Error Sources in Modeling — I can explain sources of error in measurement/modeling and how they affect conclusions. - Math 3

SECTION 1

Historical Context & Motivation

Humans have been building mathematical models for thousands of years — from ancient astronomers predicting eclipses to modern engineers designing bridges. However, every model ever created contains some degree of error, a gap between what the model predicts and what actually happens. Understanding where these errors come from is just as important as building the model itself, because unrecognized errors can lead to dangerously wrong conclusions.

Throughout history, breakthroughs in science and engineering have often come not from building better models, but from understanding the sources of error in existing ones. Let's trace some key moments when error analysis changed the course of knowledge.

~150 CE • Ptolemy's Geocentric Model

Ptolemy's model of the solar system placed Earth at the center. It worked reasonably well for predicting planetary positions, but its systematic error — a flawed assumption about Earth's position — grew larger over centuries of observation.

1809 • Gauss and Least Squares

Carl Friedrich Gauss developed the method of least squares to minimize prediction errors when tracking the orbit of the asteroid Ceres. This became the foundation for modern regression analysis.

1920s • Measurement Uncertainty in Physics

Werner Heisenberg's uncertainty principle showed that certain measurements have fundamental limits. Scientists formalized error propagation — how small measurement errors combine and grow through calculations.

1986 • Challenger Disaster

Engineers used a model that underestimated the effect of cold temperatures on O-ring seals. This tragic failure highlighted the consequences of ignoring error sources outside a model's tested range — a concept called extrapolation error.

2020s • Modern Data Science

Today, every field from medicine to climate science relies on error analysis. Researchers routinely report confidence intervals and margins of error alongside their predictions, acknowledging that no model is perfect.

The central question this lesson addresses is: Where do errors come from in mathematical models, and how do those errors affect the conclusions we draw? By the end, you'll be able to identify, classify, and reason about errors in any modeling context.

SECTION 2

Core Principles & Definitions

Before diving into specific examples, you need a clear vocabulary for talking about errors. In mathematics and science, "error" doesn't mean "mistake" — it refers to the unavoidable difference between a model's prediction and the real-world value. Every model simplifies reality, and those simplifications introduce gaps. Here are the foundational ideas you need.

Measurement Error

The difference between a measured value and the true value. Every tool — rulers, thermometers, sensors — has limited precision. These limitations introduce uncertainty into any data you collect.

Systematic Error (Bias)

An error that consistently pushes results in one direction — always too high or always too low. A scale that reads 2 pounds heavy produces a systematic error that affects every single measurement the same way.

Random Error

Unpredictable fluctuations that cause measurements to scatter above and below the true value. Unlike systematic error, random errors tend to average out with enough data points.

Model Assumptions Error

Every model makes simplifying assumptions — for instance, assuming a relationship is linear when it's actually curved. Choosing the wrong type of model introduces structural error that no amount of data can fix.

Extrapolation Error

When you use a model to make predictions beyond the range of your original data, accuracy often drops sharply. The further you extrapolate, the less reliable the prediction becomes.

✦

KEY TAKEAWAY

Think of a model like a GPS app giving you an estimated arrival time. Measurement error is like the GPS not knowing your exact speed at every moment. Systematic error is like the app always underestimating traffic. Model assumptions error is like the app assuming all roads have the same speed limit. And extrapolation error is like trusting the app's estimate for a city it's never mapped. No single source of error is always the biggest — you have to check them all.

SECTION 3

Visual Explanation — How Errors Affect a Model

The diagram below shows a scatter plot of real data (dots) along with two different model fits. The linear model (straight line) captures the general trend but misses the curvature in the data — this is a model assumptions error. The quadratic model (curve) fits the data more closely, but notice how it diverges wildly in the extrapolation zone — illustrating extrapolation error. The vertical bars on each data point represent measurement uncertainty.

The dashed cyan line is the linear fit, and the green curve is the quadratic fit. The pink bracket marks a residual — the vertical distance between a data point and the model's predicted value. The amber zone on the right is the extrapolation zone where no data exists to validate predictions.

