Inspecting a Dataset

When a new dataset arrives, the temptation is to dive straight into the interesting question — "do customers in region X spend more than customers in region Y?" — and start running analyses.

That is almost always a mistake. The very first thing every experienced analyst does is something more modest: find out what you actually have. How big is it? What are the columns? Are there missing values? Are the numbers in sensible ranges? Are there obviously wrong rows?

This page is the ritual. Memorize it.

The five-minute ritual

Each step takes seconds and gives you orientation. Skip any of them and you risk asking a meaningless question of a misunderstood dataset.

We'll use airquality — a built-in dataset of daily air quality measurements in New York, May–September 1973 — as our example.

Step 1: shape

The shape tells you what kind of analysis is even feasible. 153 rows × 6 columns is small — every operation will be instant. A million rows × 200 columns is a different problem entirely.

Step 2: peek

Glance at the first few and last few rows. You're checking:

• Did the data load correctly? (No mangled headers, no obvious parsing errors.)
• Do the values look like what you'd expect?
• Are there obvious anomalies?

It is astonishing how often a 10-second head() catches a problem that would have wasted an hour downstream.

Step 3: structure and types

str() packs an enormous amount of information into a few lines: how many rows, how many columns, every column name, every column type, and a sample of values.

What you're looking for:

• Are types right? (Is the "amount" column truly numeric, or did it accidentally get loaded as character because of a stray comma?)
• Are dates being treated as dates or as strings?
• Are categorical variables stored as factors, characters, or numbers?

Step 4: summaries

summary() gives you, per column:

• For numerics: min, 1st quartile, median, mean, 3rd quartile, max, and the count of NAs
• For factors: counts per level
• For characters: just the count

You're scanning for:

• Implausible values. Negative ages. Heights of 3 meters. Temperatures of 999 (a common "missing" sentinel).
• Suspiciously round means. A mean of exactly 0 in a column that should be variable is suspicious.
• Mean very different from median. Suggests skew or outliers.
• Lots of NAs.

In airquality, look at the Ozone column. The max is 168 but the median is around 30 — a heavy right tail. That's a real finding, not a glitch, but worth noting.

Step 5: missingness

The default summary() does report NAs, but it's often clearer to ask explicitly:

In airquality, Ozone and Solar.R have meaningful amounts of missing data. That will affect how you compute averages, how you plot, how you join with other data, everything. Knowing it upfront lets you choose your strategy.

A multi-file walkthrough: meet iris

Let's apply the ritual to a fresh dataset, with the inspection split across files so you can see each step clearly:

Test your understanding

QuestionSelect one

Which function gives you a compact view of a data frame's structure: number of rows, number of columns, every column's type, and a sample of its values?

summary()

head()

str()

dim()

QuestionSelect one

Why is it a bad idea to skip the "inspect the dataset" ritual?

It's required by R; the language won't compute without it.

Because real datasets contain surprises (missing values, wrong types, implausible values) that can silently corrupt every downstream analysis.

Because summary() is the only way to compute means.

Because it's a tradition.

QuestionSelect one

If mean(airquality$Ozone) returns NA, the most likely reason is:

The mean function is broken.

The dataset has fewer than 30 rows.

The Ozone column contains NA values, and mean() returns NA unless told otherwise.

R refuses to compute means on built-in datasets.

Inspecting tells us what's there. The next page is about extracting just the parts we want.