Worksheet 05 — Linear Regression

Building a calibration curve to predict leaf area from paper tracing mass

r-basics

project-setup

Doing linear regressions and plotting

Author

Bill Perry

Published

July 5, 2026

Regression — Predicting Leaf Area

Recap from the Quarto intro (Worksheet 4.5)

Quarto — one .qmd file holds your writing and your code and renders to Word, HTML, or PowerPoint
Three ingredients — YAML front matter, Markdown text, and uniquely named code chunks
Reproducibility — every number and figure is computed, never copied by hand
Inline code — a live result drops straight into a sentence with `r `
Carried forward (Worksheet 04) — shady leaves were significantly heavier than sunny leaves (Welch’s t-test, p < 0.001)
But: weight in grams tells us about mass — not directly about leaf surface area

Today’s Objectives

Load and inspect the paper calibration data
Make a scatter plot and visually assess linearity
Fit a linear regression with lm() and read every line of summary()
Extract slope, intercept, and R² from the model
Check all four regression assumptions (linearity, equal variance, normality of residuals)
Use the calibration equation to predict leaf area from a tracing mass
Use predict() with confidence and prediction intervals
Write a complete results paragraph following W&S Chapter 17 conventions

How to use this worksheet

Work through each part in order. Type the code into a new R script in Positron and run it line by line.

Code blocks marked ▶ Run this should be executed as written.

Blocks marked ✏️ Your turn ask you to write, modify, or interpret.

The Going further section is optional.

🔮 Predict before you run — and type, don’t paste

Before each ▶ Run this block, cover the output and predict what R will print. Then type the code yourself. Two reasons:

Predicting makes it stick. Guessing the slope, R², or a residual pattern forces you to retrieve the idea — being surprised is when it sinks in.
Typing beats pasting. Typing trains your eye for the small errors (a flipped lm() formula, a wrong newdata column name) that you would otherwise spend real time hunting.

📂 The same analysis, two ways — download both and compare

Today you get to feel the difference the Quarto intro talked about. These two files do the exact same regression — one as a plain script, one as a report:

regression_analysis_script.R — the analysis as a classic R script. Great for running the numbers. To share results you would copy them into a document by hand.
regression_analysis_report.qmd — the analysis as a Quarto report. Same code, but wrapped in writing, with figures and numbers that render straight into a Word document.

Open them side by side in Positron. As you work through this worksheet, notice which one you would rather hand in.

🧩 Chunk 1 — Set up the data (after lecture Chunk 1)

Parts 1–3: load the paper data, inspect it, and make the scatter plot.

Part 1 · Load libraries and data

Libraries

▶ Run this at the top of your script:

# Load packages at the top — always ----------------------
library(readxl)      # reading Excel files
library(tidyverse)   # data manipulation + ggplot2
library(skimr)       # fast dataset overview

The calibration dataset

The file paper_area_weights.xlsx contains 118 paper squares of known area (1–567 cm²) with their measured masses in grams. We will fit a regression to convert mass → area, then apply it to leaf tracings.

▶ Run this:

# Load the paper calibration data -----------------------
paper_df <- read_excel("data/paper_area_weights.xlsx")

✏️ Your turn: Run dim(paper_df). How many rows and columns does the dataset have?

# Write your code here:

Rows:
Columns:
Column names and their roles:
  area_cm2 →  (response / explanatory / circle one)
  mass_g   →  (response / explanatory / circle one)

Part 2 · Inspect the data

▶ Run this:

# Quick overview with skim --------------------------------
skim(paper_df)

# How many squares were measured at each area level? -----
paper_df %>%
  group_by(area_cm2) %>%
  summarize(
    n          = n(),
    mean_mass  = round(mean(mass_g), 5),
    min_mass   = round(min(mass_g),  5),
    max_mass   = round(max(mass_g),  5)
  )

✏️ Your turn: Fill in the table from the output above:

area_cm2   mean mass (g)
--------   -------------
1
4
16
64
100
400

✏️ Your turn: Is there a clear pattern — as area increases, what happens to mass? Is this what you expected from your knowledge of paper?

Your answer:

Part 3 · First scatter plot

▶ Run this:

# Scatter plot — check the relationship visually ---------
scatter_09_plot <- paper_df %>%
  ggplot(aes(x = mass_g, y = area_cm2)) +
  geom_point(alpha = 0.5, size = 2) +
  labs(
    title = "Paper Mass vs. Known Area",
    x     = "Mass (g)",
    y     = "Area (cm²)"
  ) +
  theme_minimal()

scatter_09_plot

✏️ Your turn: Describe the relationship. Is it linear? Is the variance (spread of points) roughly constant across the range of mass values?

