Regression & Model Comparison - IB Mathematics Applications and Interpretation

An economist models the unemployment rate over several years. The residual plot for their linear model shows that residuals are negative for the earliest years, positive for the middle years, and negative again for the most recent years. What does this pattern suggest?
Answer: The relationship between time and unemployment rate is non-linear, and a different model is needed. A systematic pattern in the residuals, like the negative-positive-negative sequence described, indicates that the chosen linear model is not capturing the true underlying relationship. This specific pattern often suggests that the data follows a curve (like a parabola) that the straight line cannot properly fit. Therefore, a non-linear model (e.g., quadratic) should be investigated. Randomly scattered residuals would have indicated a good fit.
A scientist is modeling the concentration of bacteria in a petri dish. The population grows quickly at first, but then slows down as resources become scarce and the population size approaches the carrying capacity of the dish. Which of the following regression models is most suitable for this situation?
Answer: Logistic model This scenario describes logistic growth. The key features are an initial phase of rapid (almost exponential) growth followed by a slowing of growth as the population approaches a maximum limit, or carrying capacity. A logistic model is specifically designed to represent this S-shaped growth curve. An exponential model (A) would show unlimited growth. A linear model (B) would show a constant rate of growth. A quadratic model (C) would show the population increasing and then decreasing (or vice versa), which doesn't fit the description of leveling off.
A student collects data on the number of pages in a textbook and its price. They fit a least squares regression line to the data. What is the theoretical sum of the residuals from this line?
Answer: It is approximately zero. When you encounter questions about least squares regression lines, remember that these lines are fitted using a specific mathematical property that minimizes the sum of squared residuals. This creates a fundamental characteristic about the residuals themselves. The least squares regression line is positioned so that it minimizes $\sum(y_i - \hat{y}_i)^2$ , where $y_i$ are the actual values and $\hat{y}_i$ are the predicted values. A key mathematical property of this optimization process is that the sum of the residuals (not squared) must equal zero: $\sum(y_i - \hat{y}_i) = 0$ . This happens because the regression line passes through the point $(\bar{x}, \bar{y})$ , ensuring that positive and negative residuals balance out perfectly. Let's examine why the other options are incorrect. Option A suggests the sum always equals 1, which has no mathematical basis in regression theory. Option B claims the sum depends on the correlation coefficient, but regardless of whether correlation is strong or weak, the sum of residuals is always zero due to the mathematical properties of least squares fitting. Option C states the sum is always positive, but this ignores that residuals can be both positive (points above the line) and negative (points below the line), and they must balance to zero. The correct answer is D - the sum is approximately zero (or exactly zero in theory, with "approximately" accounting for potential rounding in calculations). Study tip: Remember that least squares regression has two key properties: it minimizes the sum of squared residuals, and it forces the sum of regular residuals to equal zero.
The population of a town over several decades shows a period of growth, followed by a period of stability, and then a period of decline. An attempt is made to model this trend. Which type of regression would be most appropriate to capture these three distinct phases?
Answer: A cubic model The description of the population trend involves two turning points: one where growth changes to stability, and another where stability changes to decline. A cubic model, which has the general shape of $y = ax^3 + bx^2 + cx + d$ , can have up to two local extrema (turning points), making it suitable for capturing a pattern of rise, level-off, and fall. A logistic model only levels off. An exponential model only increases or decreases. A linear model shows a constant rate of change.
A researcher fits four different regression models to the same dataset to predict crop yield. To assess the goodness-of-fit for each model, the sum of squared residuals (SSR) is calculated. The results are as follows: Model A (Linear): SSR = 158.2 Model B (Quadratic): SSR = 45.7 Model C (Cubic): SSR = 47.1 Model D (Exponential): SSR = 98.6 Based solely on the SSR values, which model provides the best fit to the data?
Answer: Model B The sum of squared residuals (SSR), or $S_{res}$ , measures the total squared difference between the observed data and the values predicted by the model. A smaller SSR indicates that the model's predictions are closer to the actual data points, signifying a better fit. Among the given options, Model B (Quadratic) has the lowest SSR of 45.7, making it the best-fitting model based on this criterion.
