Evaluating Model Reasonableness — I can evaluate whether a linear model is reasonable and revise assumptions when needed. - Grade 9 Math

SECTION 1

Historical Context & Motivation

Humans have been modeling real-world patterns with mathematics for thousands of years. Ancient civilizations used simple proportional reasoning to predict floods, plan harvests, and track the movements of celestial bodies. Over time, mathematicians realized that not every relationship in the world follows a neat, straight-line pattern, and learning to recognize when a model breaks down became just as important as building the model in the first place.

The idea of model evaluation — systematically checking whether a mathematical representation actually fits reality — gained serious traction during the Scientific Revolution, when thinkers like Galileo and Kepler tested their equations against careful observations. In the modern era, evaluating model reasonableness is a core practice in data science, engineering, economics, and nearly every field that relies on quantitative reasoning.

~1600 • Galileo's Experiments

Galileo Galilei tested whether falling objects follow a constant-speed (linear) model and discovered that distance actually grows with the square of time — an early lesson in model revision.

1805 • Least Squares Method

Adrien-Marie Legendre published the method of least squares, giving scientists a formal way to measure how well a line (or curve) fits observed data.

1900s • Rise of Statistics

Karl Pearson and Ronald Fisher developed correlation coefficients and residual analysis, giving researchers precise tools to judge model fit.

2010s–Today • Data Science Era

With massive datasets now commonplace, checking whether a linear model is reasonable — or whether a more complex model is needed — is a daily task for analysts worldwide.

The central question this lesson addresses is deceptively simple: How do you know if a linear model is good enough? And when it isn't, what should you do about it? These are the skills that separate someone who can write an equation from someone who can actually solve real problems.

SECTION 2

Core Principles of Model Reasonableness

Before you can evaluate whether a linear model is reasonable, you need a clear understanding of what a linear model actually claims. A linear model asserts that one variable changes at a constant rate relative to another. That's a strong claim, and it doesn't always hold in real life. The principles below give you a framework for checking that claim against evidence.

Constant Rate of Change

A linear model assumes the slope (rate of change) stays the same across the entire domain. If the real-world rate speeds up or slows down, a straight line won't capture the pattern.

Domain Awareness

Every model applies within a certain range of inputs. A line that fits well for x = 0 to 10 might produce absurd predictions at x = 100. Always ask: Where does this model make sense?

Residual Analysis

A residual is the difference between an observed value and the value the model predicts. Examining the pattern of residuals reveals whether a line is truly a good fit.

Context Check

Does the model's prediction make sense in the real world? If your line of best fit says a person's height is negative or that a temperature drops below absolute zero, the model fails the context check.

Willingness to Revise

Good modelers don't cling to a model that doesn't work. When evidence shows the linear assumption is wrong, you revise your assumptions and try a different approach — maybe quadratic, exponential, or piecewise.

✦

KEY TAKEAWAY

Think of a linear model like a straight ruler. If you're measuring a flat table edge, the ruler is perfect. But if you try to trace the curve of a basketball with a straight ruler, you'll see gaps everywhere. Evaluating model reasonableness is about noticing those gaps before you make decisions based on bad measurements.

SECTION 3

Visualizing Good Fit vs. Poor Fit

The most powerful way to evaluate a linear model is to look at it. The scatter plot below shows two datasets side by side. On the left, the data points cluster tightly around the line — the linear model is reasonable. On the right, the data clearly curves away from the line — the linear model is not reasonable, and a different model (like a quadratic curve) would be a better fit.

Left: data points follow a straight-line trend — the linear model is reasonable. Right: data points curve, and the straight line misses the pattern — the linear model should be revised or replaced.

Notice how in the right-hand plot, the data points are below the line at first, then above it in the middle, and then below it again at the far right. This systematic pattern in the residuals (the vertical gaps between the points and the line) is a telltale sign that a linear model isn't capturing the true relationship. When residuals show a random scatter (no pattern), you can be more confident your linear model is appropriate.

SECTION 4

Mathematical Framework for Evaluation

While visual inspection is a great first step, math gives us precise tools to quantify how well a linear model fits. Here are the key formulas you'll use to evaluate reasonableness.

LINEAR MODEL

y = mx + b

Where m is the slope (constant rate of change), b is the y-intercept, x is the independent variable, and y is the predicted value.

RESIDUAL

Residual = y_observed − y_predicted

A positive residual means the actual value is above the line; a negative residual means it's below the line. If the model is good, residuals should be small, roughly balanced between positive and negative, and show no clear pattern.

PERCENT ERROR

Percent Error = |y_observed − y_predicted| ÷ y_observed × 100%

Percent error lets you judge how large a residual is relative to the actual value. A residual of 5 is small if y_observed is 500, but enormous if y_observed is 10.

You may also encounter the correlation coefficient (often written as r), which measures the strength and direction of a linear relationship. Values of r close to 1 or −1 suggest a strong linear pattern, while values near 0 suggest little or no linear association. However, even a high r doesn't guarantee the model is reasonable — you still need to check residuals and context.

