EDA in R: A 7-Step Framework That Works on Every Dataset You'll Encounter

Exploratory Data Analysis (EDA) is the process of examining a dataset before building any model or running any test, you look at structure, spot missing values, check distributions, flag outliers, and uncover relationships so that every downstream decision rests on evidence, not assumptions.

Most tutorials treat EDA as a grab bag of random plots and summary tables. This tutorial gives you a repeatable 7-step framework you can apply to any dataset: structure, missingness, distributions, outliers, correlations, group comparisons, and time patterns. We'll work through every step with R's built-in airquality dataset and the tidyverse.

What Does Your Data Actually Look Like?

Every EDA starts with the same question: what am I working with? Before you plot a single chart or compute a single statistic, you need to know how many rows and columns you have, what type each variable is, and whether R read them in correctly.

Right away you can see three important things. First, the dataset has 153 rows and 6 columns, a manageable size for learning. Second, Ozone and Solar.R already show NA values, which tells you missing data will matter. Third, Month and Day are stored as integers, not as date or factor types, something you might want to fix later.

The str() function gives you a more compact view that emphasises storage types and shows the first few values of each column.

Now Month is a factor with readable labels instead of bare integers. This small fix pays off in every plot and summary from here on, ggplot2 will label axes automatically instead of showing "5, 6, 7, 8, 9".

Where Is Your Data Missing, and Does It Matter?

Missing data isn't just an inconvenience, it can silently bias every analysis you run. A correlation computed on only complete cases might overrepresent certain months. A mean calculated after dropping NAs might miss the fact that values are missing because they were extreme. The pattern of missingness matters as much as the amount.

Let's start by counting exactly how many values are missing in each column.

Ozone has 37 missing values, that's 24% of the dataset. Solar.R has 7 (5%). Wind, Temp, Month, and Day are complete. This immediately tells you that any analysis involving Ozone needs to handle missing values carefully.

Next, let's see where those gaps fall. Are they random, or do they cluster in certain months?

The heatmap reveals that Ozone's missing values aren't perfectly random, there are clusters, especially around rows 25-30 and rows 95-100. This kind of pattern might indicate sensor failures on consecutive days.

About 72.5% of rows are complete. If you naively dropped all incomplete rows, you'd lose over a quarter of your data. That's a steep price, especially if the missing data is concentrated in certain months, which would bias your seasonal analysis.

How Are Your Variables Distributed?

Distributions tell you the shape of your data, is it symmetric, skewed, spread out, or concentrated? This matters because many statistical methods assume a particular shape (often a bell curve), and blindly applying them to skewed data gives misleading results.

Let's start with a histogram of Ozone, the variable with the most interesting behaviour.

The histogram shows a clear right skew, most days have low Ozone (under 50 ppb), but some days spike above 100 ppb. This skew means the mean will be pulled higher than the median, and any method assuming normality will be off.

Figure 1: Decision tree for choosing the right distribution plot based on variable type and sample size.

Now let's compare how Temperature is distributed across months. A density plot works well here because it smooths out the histogram bumps and makes overlapping distributions easier to compare.

The density plot reveals a clear seasonal shift, May's distribution sits around 60-70F, while July and August peak around 80-85F. September's distribution is wider, reflecting the transition from summer to autumn.

The summary() function gives you the five-number summary (min, Q1, median, Q3, max) plus the mean in one call.

Notice how Ozone's mean (42.1) is well above its median (31.5), that confirms the right skew we saw in the histogram. When mean and median diverge like this, always report the median as the "typical" value.

Which Values Are Outliers, and What Should You Do About Them?

An outlier is a data point that sits far from the rest. It might be a sensor malfunction, a data entry mistake, or a genuinely extreme event (like a heat wave). The important question isn't "is it an outlier?" but "why is it an outlier?", because the answer determines what you do about it.

The boxplot is the classic outlier detection tool. Points beyond the whiskers (1.5 times the interquartile range) are flagged automatically.

The boxplot shows outlier dots above the whiskers in several months. August has the highest median Ozone, and May has a couple of extreme high values that stand out from an otherwise low-Ozone month.

Let's detect those outliers programmatically using the IQR method.

Only two values (135 and 168 ppb) exceed the upper bound of 131.1. These aren't impossible, real Ozone spikes happen during heat waves. But let's see how much they affect the mean.

Figure 2: Decision flow for handling outliers: fix errors, keep meaningful extremes, or test both ways.

The two outliers shift the mean by only 1.7 ppb, a small effect. In this case, keeping them is reasonable because they represent real atmospheric events, not errors. If the difference were much larger, you'd want to run your downstream analysis both ways and report the sensitivity.

Which Variables Are Related to Each Other?

Correlation measures the strength and direction of a linear relationship between two numeric variables. It ranges from -1 (perfect negative, as one goes up, the other goes down) to +1 (perfect positive, they move together). A value near 0 means no linear relationship, but there might still be a curved one.

Let's compute the correlation matrix for all numeric columns.

Three relationships jump out. Ozone and Temp have a strong positive correlation (0.70), hotter days produce more Ozone. Ozone and Wind show a moderate negative correlation (-0.60), windier days have lower Ozone, likely because wind disperses pollutants. Solar.R has a weak positive link to Ozone (0.35), which makes physical sense since sunlight drives Ozone formation.

Let's visualize the strongest relationship with a scatter plot.

The scatter plot confirms the positive relationship, but the loess smoother reveals something the correlation number hides: the relationship isn't perfectly linear. Below 75F, Ozone is relatively flat. Above 80F, it accelerates sharply. This non-linearity means a simple correlation of 0.70 actually understates how strongly temperature drives Ozone on hot days.