An analyst is comparing two models for a dataset. Model A (a quadratic function) has an $R^2$ of 0.95, but its residual plot shows a clear wave-like pattern. Model B (a cubic function) has an $R^2$ of 0.92, and its residual plot shows a random scatter of points around zero. Which model should be preferred and why?
Answer: Model B, because a random residual plot indicates that the model has captured the underlying trend appropriately. When evaluating statistical models, you need to consider both goodness of fit (like $R^2$ ) and model assumptions. A key assumption is that residuals should be randomly distributed around zero, indicating the model has captured all systematic patterns in the data. Model B is the better choice because its random residual plot shows it has successfully captured the underlying relationship in the data. When residuals are randomly scattered, it means there's no remaining pattern the model failed to account for. The slightly lower $R^2$ of 0.92 is still excellent and represents genuine explanatory power. Option A is wrong because while Model A's $R^2$ of 0.95 appears better, the wave-like residual pattern reveals a critical flaw: the quadratic model is systematically missing parts of the data's structure. This pattern indicates the model is inadequate despite its high $R^2$ . Option B incorrectly suggests the models are equivalent. The residual patterns make them fundamentally different in quality - one meets key assumptions while the other violates them. Option C misunderstands the situation. Model B actually demonstrates the ideal combination: high $R^2$ with random residuals. The statement incorrectly implies neither model is acceptable. Remember: $R^2$ alone can be misleading. A model with systematic residual patterns is missing something important about the data's structure, even if it explains a high percentage of variance. Always check residual plots to verify your model assumptions are met - random scatter around zero is what you want to see.
A financial analyst models a stock's price based on the company's quarterly earnings. For a quarter where the earnings were $2.50 per share, the model had a residual of -$1.75. What is the correct interpretation of this residual?
Answer: The model over-predicted the stock's price by $1.75. A residual is calculated as (actual value) - (predicted value). A negative residual means that the predicted value was greater than the actual value. In this context, a residual of -$1.75 means $\text{Actual Price} - \text{Predicted Price} = -\$1.75$ , or $\text{Predicted Price} = \text{Actual Price} + \$1.75$ . Therefore, the model over-predicted the stock's price by $1.75. Distractor A corresponds to a positive residual. Distractors C and D mistake the residual for the actual or predicted price itself.
The researcher uses the model to predict the test score for a student who studied for 20 hours. What is the main concern with this prediction?
Answer: The prediction is an extrapolation, which may be unreliable as it is outside the domain of the original data. The model was built using data for study hours between 1 and 10. Predicting a score for 20 hours of study is an extrapolation because it falls far outside this range. Even if a model has a high $R^2$ value for the data it was built on, its behavior outside that range is unknown and potentially very inaccurate. For example, a quadratic model might predict a decreasing score after a certain number of hours, which may or may not be realistic. The primary issue is the unreliability of extrapolation. Distractor A ignores this issue. Distractor C is too absolute; a quadratic model could be appropriate within a certain domain. Distractor D is irrelevant as calculators handle the computation.
An agronomist models the height of a plant over time. The data shows that the plant's growth rate slows as it gets taller. A plot of $\ln(H)$ versus $t$ , where $H$ is height and $t$ is time, yields a straight line. What type of model does this suggest for the original relationship between $H$ and $t$ ?