Shape of relationship (linear / curved):
Variance (constant / increases / decreases with X):
Any obvious outliers?  Y / N

✏️ Your turn: In this study, which variable is the explanatory (X) and which is the response (Y)? Why does that direction make sense for our goal of predicting leaf area?

Explanatory (X):
Response (Y):
Reason for this direction:

💡 Key idea: In regression, X is what you measure (mass — easy with a balance), and Y is what you want to know (area — harder to measure on a complex leaf shape). We measure X to predict Y.

🧩 Chunk 2 — Fit & read the model (after lecture Chunk 2)

Parts 4–7: fit lm(), decode summary(), pull the equation and R².

Part 4 · Fit the regression with `lm()`

🔮 Predict first: area rises with mass. Before you read summary() — will the slope be positive or negative, and roughly how many cm² per gram?

▶ Run this:

# Fit the least-squares linear model --------------------
# lm(response ~ predictor, data = data_frame)
paper_lm_model <- lm(area_cm2 ~ mass_g, data = paper_df)

# Full summary — every number matters -------------------
summary(paper_lm_model)

✏️ Your turn: Copy the Coefficients table output here:

Paste the Coefficients section of the summary() output:

Part 5 · Decode the `summary()` output line by line

✏️ Your turn: Match each piece of output to its meaning. Fill in the blanks:

(Intercept) estimate = _____
  → This is "a" in Ŷ = a + bX.
    It is the predicted area when mass = _____.
    Does this make physical sense? (a piece of paper with zero mass has zero area)  Y / N

mass_g estimate = _____
  → This is "b" (the slope).
    It means: for every additional 1 gram of paper, area increases by _____ cm².
    Biologically, 1 gram of this paper stock has _____ cm² of surface area.

Std. Error for mass_g = _____
  → The uncertainty in the slope estimate.

t value for mass_g = _____
  → slope / SE.  This tests H₀: β = _____.

Pr(>|t|) for mass_g = _____
  → The p-value for the slope test. Is it significant at α = 0.05?  Y / N

Multiple R-squared = _____
  → _____ % of the variation in area is explained by mass.

F-statistic = _____  on _____ and _____ df
  → Overall test of the model.  p = _____

📖 Whitlock & Schluter §17.3 p. 551–553: The slope is tested with a t-statistic: t = b / SE_b, with df = n − 2. If H₀: β = 0 is rejected, mass significantly predicts area.

Part 6 · Extract the calibration equation

▶ Run this:

# Pull intercept and slope from the model ---------------
a_intercept <- coef(paper_lm_model)[1]
b_slope     <- coef(paper_lm_model)[2]

cat("Intercept (a) =", round(a_intercept, 4), "cm²\n")
cat("Slope (b)     =", round(b_slope, 2), "cm²/g\n")

✏️ Your turn: Write the calibration equation for predicting area from mass:

area_cm2 = _____ × mass_g + _____

✏️ Your turn: Use the equation by hand to predict the area of a paper tracing that weighs 0.120 g. Show your work:

area = _____ × 0.120 + _____
     = _____  cm²

Part 7 · Extract and interpret R²

🔮 Predict first: For uniform paper, predict R² — near 0, 0.5, or 1? Write your guess before you run it.

▶ Run this:

# Pull R-squared from the model summary -----------------
r2_val <- summary(paper_lm_model)$r.squared
cat("R² =", round(r2_val, 6), "\n")

✏️ Your turn: What does R² = 0.9999 mean in plain language for our calibration?

Your answer:

✏️ Your turn: In ecology, you might see regression results with R² = 0.45. Is that a useful result? What does it mean?

Your answer:

📖 Whitlock & Schluter §17.4 p. 555: “R² measures the fraction of variation in Y that is explained by X.” A high R² indicates a tight fit; a low R² indicates high scatter around the line — but a low R² does not mean the relationship is not real or useful.

🧩 Chunk 3 — Visualize & check assumptions (after lecture Chunk 3)

Parts 8–10: plot the curve, then check the residual and QQ/Shapiro assumptions.