⚠

Watch Out

A common mistake is trusting a linear model just because the correlation coefficient is high. A dataset can have r ≈ 0.95 and still curve in a way that a straight line completely misses. Always pair numerical checks with a visual inspection of the residual plot.

SECTION 5

Reading Residual Plots

A residual plot graphs each residual (y-axis) against the corresponding x-value (x-axis). The horizontal line at zero represents the model's predictions. Points scattered randomly above and below that line indicate a good fit; a clear curved pattern indicates the model's assumptions need revision.

Left: residuals scatter randomly around zero — the linear model is reasonable. Right: residuals form a U-shape — the linear model should be revised.

Here's a quick guide to reading residual patterns. If the residuals form a U-shape or an arc, the true relationship is likely curved, and a quadratic or other nonlinear model might work better. If the residuals fan out (getting wider as x increases), the variability in the data is not constant, which can also undermine the model's predictions. And if one or two residuals are dramatically larger than the rest, those outliers deserve a closer look — they might indicate data errors or special circumstances.

Residual Pattern	What It Means	Action to Take
Random scatter	Linear model is a reasonable fit	Keep the linear model
U-shape or arc	True relationship is curved	Try a quadratic or other curved model
Fan shape (spreading out)	Variability isn't constant	Consider data transformation or weighted model
One or two extreme values	Possible outliers or errors	Investigate those data points individually

Common residual patterns and how to respond

SECTION 6

Worked Example: Cell Phone Data Plan

Suppose you're analyzing data about how much data (in GB) your friend uses on their phone each month, and you've been asked to predict future usage. The first six months of data look like this:

Month (x)	Data Used in GB (y)
1	2.0
2	2.8
3	3.5
4	5.0
5	7.2
6	10.5

Monthly cell phone data usage

A classmate suggests using the linear model y = 1.7x + 0.1. Let's evaluate whether this model is reasonable.

Evaluating the Linear Model y = 1.7x + 0.1

1
Step 1 — Calculate Predicted Values

Plug each x-value into the equation y = 1.7x + 0.1 to find the predicted values. For month 1: y = 1.7(1) + 0.1 = 1.8. For month 2: y = 1.7(2) + 0.1 = 3.5. Continue for all six months.

Predicted values: 1.8, 3.5, 5.2, 6.9, 8.6, 10.3
2
Step 2 — Calculate Residuals

Subtract each predicted value from the observed value. Month 1: 2.0 − 1.8 = 0.2. Month 2: 2.8 − 3.5 = −0.7. Month 3: 3.5 − 5.2 = −1.7. Month 4: 5.0 − 6.9 = −1.9. Month 5: 7.2 − 8.6 = −1.4. Month 6: 10.5 − 10.3 = 0.2.

Residuals: +0.2, −0.7, −1.7, −1.9, −1.4, +0.2
3
Step 3 — Look for a Pattern in the Residuals

The residuals start positive, dip negative in the middle, and then return toward positive at the end. This is the classic U-shape pattern we discussed. The model overestimates in the middle months and underestimates at the extremes.
4
Step 4 — Context Check

Data usage that grows slowly at first and then accelerates is common behavior — people discover new apps, start streaming more, etc. This kind of growth often follows an exponential or quadratic pattern rather than a linear one.
5
Step 5 — Revise the Assumption

Because the residuals show a curved pattern and the context supports accelerating growth, the linear model is not reasonable. We should revise our assumption of constant growth and try a model that allows the rate of change to increase — for example, y = 0.5x² + 0.3x + 1.2 or an exponential model.

Conclusion: The linear model is unreasonable. A curved model better fits this data.

SECTION 7

When Linear Models Shine and When They Don't

Linear models are among the most powerful and widely used tools in mathematics precisely because they are simple. But simplicity is both their greatest strength and their greatest limitation. Understanding when to trust a linear model — and when to move on — is a skill you'll use throughout high school and beyond.

Strengths	Limitations
Easy to calculate and interpret — slope and intercept have clear meanings	Cannot capture curves, acceleration, or exponential growth
Works well when the rate of change is approximately constant	May produce absurd predictions outside the data range (extrapolation risk)
Requires minimal data — you only need two points to define a line	Sensitive to outliers; one extreme value can distort the line
A strong foundation for more complex models (many start with linearity)	Assumes a relationship exists; a line through unrelated data is meaningless

Strengths vs. limitations of linear models

✦

KEY TAKEAWAY

A linear model is like cruise control on a highway. It's perfect when the road is straight and flat, but the moment you hit hills or sharp curves, you need to switch to manual control (a different model) or you'll miss the turns. Knowing when to switch is what makes you a good driver — and a good mathematician.

SECTION 8

Connecting to Nonlinear and Advanced Models

When you determine that a linear model isn't reasonable, the next step is choosing a better model. In later math and science courses, you'll learn about several alternatives, each suited to different real-world patterns. The table below compares linear models to a few common nonlinear alternatives you may encounter.