Answer: Exponential model, $H = ab^t$ When you encounter a problem about transforming data to create linear relationships, you're dealing with linearization techniques that help identify the underlying mathematical model. The key insight here is working backwards from the linear transformation. If plotting $\ln(H)$ versus $t$ produces a straight line, then we can write: $\ln(H) = mt + c$ for some constants $m$ and $c$ . To find the original relationship, we need to solve for $H$ by taking the exponential of both sides: $H = e^{mt + c} = e^c \cdot e^{mt}$ . Since $e^c$ is just a constant (let's call it $a$ ), and $e^{mt} = (e^m)^t$ (let's call $e^m = b$ ), we get $H = ab^t$ . This is exactly the exponential model in option A. Let's examine why the other options fail: Option B (linear model) would give a straight line when plotting $H$ directly versus $t$ , not $\ln(H)$ versus $t$ . Option C (power model) would produce a straight line when plotting $\ln(H)$ versus $\ln(t)$ , not versus $t$ itself. Option D (logarithmic model) would create a straight line when plotting $H$ versus $\ln(t)$ . Remember this pattern: the variables in your linear plot tell you the model type. $\ln(y)$ vs. $x$ linear → exponential model. $\ln(y)$ vs. $\ln(x)$ linear → power model. $y$ vs. $\ln(x)$ linear → logarithmic model. Master these transformations to quickly identify mathematical relationships from linearized data.
Which of the following statements is the most accurate assessment of the two predictions?
Answer: Prediction A is an interpolation and is likely to be more reliable than Prediction B, which is an extrapolation. Prediction A is for $x=65$ , which is within the range of the original data (20 to 100). This is called interpolation. Prediction B is for $x=125$ , which is outside the range of the original data. This is called extrapolation. Models are generally reliable for interpolation but can be very unreliable for extrapolation, as the underlying trend may change outside the observed domain. Therefore, Prediction A is more likely to be reliable.
What is the most likely issue with using this 5th-degree polynomial model for prediction?
Answer: The model is likely overfitted to the noise in the data and will have poor predictive accuracy for new data. When you encounter regression problems involving polynomial models, the key concept to consider is the balance between model complexity and data size. This question tests your understanding of overfitting, a critical issue in statistical modeling. With only 6 data points, a 5th-degree polynomial has exactly enough parameters to pass through every single point perfectly. This creates a dangerous situation: the model is essentially memorizing the specific data rather than learning the underlying relationship between variables. While $R^2 = 1$ looks impressive, it's actually a red flag when you have so few data points relative to model complexity. Option A correctly identifies that this model is overfitted to the noise in the data and will perform poorly on new, unseen data points. The polynomial will likely create wild oscillations between the original points, making unreliable predictions. Option B falls into the common trap of thinking perfect fit equals perfect model. A model that fits training data perfectly but fails on new data is essentially useless for prediction. Option C incorrectly focuses on computational complexity, which isn't the real issue here. Modern calculators can easily handle 5th-degree polynomials. Option D shows a misunderstanding of $R^2$ . It's absolutely possible to achieve $R^2 = 1$ when you have enough parameters to fit all data points exactly. Key takeaway: Always be suspicious of complex models with perfect fit on small datasets. In IB Mathematics, look for the general rule that your model should be simpler than your data – fewer parameters than data points prevents overfitting.
A physicist is modeling the decay of a radioactive substance. Based on well-established physical laws, the decay should follow an exponential model. The physicist fits two models to the experimental data: Model A (Exponential): $R^2 = 0.98$ Model B (Quartic Polynomial): $R^2 = 0.99$ Which model is the more sensible choice for making predictions, and why?
Answer: Model A, because it is consistent with the established scientific theory for the phenomenon, and avoids potential overfitting. When evaluating mathematical models in real-world contexts, you need to balance statistical fit with theoretical soundness and predictive reliability. This question tests whether you understand that a higher $R^2$ value doesn't automatically mean a better model. Model A is the sensible choice because radioactive decay is governed by well-established physical laws that produce exponential behavior. While its $R^2 = 0.98$ is slightly lower than Model B's, it aligns with the underlying physics and is more likely to make accurate predictions beyond the observed data range. The quartic polynomial, despite its higher $R^2 = 0.99$ , represents overfitting—using excessive complexity to capture noise rather than the true signal. Looking at the wrong answers: Option B falls into the common trap of assuming higher $R^2$ always means better modeling. This ignores overfitting risks and theoretical considerations. Option C incorrectly states both models are equally good when they clearly differ in their theoretical foundation and predictive reliability. Option D misunderstands that $R^2 = 1$ would indicate perfect fit to data points, which is unrealistic with real experimental data containing measurement errors and natural variation. Key strategy: On IB Mathematics questions involving model selection, remember that the "best" model balances statistical fit with theoretical appropriateness. A model consistent with established scientific principles that fits reasonably well will typically outperform a purely statistical model with slightly better fit statistics, especially for predictions.