Part 8 · Plot the calibration curve

▶ Run this:

# Regression line with 95% confidence band --------------
calibration_plot <- paper_df %>%
  ggplot(aes(x = mass_g, y = area_cm2)) +
  geom_point(alpha = 0.5, size = 1.8) +
  geom_smooth(method = "lm", se = TRUE,
              color = "steelblue", fill = "lightblue") +
  labs(
    title    = "Paper Calibration Curve",
    subtitle = paste0("area = ", round(b_slope, 2),
                      " × mass + ", round(a_intercept, 3),
                      "  |  R² = ", round(r2_val, 4)),
    x        = "Paper Mass (g)",
    y        = "Area (cm²)"
  ) +
  theme_minimal()

calibration_plot

✏️ Your turn: Where is the shaded confidence band narrowest? Where is it widest? Why does the width vary?

Narrowest at:
Widest at:
Why it varies:

Save the plot

▶ Run this:

ggsave("figures/paper_calibration_curve.png",
       plot   = calibration_plot,
       width  = 5, height = 5, units = "in", dpi = 300)

Part 9 · Check assumptions — residual plot

🔮 Predict first: If the line fits well, what should the residuals-vs-fitted plot look like — a flat random cloud, a curve, or a funnel? Predict, then run.

▶ Run this:

# Add fitted values and residuals to the data ----------
paper_resid_df <- paper_df %>%
  mutate(
    fitted    = fitted(paper_lm_model),
    residuals = residuals(paper_lm_model)
  )

# Residuals vs Fitted — checks linearity and equal variance
resid_09_plot <- paper_resid_df %>%
  ggplot(aes(x = fitted, y = residuals)) +
  geom_point(alpha = 0.5) +
  geom_hline(yintercept = 0, linetype = "dashed", color = "red") +
  labs(
    title = "Residuals vs Fitted Values",
    x     = "Fitted Values (cm²)",
    y     = "Residuals (cm²)"
  ) +
  theme_minimal()

resid_09_plot

✏️ Your turn: Describe the residual plot. Do the points form a random cloud around zero, or is there a pattern?

Pattern or random cloud?:
Any evidence of non-linearity (curved pattern)?  Y / N
Any evidence of unequal variance (funnel shape)?  Y / N

⚠️ Watch out! A funnel shape (residuals getting larger at higher fitted values) means variance is not constant — a violation of the equal variance assumption. A curved pattern means the relationship is not linear.

📖 Whitlock & Schluter §17.5 p. 559 (Figure 17.5-4): “A residual plot should show a roughly symmetric cloud of points with no pattern.”

Part 10 · Check assumptions — QQ plot and Shapiro-Wilk

▶ Run this:

# QQ plot of residuals — checks normality assumption ---
qq_resid_09_plot <- paper_resid_df %>%
  ggplot(aes(sample = residuals)) +
  stat_qq() +
  stat_qq_line(color = "red", linewidth = 0.8) +
  labs(
    title = "Normal QQ Plot of Residuals",
    x     = "Theoretical Quantiles",
    y     = "Sample Quantiles"
  ) +
  theme_minimal()

qq_resid_09_plot

# Shapiro-Wilk test on residuals — NOT on raw data -----
shapiro.test(residuals(paper_lm_model))

✏️ Your turn: Record the Shapiro-Wilk result:

W = _____   p = _____
Decision (normal / not normal):

✏️ Your turn: An important distinction — in regression, normality is checked on the residuals, not on Y or X directly. Why does that matter?

Your answer:

🧩 Chunk 4 — Predict & report (after lecture Chunk 4)

Parts 11–12: predict leaf area from tracing mass, then write the results paragraph.

Part 11 · Predict leaf area from tracing mass

🔮 Predict first: sunny tracing 0.092 g, shady 0.138 g. Which predicts the larger area? Does that match “shady leaves are bigger”? Predict before running.

Method 1 — Direct calculation

▶ Run this:

# Direct calculation using the calibration equation ----
mass_sunny <- 0.092   # g — example sunny leaf tracing
mass_shady <- 0.138   # g — example shady leaf tracing

area_sunny <- b_slope * mass_sunny + a_intercept
area_shady <- b_slope * mass_shady + a_intercept

cat("Sunny tracing:", mass_sunny, "g →",
    round(area_sunny, 2), "cm²\n")
cat("Shady tracing:", mass_shady, "g →",
    round(area_shady, 2), "cm²\n")

✏️ Your turn: Does the shady leaf have a larger predicted area? In Worksheet 04 we found shady leaves were significantly heavier. Does a larger predicted area match that finding? What does this tell you about the relationship between leaf weight and leaf area?

Sunny predicted area:
Shady predicted area:
Larger side:
Does larger area match the heavier weight from Worksheet 04?  Y / N
What does this tell you about the relationship between weight and area in leaves?