Feature	Linear (y = mx + b)	Quadratic (y = ax² + bx + c)	Exponential (y = a · bˣ)
Shape of graph	Straight line	Parabola (U-shape)	Rapid curve upward or downward
Rate of change	Constant	Changes at a steady pace	Multiplies by a constant factor
Good for	Constant-speed motion, fixed-rate costs	Projectile motion, area problems	Population growth, compound interest
Residual clue when misapplied	Random (if correct model)	U-shape in residuals	Residuals grow rapidly at high x

Comparison of common model types

In Grade 9, you aren't expected to master exponential or quadratic regression yet. But you are expected to recognize when a linear model fails and to suggest that a different type of model might work better. This critical thinking skill connects directly to Algebra 2, Statistics, and AP courses you'll take later.

ℹ

Looking Ahead

In statistics courses (like AP Statistics), you'll learn to use r² (the coefficient of determination) to quantify exactly how much of the variation in your data a linear model explains. You'll also learn formal hypothesis tests for whether the relationship is truly linear. For now, focus on building strong intuition through residual analysis and context checks.

SECTION 9

Practice Problems

Question 1: CONCEPTUAL

A student plots residuals from a linear model and sees them scattered randomly above and below zero with no visible pattern. What should the student conclude about the linear model, and why?

Show answer

The student should conclude that the linear model is a reasonable fit for the data. When residuals are randomly scattered with no pattern, it means the line captures the overall trend without systematically over- or under-predicting. There's no evidence of curvature or other structure the model is missing. However, the student should still perform a context check to make sure the predictions make physical or practical sense within the domain.

Question 2: BASIC CALCULATION

A linear model predicts y = 3x + 5. For x = 4, the observed value is y = 19. Calculate the residual and the percent error.

Show answer

First, find the predicted value: y = 3(4) + 5 = 17. Then calculate the residual: 19 − 17 = 2. The model underestimates the actual value by 2. For percent error: |19 − 17| ÷ 19 × 100% = 2 ÷ 19 × 100% ≈ 10.5%. This is a moderate error — whether it's acceptable depends on the context. For something like estimating room temperatures, 10.5% might be fine; for dosing medicine, it would not be.

Question 3: INTERMEDIATE

A company models its monthly revenue (in thousands of dollars) with y = 5x + 10, where x is the number of months since opening. The observed revenues for months 1 through 5 are: 14, 20, 27, 38, 52. Calculate all five residuals and determine whether the linear model is reasonable.

Show answer

Predicted values: y(1) = 15, y(2) = 20, y(3) = 25, y(4) = 30, y(5) = 35. Residuals: 14 − 15 = −1; 20 − 20 = 0; 27 − 25 = +2; 38 − 30 = +8; 52 − 35 = +17. The residuals start near zero but become increasingly positive — they're growing, not scattered randomly. This pattern suggests the revenue is accelerating (growing faster over time), which a linear model cannot capture. The linear model is not reasonable for this data; an exponential or quadratic model should be considered instead.

Question 4: APPLIED

A city uses a linear model to predict daily water usage (in millions of gallons) based on the high temperature (°F): y = 0.15x − 5. On a 110°F day, the model predicts 11.5 million gallons. On a 20°F winter day, what does the model predict? Is the model reasonable across this full range? Explain what you would revise.

Show answer

For x = 20: y = 0.15(20) − 5 = 3 − 5 = −2 million gallons. A negative water usage is physically impossible, which immediately tells us the model fails the context check at low temperatures. The model might be reasonable within a limited domain (perhaps 60°F to 110°F, where the data was collected), but it should not be used to extrapolate to cold temperatures. A revised model might include a minimum usage floor (people still need water in winter) or use a piecewise model with different equations for different temperature ranges.

Question 5: CRITICAL THINKING

Two students fit different models to the same dataset about plant growth over 8 weeks. Student A uses a linear model (y = 2.5x + 1) and reports r = 0.94. Student B uses a quadratic model (y = 0.4x² + 0.3x + 1.5) and notes that the residual plot for the linear model shows a clear U-shape. Which student's conclusion is more defensible, and what does this situation teach you about relying solely on r-values?

Show answer

Student B's conclusion is more defensible. Even though Student A's correlation coefficient (r = 0.94) is high, it only measures the strength of the linear association — it does not reveal whether the data follows a curve. The U-shaped residual pattern that Student B identified is direct evidence that the relationship is nonlinear. This situation teaches a critical lesson: a high r-value can be misleading if you don't also examine the residual plot. A model can explain most of the variation in the data (high r) while still being the wrong type of model. Always combine numerical measures with visual residual analysis and context checking to make a well-supported conclusion.

SUMMARY

Lesson Summary

Summary

Evaluating model reasonableness means checking whether a linear model truly captures the pattern in your data. You do this by computing residuals (observed minus predicted values), plotting them, and looking for patterns. Random scatter in the residuals means the linear model is a reasonable fit. A curved pattern in the residuals (such as a U-shape or fan shape) signals that the relationship is nonlinear and the model needs revision.

Beyond residual analysis, always perform a context check: does the model produce predictions that make sense in the real world? Watch out for negative values where negatives are impossible, or predictions that explode when you extrapolate far beyond the data. Remember that a high correlation coefficient alone does not guarantee reasonableness — always pair it with visual and contextual analysis. When evidence shows the linear assumption fails, revise your assumptions and consider quadratic, exponential, or piecewise models.