Notice how the linear model's residuals show a clear pattern — they are negative in the middle and positive at the ends. This pattern is a red flag that the model's structure is wrong, not just that the measurements are imprecise. When residuals show no pattern and scatter randomly, the model structure is likely appropriate and the remaining error is due to random measurement noise.

SECTION 4

Mathematical Framework — Quantifying Errors

To move beyond simply identifying errors, you need tools to measure how large they are. These formulas let you put a number on the error so you can compare models, assess reliability, and communicate how trustworthy a conclusion is.

ABSOLUTE ERROR

Absolute Error = |Measured Value − True Value|

The absolute value bars ensure the error is always positive. This tells you how far off a single measurement is, regardless of direction.

PERCENT ERROR

Percent Error = (|Measured − True| ÷ |True|) × 100%

Percent error puts the error in context. An error of 5 grams means very different things if the true value is 10 grams (50% error) versus 1000 grams (0.5% error).

RESIDUAL

Residual = y_observed − y_predicted

A residual is the difference between an actual data point and the value the model predicts at the same x-value. Positive residuals mean the model underestimates; negative residuals mean it overestimates.

SUM OF SQUARED RESIDUALS (SSR)

SSR = Σ(yᵢ − ŷᵢ)²

Squaring each residual makes all values positive and penalizes larger errors more heavily. A smaller SSR indicates a better-fitting model. The regression line of best fit is the one that minimizes the SSR.

💡

Why Square the Residuals?

If you simply added up the residuals without squaring, the positive and negative values would cancel each other out, giving you a misleadingly small total. Squaring eliminates the cancellation and also makes the math work nicely for finding the best-fit line using algebra.

SECTION 5

Classifying Error Sources — A Deeper Look

Now that you know the major categories of error, let's see how they compare in terms of their cause, their effect on your conclusions, and what you can do about them. The diagram below organizes the five main error sources into a visual flowchart, and the table that follows provides a detailed comparison.

This flowchart classifies all model error into two main branches: data/measurement error and model/structural error. Each sub-type shows its effect on conclusions and how it can be addressed. Human blunders sit outside the classification because they are preventable mistakes rather than inherent error sources.

Error Type	Cause	Effect on Conclusions	How to Reduce
Random Error	Natural variation in measurements; unpredictable fluctuations	Reduces precision; widens confidence intervals; averages out with more data	Take more measurements; average repeated trials
Systematic Error	Flawed instrument calibration; consistent procedural bias	Shifts results in one direction; creates false conclusions even with large datasets	Calibrate instruments; compare with known standards; use controls
Model Assumption Error	Choosing the wrong type of function (e.g., linear for curved data)	Systematic pattern in residuals; consistently wrong predictions	Check residual plots; try alternative model types (quadratic, exponential)
Extrapolation Error	Predicting beyond the range of collected data	Error can grow rapidly; predictions may be wildly inaccurate	Collect data across a wider range; flag extrapolated results as uncertain
Human Error (Blunder)	Mistakes in reading, recording, or entering data	Creates outliers; can distort the model fit dramatically	Double-check data entry; use automated collection when possible

Summary of error types, causes, effects, and remedies

SECTION 6

Worked Example — Identifying and Quantifying Error

A student collects data on the height of a bouncing ball versus the drop height. She uses a linear model to fit the data and wants to predict the bounce height for a drop height outside her data range. Let's walk through how to identify and quantify the errors in her work.

Ball Bounce Height — Error Analysis

1
Step 1 — Identify the Data and Model

The student drops a ball from heights of 50, 100, 150, 200, and 250 cm, measuring bounce heights. She fits a linear model: ŷ = 0.62x + 3.1, where x is drop height (cm) and ŷ is predicted bounce height (cm). The r² value is 0.97.
2
Step 2 — Calculate a Residual

At x = 200 cm, the actual bounce height was 121 cm. The model predicts ŷ = 0.62(200) + 3.1 = 127.1 cm. The residual is: 121 − 127.1 = −6.1 cm. This means the model overestimated the bounce height by 6.1 cm at this point.

Residual = −6.1 cm
3
Step 3 — Calculate Percent Error at This Point

Percent Error = (|121 − 127.1| ÷ 121) × 100% = (6.1 ÷ 121) × 100% ≈ 5.04%. This is a moderate error — the model is within about 5% at this point.

