Age Ad_Spend Clicks Purchases Revenue
Min. :18.00 Min. : 1.83 Min. : 0.00 Min. : 0.000 Min. : 0.00
1st Qu.:31.00 1st Qu.: 96.86 1st Qu.: 10.00 1st Qu.: 0.000 1st Qu.: 0.00
Median :43.00 Median : 172.25 Median : 19.00 Median : 1.000 Median : 26.86
Mean :42.51 Mean : 216.78 Mean : 24.17 Mean : 1.047 Mean : 41.21
3rd Qu.:54.00 3rd Qu.: 278.74 3rd Qu.: 32.00 3rd Qu.: 2.000 3rd Qu.: 61.28
Max. :65.00 Max. :3875.35 Max. :322.00 Max. :17.000 Max. :702.61 Exploratory Data Analysis: Covariation
MKT 566 · Fall 2026
Task 3: the distribution of ad spend
What shape is this, and what marketing strategy could produce it?
Task 3: reading it
- Right-skewed: the mean is to the right of the median, as the summary table suggested
- A likely strategy: a basic spend on everyone, plus extra budget on a few customers the firm thinks are high-value (retargeting, lookalike audiences)
Task 4: the same chart on a log scale
What can you see now that was invisible before?
Task 4: reading it
- On the log scale the distribution is roughly symmetric. The “typical” customer is easy to see: $150–200
- The low-spend customers were all piled up on the left before. Now they are spread out. Logs stretch small values and squeeze large ones
Task 5: the channel mix
Balanced? Does the mix make sense for an online retailer?
Task 5: reading it
- Not balanced: Search is a third of customers, Email one in twenty
- It makes sense: search reaches customers who already want to buy, social and display find new customers, and email only reaches customers the firm already knows
Task 6: hunting outliers
How many outliers, on which side, and should we drop them?
Task 6: reading it
- The 1.5 × IQR rule flags 39 customers, all on the high side. On the log scale only a few remain. Most “outliers” are just the tail of a skewed distribution, not errors
- Do not drop them without a reason. Run later analyses with and without them, say what you did, and prefer medians to means
Task 7: your own question (two examples)
Same two groups, two different variables, both on a log scale. What is different between the two panels?
Task 7: reading it
- Spend is the same on both devices; clicks are higher on mobile. The firm spends the same but gets more clicks on mobile
- Why “clicks + 1”? Six customers have zero clicks, and the log of zero does not exist. Adding 1 keeps them on the chart
Sales vs. total ad spend
marketing$total_ad_spend <-
marketing$youtube + marketing$facebook + marketing$newspaper
ggplot(marketing, aes(x = total_ad_spend, y = sales)) +
geom_point(alpha = 0.6, color = "darkgreen") +
geom_smooth(method = "lm", se = FALSE) +
labs(title = "Sales vs. Total Ad Spend",
x = "Total ad spend (thousands $)",
y = "Sales (thousands of units)") +
theme_minimal()
Each point is one campaign. What do you see?
Sales vs. total ad spend: reading it
- A clear positive relationship: campaigns that spend more sell more
- Roughly a straight line, with a few high-spend campaigns that sold little
- The fitted line (
geom_smooth(method = "lm")) is a preview of next week: regression
Can we do a more informative viz?
marketing_long <- pivot_longer(marketing,
cols = c(youtube, facebook, newspaper),
names_to = "channel",
values_to = "ad_spend")
ggplot(marketing_long,
aes(ad_spend, sales, color = channel)) +
geom_point(alpha = 0.6) +
geom_smooth(method = "lm", se = FALSE) +
facet_wrap(~channel, scales = "free_x") +
labs(title = "Sales vs. Spend by Channel",
x = "Ad spend (thousands $)",
y = "Sales") +
theme_minimal() +
theme(legend.position = "none")
Same 200 campaigns, one panel per channel. What changed?
One panel per channel: reading it
- YouTube: strong, tight relationship. Facebook: positive but noisier. Newspaper: almost flat
- The total-spend chart hid this difference. Splitting a chart by a categorical variable (
facet_wrap) is the most useful move in EDA
Overplotting: when you have too many points
ggplot2’s diamonds dataset, 53,940 rows. What is the problem with this chart?
Overplotting: two fixes
- Transparency (
alpha) + a smooth curve shows where most points are and the curved shape of the relationship - 2D bins (
geom_bin2d) count the points in each cell: a heatmap. Use either one when a scatter plot turns into a blob (RateBeer: 290,000 reviews)
Purchases by gender
avg_gender <- df %>%
group_by(Gender) %>%
summarise(avg_purchases = mean(Purchases),
se_purchases = sd(Purchases) / sqrt(n()))
ggplot(avg_gender, aes(x = Gender, y = avg_purchases, fill = Gender)) +
geom_col(width = 0.6) +
scale_fill_manual(values = c(F = "#E07A5F", M = "#3D5A80")) +
labs(title = "Average Purchases by Gender",
x = NULL, y = "Average purchases") +
theme_minimal() + theme(legend.position = "none")
First summarize (one row per group), then chart the summary. What do you conclude?
Purchases by gender: reading it
- Men average 1.09 purchases, women 1.01. Is the difference real, or would another 1,000 customers flip it?
- The bars cannot tell you. We need to know how precise each average is
Add error bars
The standard error, \(SE = sd / \sqrt{n}\), tells us how precise the average is. Mean ± 1.96 × SE is a 95% confidence interval: the range where the true average most likely is. Now what do you conclude?
Error bars: reading it
- The two intervals overlap a lot: the data cannot tell men and women apart on purchases
- Rule of thumb: bars of averages without error bars make small differences look real. Always ask for error bars
Add a third variable: the ad channel
facet_wrap(~Channel) again. What do you see now?
Gender × channel: reading it
- The gap changes by channel: men buy more via Search (1.31 vs. 1.09), women via Video (1.20 vs. 0.94)
- Email has the widest error bars: only 54 customers. Small group, wide interval. Always check how many people are behind a bar
Averages hide distributions
One boxplot per group, sorted by median. Do the channels differ, and how?
Boxplots by group: reading it
- The medians are almost the same: the typical customer gets the same spend in every channel
- The differences are in the tails: Video and Social have more high-spend outliers. An average would say “Video spends more”. The boxplot says “Video spends more on a few customers”. A different conclusion
Channel and device: how many customers in each cell?
A heatmap: color = count, number on top. What do you see?
Counts heatmap: reading it
- Mobile is the majority in every channel (65% overall)
- Video is the most mobile-heavy (125 vs. 53): video ads show up where people watch video
- A heatmap is just a table with color. For two categorical variables, that is usually all you need
Add a third variable: average ad spend per cell
Same grid, the color is now the average spend in the cell. Anything stand out?
Spend heatmap: reading it, and the EDA loop
- Video / Desktop at $315 is far above every other cell. Is it a real pattern, or a few outliers in a cell with only 53 customers?
- To find out, go back and make a boxplot of that cell (it is in the class script). This is the EDA loop: notice → ask → check → refine