A researcher is modeling the concentration of a medication in the bloodstream over time. The concentration increases after administration, reaches a peak, and then decreases. Which type of regression model would be most appropriate to capture this entire process?
Answer: A quadratic model The description indicates a relationship with a single turning point (a maximum). A quadratic model, specifically a parabola opening downwards (with a negative coefficient for the squared term), is well-suited to model a quantity that rises to a peak and then falls. A linear model cannot capture a turning point. An exponential decay model only describes a decrease. A logistic model describes a value that rises and then levels off, but does not decrease.
A marketing team analyzes the number of users of a new mobile application over its first year. They fit three different models to the data, where $t$ is the number of months since launch. The coefficient of determination, $R^2$ , for each model is calculated. Linear Model: $R^2 = 0.82$ Quadratic Model: $R^2 = 0.96$ Logistic Model: $R^2 = 0.99$ Given that the user growth was rapid initially and is now slowing as it approaches market saturation, which model should the team choose and why?
Answer: The logistic model, because it has the highest $R^2$ and its shape correctly models growth that slows and approaches a limit. The best model is chosen based on both statistical measures (like $R^2$ ) and the underlying context. The logistic model not only has the highest $R^2$ value (closest to 1), indicating the best statistical fit, but its S-shaped curve is also theoretically appropriate for situations like market saturation where growth slows as it approaches a carrying capacity. The quadratic model, while having a high $R^2$ , would eventually predict a decline in users, which may not be realistic. The linear model is a poor fit compared to the others.
A power model of the form $y = ax^b$ is often used in biology. To fit such a model, data is often transformed. If a researcher plots $\log(y)$ against $\log(x)$ and observes a strong linear relationship, what does this confirm?
Answer: A power model of the form $y=ax^b$ is appropriate. To see why, take the logarithm of both sides of the power equation $y = ax^b$ . This gives $\log(y) = \log(ax^b) = \log(a) + \log(x^b) = \log(a) + b\log(x)$ . If we let $Y = \log(y)$ , $X = \log(x)$ , and $c = \log(a)$ , this equation becomes $Y = c + bX$ , which is the equation of a straight line. Therefore, a linear relationship between $\log(y)$ and $\log(x)$ confirms that a power model is appropriate for the original variables $x$ and $y$ .
An environmental scientist fits a linear regression model to data relating the concentration of a pollutant in a river (in ppm) to the distance downstream from a factory (in km). The resulting residual plot shows a distinct U-shape. What is the most appropriate conclusion?
Answer: A non-linear model, such as a quadratic function, would likely provide a better fit for the data. A pattern, such as a U-shape, in a residual plot indicates that the chosen model (in this case, linear) is not capturing the underlying structure of the data. A non-linear model is needed. A U-shaped residual plot often suggests a quadratic relationship. Distractor A is incorrect because a pattern indicates the model type is wrong, not just that the correlation is weak. Distractor C describes heteroscedasticity, which typically appears as a fan or funnel shape in the residual plot, not a U-shape. Distractor D is incorrect because the existence of a pattern, even in the residuals, implies there is a relationship.
A linear regression model is created for a set of data, and the sum of squared residuals (SSR) is calculated to be 250. A new data point is added that lies very far from the original regression line. How will the SSR of the new regression model (re-calculated with the new point) most likely compare to the original SSR?
Answer: It will be significantly larger than 250. The sum of squared residuals (SSR) is a measure of the total error of the model. An outlier, a point that lies far from the general trend of the data, will have a large residual. The square of this large residual will add a substantial amount to the SSR. Even though the regression line itself will shift slightly to accommodate the new point, the overall sum of squared errors will almost certainly increase, and likely significantly, due to the presence of the outlier.

Start this quiz