; and uncertainty with confidence intervals rather than bare means. Showing the raw data points alongside summaries, when sample size allows, helps readers judge the result themselves.

Finished figures use readable fonts at print size, direct labels instead of distant legends, colorblind-safe palettes that also work in grayscale, and consistent axes across panels that will be compared. We avoid 3D effects, truncated bar axes and dual y-axes that exaggerate or obscure patterns.