Percent Error ≈ 5.04%
4
Step 4 — Check for Systematic Pattern in Residuals

Looking at all five residuals: +1.2, +0.5, −2.3, −6.1, −8.8. The residuals start positive and become increasingly negative. This curved pattern in the residuals suggests the relationship is not truly linear — there is a model assumptions error. A quadratic or square root model might fit better.

Model assumptions error detected — patterned residuals indicate the wrong model type
5
Step 5 — Assess Extrapolation Risk

The student wants to predict bounce height at x = 500 cm (double her largest value). Using the linear model: ŷ = 0.62(500) + 3.1 = 313.1 cm. But since the residuals already show increasing negative error, extrapolating this far is likely to produce an even larger overestimate. This prediction should be flagged as highly unreliable due to both model assumptions error and extrapolation error compounding together.

Extrapolated prediction = 313.1 cm — unreliable due to compounding error sources

SECTION 7

Strengths & Limitations of Error Analysis

Understanding error sources is powerful, but error analysis itself has limitations. The table below compares what error analysis can and cannot do for you as a modeler.

Strengths	Limitations
Helps you choose between competing models by comparing residuals and SSR values	Cannot tell you if you've missed a variable entirely — an unknown factor could be driving the data
Quantifies uncertainty so you know how confident to be in a prediction	Requires enough data to be meaningful; a handful of data points can give misleading error estimates
Residual plots reveal when a model type is inappropriate	A good-looking residual plot doesn't guarantee the model is correct — it only means no obvious structural problem was detected
Helps communicate reliability to others through margins of error and confidence intervals	People often misinterpret margins of error as guarantees rather than probabilistic ranges

What error analysis can and cannot accomplish

✦

KEY TAKEAWAY

Error analysis is like a doctor's diagnostic toolkit — it can identify many problems and measure how severe they are, but it can't catch a disease the doctor has never heard of. Similarly, error analysis can flag issues with the model you have, but it can't alert you to an entirely different variable or mechanism you haven't considered. The best practice is to combine error analysis with domain knowledge — your understanding of the real-world situation being modeled.

SECTION 8

Connection to Advanced Methods

The error analysis concepts you've learned here form the foundation for more sophisticated techniques used in college statistics, data science, and scientific research. The table below previews how each concept you've learned connects to its advanced counterpart.

What You Learned	Advanced Version	Where It's Used
Residuals and SSR	Coefficient of determination (R²), adjusted R², root mean square error (RMSE)	Statistics courses, data science, machine learning model evaluation
Percent error	Confidence intervals, standard error of the estimate	Scientific research, polling, clinical trials
Residual pattern analysis	Hypothesis testing on residuals, Durbin-Watson test, heteroscedasticity checks	Econometrics, regression analysis, quality control
Systematic vs. random error	Bias-variance tradeoff in machine learning	Artificial intelligence, predictive modeling, deep learning
Extrapolation caution	Cross-validation, train/test split, out-of-sample prediction	Data science, weather forecasting, financial modeling

From high school concepts to advanced applications

You don't need to master these advanced tools right now, but recognizing that error analysis scales up is valuable. The habit of asking "Where could this model be wrong?" is the same whether you're fitting a line in Math 3 or training a neural network in a computer science lab. The logic stays the same — only the tools get more powerful.

SECTION 9

Practice Problems

Question 1: CONCEPTUAL

A thermometer consistently reads 2°F too high. A student uses it to collect temperature data for a science experiment and builds a linear model from the data. Is the error in this situation random or systematic? Will taking more measurements fix the problem? Explain your reasoning.

Show answer

This is a systematic error because the thermometer consistently reads 2°F too high — every measurement is shifted in the same direction by the same amount. Taking more measurements will not fix this problem because the bias doesn't cancel out when you average. Every reading is 2°F too high, so the average will also be 2°F too high. The student needs to calibrate the thermometer or subtract 2°F from each reading to correct for the systematic bias.

Question 2: BASIC CALCULATION

A model predicts that a car traveling at 60 mph will take 4.2 hours to complete a trip. The actual time was 4.5 hours. Calculate the absolute error and the percent error of the model's prediction.