Method 2 — Using `predict()` with intervals

▶ Run this:

# predict() gives uncertainty around predictions ------
new_masses <- tibble(mass_g = c(0.092, 0.138))

# 95% confidence interval for the MEAN area
predict(paper_lm_model,
        newdata  = new_masses,
        interval = "confidence") %>%
  as_tibble() %>%
  mutate(mass_g = c(0.092, 0.138), .before = 1) %>%
  round(3)

# 95% prediction interval for a SINGLE observation
predict(paper_lm_model,
        newdata  = new_masses,
        interval = "prediction") %>%
  as_tibble() %>%
  mutate(mass_g = c(0.092, 0.138), .before = 1) %>%
  round(3)

✏️ Your turn: The prediction interval is wider than the confidence interval. In your own words, explain why — what additional source of uncertainty does a prediction interval capture?

Your answer:

📖 Whitlock & Schluter §17.2 p. 549: “Confidence bands measure the precision of the predicted mean Y for each value of X. Prediction intervals measure the precision of the predicted single Y-values for each X.”

Part 12 · Write a scientific results paragraph

✏️ Your turn: Write a complete results paragraph using the information below. Follow the format from Lecture 05.

Information to include: - What the regression tested - F-statistic, df₁, df₂, and p-value (from the F-statistic line in summary()) - R² - The calibration equation (slope and intercept) - 95% confidence interval for the slope (use confint(paper_lm_model)) - Brief mention of assumption checks

# Get the 95% CI for the slope -------------------------
confint(paper_lm_model)

Write your results paragraph here:

✍️ Now put it in a report

Open regression_analysis_report.qmd. Its Discussion section already pulls the slope, R², and p-value into the sentences with inline code. Paste your results paragraph in, press Render, and you have a Word document — no numbers copied by hand.

Part 13 · Review and checkpoint

At this point you should be able to:

✏️ Your turn — before you move on: Run your entire script with Ctrl/Cmd + Shift + Enter. Does it run from top to bottom without errors?

Ran cleanly?  Y / N
If not, what error appeared:

Part 14 · Going further

This section is optional — work through it if you finish early or want to push deeper.

Explore the fit at each area group

▶ Try this:

# How well does the model fit within each area group? --
paper_df %>%
  mutate(
    predicted = fitted(paper_lm_model),
    residual  = residuals(paper_lm_model)
  ) %>%
  group_by(area_cm2) %>%
  summarize(
    n             = n(),
    mean_resid    = round(mean(residual), 4),
    max_abs_resid = round(max(abs(residual)), 4)
  )

✏️ Your turn: Are the residuals larger at bigger area values? What does that pattern (or lack of it) tell you about the equal-variance assumption?

Your observation:

Prediction for a real leaf tracing

Suppose you traced a leaf onto the same type of paper used in the calibration, cut it out, and weighed it. The tracing weighs 0.175 g.

▶ Try this:

# Predict area for your leaf tracing -------------------
my_tracing <- tibble(mass_g = 0.175)

predict(paper_lm_model,
        newdata  = my_tracing,
        interval = "prediction") %>%
  round(2)

✏️ Your turn: What is the predicted leaf area? What is the 95% prediction interval? How confident are you in this prediction?

Predicted area:
95% prediction interval:  _____ to _____ cm²
Confidence in prediction:

Extrapolation warning

The calibration data ranges from 1 to 567 cm². Suppose a very large leaf tracing weighs 5.2 g.

✏️ Your turn: Would you trust a prediction for a tracing mass of 5.2 g? Why or why not? (Hint: what is the largest mass in the calibration data?)

Max mass in calibration data:
Is 5.2 g within the calibration range?  Y / N
Should you predict at 5.2 g?  Y / N
Why:

📖 Whitlock & Schluter §17.2 p. 550: “Extrapolation is the prediction of Y at values of X beyond the range of X-values in the data. Extrapolation is problematic because there is no way to ensure the relationship between X and Y continues to be linear.”

Connect to Worksheet 04

In Worksheet 04 you ran a Welch’s t-test comparing leaf weight (grams) between sunny and shady sides and found a significant difference (p < 0.001). Now you have a calibration curve that converts tracing weight to area (cm²) — a more biologically meaningful measurement. You could use the regression equation as a mutate() step to add predicted area to a leaf data frame, then redo the t-test on area rather than weight.