A pairs plot gives you every pairwise scatter plot at once, useful for quickly scanning all relationships.

The pairs plot confirms what we found: Ozone-Temp is the strongest visual pattern, Ozone-Wind shows a clear downward trend, and Solar.R has weaker relationships with everything else.

How Do Groups Differ Across Your Variables?

Grouping reveals structure that global summaries hide. A variable that looks well-behaved overall might behave completely differently in subgroups. In the airquality data, Month is the natural grouping variable, summer months behave differently from spring and autumn.

Let's start with a grouped summary of Ozone by Month.

July and August have Ozone levels roughly double those of May and September. The standard deviation is also much higher in summer, not only is Ozone worse on average, but it's more unpredictable. September drops back down, roughly matching June despite higher temperatures. This suggests temperature alone doesn't explain Ozone patterns.

Side-by-side boxplots make these group differences visually immediate.

The boxplot reveals a clear seasonal arc, temperatures climb steadily from May through August, then drop slightly in September. The spread (box height) is similar across months, meaning temperature variability is roughly constant regardless of season.

Faceted histograms let you compare the shape of distributions across groups, not just their centres.

The faceted view reveals something the grouped summary didn't make obvious: May and June have most values near zero with a few spikes, while July and August are spread across a much wider range. The shape of the distribution changes across months, not just the centre.

Are There Patterns Over Time?

If your data has a time dimension, trends and seasonality can dominate everything else. Two variables might look correlated, but the real story is that both are simply rising (or falling) over time. The airquality dataset spans May through September 1973, giving us a clear temporal axis.

Let's construct a proper date column and plot Ozone over time.

The time series shows high day-to-day variability, but a clear seasonal envelope, Ozone peaks in July and August, then drops in September. The gaps in the line correspond to the missing values we identified in Step 2.

Adding a smoothed trend line makes the seasonal pattern easier to see.

Temperature follows a smooth seasonal arc, rising steadily from May through late July, plateauing in August, then declining in September. Compare this shape to the Ozone time series above: they move together, confirming the strong positive correlation we found in Step 5.

To compare multiple variables on the same time scale, reshape the data and facet.

The multi-panel view delivers the final insight: Ozone and Temperature share a seasonal pattern (both peak mid-summer), but Wind shows no clear seasonal trend, it's variable throughout the entire period. This tells you that Wind's negative correlation with Ozone is driven by day-to-day weather, not seasonal cycles.

Practice Exercises

Exercise 1: Run the first 3 steps on mtcars

Apply Steps 1-3 of the EDA framework to the mtcars dataset: examine its structure with glimpse(), check for missing values with colSums(is.na()), and create a histogram of mpg. Write one sentence summarising each finding.

Exercise 2: Build an EDA summary function

Create a function eda_summary() that takes a numeric vector, handles NAs, and returns a named list with: mean, median, sd, n_missing, n_outliers (IQR method), and skewness direction ("left", "symmetric", or "right" based on whether mean < median, mean ≈ median, or mean > median). Test it on every numeric column in airquality.

Exercise 3: Mini EDA report on iris

Build a mini EDA report for the iris dataset: (1) structure check with glimpse(), (2) faceted histogram of Sepal.Length by Species, (3) correlation matrix of all four numeric columns, and (4) a grouped summary table with mean and sd of all numeric columns by Species.

Putting It All Together

Let's run all 7 steps in a single cohesive analysis of airquality, producing a written summary at the end. This mirrors what a real EDA section of a report looks like.

Here's the written summary that would go into a report:

The airquality dataset contains 153 daily observations of 6 variables from May-September 1973 in New York. Ozone has substantial missing data (24%), concentrated in June (33%). Ozone is right-skewed with 2 IQR outliers (135 and 168 ppb). Temperature and Ozone are strongly correlated (r = 0.70) with a non-linear acceleration above 80F. Wind is negatively correlated with Ozone (r = -0.60). Monthly grouping reveals summer peaks in Ozone (July-August, ~60 ppb) coinciding with the temperature peak, while Wind shows no seasonal pattern. Any model of Ozone should include Temperature and Wind as predictors and account for the non-linear temperature effect.

Summary

Figure 3: The complete 7-step EDA framework as a sequential pipeline.

The framework is a checklist, not a strict sequence. Skip Step 7 if there's no time dimension. Spend extra time on whichever step reveals the most interesting patterns. The goal is to build intuition about your data before you commit to any model or test.

References

1. Tukey, J.W., Exploratory Data Analysis. Addison-Wesley (1977). The foundational text that introduced EDA as a discipline.
2. Wickham, H. & Grolemund, G., R for Data Science (2e), Chapter 10: Exploratory Data Analysis. Link
3. R Core Team, airquality dataset documentation. Link
4. Wickham, H., ggplot2: Elegant Graphics for Data Analysis (3e). Springer (2024). Link
5. dplyr documentation, group_by() and summarise(). Link
6. NIST/SEMATECH, e-Handbook of Statistical Methods: Exploratory Data Analysis. Link
7. Grolemund, G., Hands-On Programming with R. O'Reilly (2014). Link

Continue Learning

• Automated EDA in R, Packages that generate EDA reports automatically, so you can compare hand-coded EDA with automated output.
• Missing Data Visualization with naniar, Deep dive into visualizing and understanding missing data patterns beyond what base R offers.
• Outlier Detection in R, Statistical and visual methods for outlier detection, expanding on Step 4 of this framework.