Show answer

Absolute Error = |4.2 − 4.5| = 0.3 hours. Percent Error = (0.3 ÷ 4.5) × 100% ≈ 6.67%. The model underestimated the travel time by about 6.67%. This tells us the model's prediction was fairly close but not exact — the error might be due to factors the model didn't account for, like traffic or stops. In this context, a 6.67% error translates to about 18 minutes, which could be significant for time-sensitive planning.

Question 3: INTERMEDIATE

A student fits a linear model to the following (x, y) data: (1, 3), (2, 5), (3, 9), (4, 14), (5, 22). The linear model is ŷ = 4.4x − 2.8. Calculate the residuals for each data point. Do the residuals show a pattern? What does this suggest about the model?

Show answer

Predicted values: ŷ(1) = 1.6, ŷ(2) = 6.0, ŷ(3) = 10.4, ŷ(4) = 14.8, ŷ(5) = 19.2. Residuals: 3 − 1.6 = +1.4, 5 − 6.0 = −1.0, 9 − 10.4 = −1.4, 14 − 14.8 = −0.8, 22 − 19.2 = +2.8. The residuals follow a curved pattern — positive, then negative, then positive again (a U-shape). This pattern indicates a model assumptions error: the data likely follows an exponential or quadratic relationship, not a linear one. The linear model systematically misrepresents the shape of the data, and no amount of additional linear data will fix this structural problem.

Question 4: APPLIED

A city planner uses 10 years of population data (2010–2020) to build a linear growth model: P(t) = 2,400t + 185,000, where t is years since 2010. She uses it to predict the population in 2050 (t = 40). Identify at least two distinct sources of error in this prediction and explain how each one affects the reliability of the 2050 estimate.

Show answer

First, there is extrapolation error: the model was built on data from t = 0 to t = 10, but the prediction at t = 40 is four times beyond the data range. Population growth often follows exponential or logistic patterns rather than staying linear, so a linear extrapolation 30 years out could significantly overestimate or underestimate the real population. Second, there is model assumptions error: assuming constant linear growth ignores factors like economic changes, migration patterns, housing capacity, and natural disasters that could alter the growth rate. Together, these errors mean the 2050 prediction should be treated with very low confidence — it's a rough projection at best, not a reliable forecast.

Question 5: CRITICAL THINKING

Two students model the relationship between study hours and test scores. Student A uses a linear model with SSR = 120 and Student B uses a quadratic model with SSR = 45. Student A argues that lower SSR doesn't necessarily mean Student B's model is better. Under what circumstances could Student A be correct? Discuss at least two reasons why a lower SSR might be misleading.

Show answer

Student A could be correct for several reasons. First, overfitting: a more complex model (like a quadratic) has more parameters and will almost always fit the training data better, but that doesn't mean it captures the true relationship — it might just be fitting the noise. If the dataset is small, the quadratic could look great on the existing data but perform terribly on new data. Second, interpretability and domain knowledge: a quadratic model implies that test scores eventually decrease as study hours increase (the parabola turns downward). While this might be true due to burnout, if there's no logical reason for scores to drop, the quadratic shape could be an artifact of noise rather than a real phenomenon. In modeling, the simplest model that adequately explains the data is often preferred — a principle called parsimony. Lower SSR is one factor in model selection, but it must be weighed against model complexity and real-world plausibility.

SUMMARY

Lesson Summary

Summary

Every mathematical model contains error — the key skill is knowing where it comes from and how it affects your conclusions. Measurement error arises from the limited precision of tools and splits into random error (unpredictable scatter that averages out) and systematic error (consistent bias that does not average out). Model assumptions error occurs when you choose the wrong type of model for the data, and extrapolation error grows when you predict beyond the range of your data.

You can quantify errors using absolute error, percent error, and residuals. Examining the residual plot reveals whether errors are random (scattered) or patterned (indicating a structural problem). The sum of squared residuals (SSR) provides a single number to compare model fits, though a lower SSR alone doesn't guarantee a better model — you must also consider model complexity and real-world plausibility. Always ask: where could this model be wrong?