▶ Sketch the code (do not necessarily run it — you would need actual tracing masses):

# How you WOULD apply the calibration to the leaf data
# (assuming trace_df has columns: side, tracing_mass_g)

# trace_df <- trace_df %>%
#   mutate(predicted_area_cm2 = b_slope * tracing_mass_g + a_intercept)
#
# t.test(predicted_area_cm2 ~ side, data = trace_df,
#        var.equal = FALSE, alternative = "two.sided")

✏️ Your turn: What is the advantage of working in units of area (cm²) rather than raw weight (g) when comparing sunny vs. shady leaves?

Your answer:

What your `figures/` folder should contain after this worksheet

figures/
├── paper_calibration_curve.png     ← from Part 8

Getting unstuck

lm() error: the formula must be lm(Y ~ X, data) — response on the left, predictor on the right. Check column names with names(paper_df).
coef() gives two values: coef(model)[1] is the intercept, coef(model)[2] is the slope for the first predictor.
residuals() vs raw data: always apply shapiro.test() and ggplot(aes(sample = residuals)) to residuals(model), not to the raw Y column.
predict() needs a tibble: newdata must be a data frame or tibble with the exact same column name as the predictor (mass_g). A typo here is the most common error.
geom_smooth(se = TRUE) shows confidence band, not prediction interval. For prediction intervals, use predict() manually.
Cheat sheets — https://posit.co/resources/cheatsheets/

💡 Key idea: The five-step logic of regression (plot → fit → check assumptions → interpret → predict) is the same framework you’ll use for multiple regression, ANOVA, and every other linear model in your career.

End of Worksheet 05. Next: Worksheet 06 — applying the calibration curve to measured leaf tracings and comparing sunny vs. shady leaf areas with a t-test.

Other Formats

Worksheet 05 — Linear Regression

Regression — Predicting Leaf Area

Recap from the Quarto intro (Worksheet 4.5)

Today’s Objectives

How to use this worksheet

🧩 Chunk 1 — Set up the data (after lecture Chunk 1)

Part 1 · Load libraries and data

Libraries

The calibration dataset

Part 2 · Inspect the data

Part 3 · First scatter plot

🧩 Chunk 2 — Fit & read the model (after lecture Chunk 2)

Part 4 · Fit the regression with `lm()`

Part 5 · Decode the `summary()` output line by line

Part 6 · Extract the calibration equation

Part 7 · Extract and interpret R²

🧩 Chunk 3 — Visualize & check assumptions (after lecture Chunk 3)

Part 8 · Plot the calibration curve

Save the plot

Part 9 · Check assumptions — residual plot

Part 10 · Check assumptions — QQ plot and Shapiro-Wilk

🧩 Chunk 4 — Predict & report (after lecture Chunk 4)

Part 11 · Predict leaf area from tracing mass

Method 1 — Direct calculation

Method 2 — Using `predict()` with intervals

Part 12 · Write a scientific results paragraph

Part 13 · Review and checkpoint

Part 14 · Going further

Explore the fit at each area group

Prediction for a real leaf tracing

Extrapolation warning

Connect to Worksheet 04

What your `figures/` folder should contain after this worksheet

Getting unstuck

Other Formats

Regression — Predicting Leaf Area

Recap from the Quarto intro (Worksheet 4.5)

Today’s Objectives

How to use this worksheet

🧩 Chunk 1 — Set up the data (after lecture Chunk 1)

Part 1 · Load libraries and data

Libraries

The calibration dataset

Part 2 · Inspect the data

Part 3 · First scatter plot

🧩 Chunk 2 — Fit & read the model (after lecture Chunk 2)

Part 4 · Fit the regression with lm()

Part 5 · Decode the summary() output line by line

Part 6 · Extract the calibration equation

Part 7 · Extract and interpret R²

🧩 Chunk 3 — Visualize & check assumptions (after lecture Chunk 3)

Part 8 · Plot the calibration curve

Save the plot

Part 9 · Check assumptions — residual plot

Part 10 · Check assumptions — QQ plot and Shapiro-Wilk

🧩 Chunk 4 — Predict & report (after lecture Chunk 4)

Part 11 · Predict leaf area from tracing mass

Method 1 — Direct calculation

Method 2 — Using predict() with intervals

Part 12 · Write a scientific results paragraph

Part 13 · Review and checkpoint

Part 14 · Going further

Explore the fit at each area group

Prediction for a real leaf tracing

Extrapolation warning

Connect to Worksheet 04

What your figures/ folder should contain after this worksheet

Getting unstuck

Part 4 · Fit the regression with `lm()`

Part 5 · Decode the `summary()` output line by line

Method 2 — Using `predict()` with intervals

What your `figures/` folder should contain after this worksheet