NONLINEAR MODELING: SOMETHING FISHY

Douglas Whitaker, Mount Saint Vincent University

Published: April 2021

Overview of Lesson
In this lesson students explore nonlinear regression models to explain fish weight using fish length,
using both transformation of the response variable and polynomial regression. Geometric
interpretations of variables are leveraged to suggest nonlinear models to fit. The intention of this lesson
is for students to perform two or three linear regression analyses that feel like others that they have
done before: the difference is that they draw on prior knowledge of geometric/physical relationships to
suggest a modification to the first analysis to improve it. Because most of the nonlinear models
considered in this lesson have only a single predictor variable, students’ familiarity with simple linear
regression can be extended to nonlinear modeling. If students are familiar with multiple linear
regression, then two additional polynomial regression models can be included.

Type of Data
 Two quantitative variables
 Static dataset provided by lesson plan authors

Learning Objectives
 Students will propose and evaluate nonlinear models (transformation of response variable
and/or polynomial regression)

Audience
 Students in a course that has a unit/chapter on linear regression in a statistical context, likely
students in grades 9-14.
 Prerequisites: Prior to this lesson, students should have experience with graphing polynomials,
radicals (square roots and cube roots), simple linear regression, and assessing linear regression
fit (at least heuristically).

Time Required
60 to 90 minutes, depending on the how familiar students are with regression modeling and the
software you use.

Technology and Other Materials

 Technology: Data analysis software appropriate for conducting simple linear regression. To
accommodate a variety of software tools and modeling approaches, several versions of the data
files are included. These are described later. If no technology for analysis is available, sample
computer output is included that can instead be provided to students.
 “Low-tech” materials required: none, though depending on the students, it may be beneficial to
have base-10 blocks or other cubes available to remind them of relationships among length,
area, and volume, and to motivate the relationship between volume and weight.

1 of 20
 Lesson Plan
The goal of this lesson is to develop an appropriate model for fish weight using fish length as a
predictor. This lesson is appropriate to use after students are familiar with analyzing bivariate datasets
using simple linear regression and heuristically assessing model fit. Students will begin by analyzing a
dataset using simple linear regression, assessing its fit, and determining that a model of the form 𝑦 =
𝑏 + 𝑏 𝑥 is NOT appropriate. Then, rather than stopping, a short review of geometric concepts is
conducted. The idea of this review is for students to recognize two things: 1) a fish’s weight is closely
related to the fish’s volume and 2) the units of volume are cubic millimeters. Together, this suggests
we should explain fish weight with fish volume – a variable we don’t have – but re-expressing our
explanatory variable so that it has the units of volume might be a productive path forward.

Note: Throughout the lesson, the term weight is used for what is more properly described as mass. The
original dataset uses the term weight but measures the fish using metric units for mass. The distinction
is irrelevant to the lesson, but it would be reasonable to make the changes throughout the lesson if the
distinction matters to you.

Description of Dataset

Scientists are interested in monitoring the health of trout perch in the

Oil Sands Region of Canada.1,2 As part of a larger study, fish were
collected from the Athabasca River and Peace River, and several
characteristics were measured, including the Weight of each fish (in
grams) and the Length of each fish (in millimeters). The dataset used in
this analysis includes data from 2088 trout perch. The first six rows of
this dataset are shown to illustrate what the file looks like.

Initial Analysis (Simple Linear Regression)

Ask the students to examine the relationship between fish Weight and
fish Length using simple linear regression and to assess the
appropriateness of the model. This can be done using any tools/technology that students have used in
their previous experiences with regression.

Students work in groups to analyze the dataset using simple linear regression and assess the model fit.
Students record their decisions, results, and conclusions in the Simple Linear Regression section of the
student handout. The questions they are asked to answer are:

 What regression model will you fit to explain Weight using Length as a predictor?
 What is the regression equation3?

1
Environment and Climate Change Data Canada [Link]
contaminants-oil-sands-region/wild-fish-health-oil-sands-region/
Data License: Open Government Licence - Canada
2
Trout perch illustration by Ellen Edmonson and Hugh Chrisp - [Link] Public
Domain, [Link]
3
Feel free to substitute the term prediction equation or fitted model depending on what has been used in class.

Statistics Teacher/STatistics Education Web: Online Journal of K-12 Statistics Lesson Plans 2
[Link] or [Link]
Contact Author for permission to use materials from this lesson in a publication
  What is the predicted (fitted) value for the fish with ID 4394? (See data excerpt on pg. 1)
 What is the residual for the fish with ID 4394?
 Interpret the slope.
 Interpret the 𝑅 value.
 Examine the residual plots and comment on any potential problems.

Depending on the type of statistics course this lesson is being used in, students may have informal or
formal ways for assessing model fit. At a basic level, a scatterplot exhibits a clear curved (nonlinear)
relationship between length and weight, as seen in this graph:

Note that the weights of the shortest and longest fish tend to be underestimated by the regression
model (because the regression line is below the points for the smallest and largest values of Length);
the weights of average length fish tend to be overestimated by the model, though this is far less clear in
this graph because of how many observations are in the dataset. 4

This can more clearly be seen in a Residuals vs. Fits graph.5 This is a type of residual plot (sometimes
called a residual diagnostic plot).6 While many software packages include this as a standard regression
4
See the Reflections and Additional Recommendations section at the end of the lesson plan for an extension to the lesson
that focuses on the underestimation/overestimation for certain fish lengths.
5
For simple linear regression, a scatterplot of the data with a fitted regression line and a Residuals vs. Fits graph can reveal
essentially the same insights. The real advantage of residual plots is when more sophisticated models are used that cannot
be easily visualized directly. Depending on what software you and your students are using, fitting models E and F below
may result in difficulties visualizing the model fit. For these models, a Residuals vs. Fits graph is powerful tool for gaining
insight about the model fit. Sometimes these graphs are referred to by similar names, such as Versus Fits (the term used in
the Minitab output).
6
See note at the end of the lesson about other residual plots.

Statistics Teacher/STatistics Education Web: Online Journal of K-12 Statistics Lesson Plans 3
[Link] or [Link]
Contact Author for permission to use materials from this lesson in a publication
plot, they can be constructed manually. To construct a Residuals vs. Fits graph, the fitted (predicted)
value and residual must be calculated for each observation – a tedious task best-suited for computers.
Then, a scatterplot is made with the residuals on the Y-axis and the fitted values on the X-axis; a
horizontal line at 𝑦 = 0 is usually added.

The ideal Residuals vs. Fits graph should have a horizontal scattering of points with no discernible
pattern, with points approximately equally above and below the horizontal line throughout the entire
range of the fitted values; a Residuals vs. Fits graph illustrating an ideal relationship using simulated
data (randomly generated so that it has a linear relationship) is shown below. The Residuals vs. Fits
graph for the original simple linear regression model differs from the ideal in two ways. First, there is a
clear curved pattern, indicating the same nonlinearity observed in the scatterplot. Second, the spread of
the residuals for low fitted values is smaller than the spread of the residuals for large fitted values – an
issue known as non-constant variance.

Two Residuals vs. Fits graphs: on the left, an example of one for data with an ideal linear relationship
(simulated data); on the right, the graph created using the residuals for Model 0.

Students can share their results and come to consensus about the appropriateness of the model in a
manner that is typical for your class. Once students have analyzed the data using simple linear
regression and determined that a linear model is likely NOT the most appropriate model to use, the
lesson continues to the next stage.

Activating Prior Knowledge and Planning

The purpose of this stage is to activate students’ prior knowledge and determine a path forward in the
analysis. Students continue in their same groupings to answer the questions in the Background section
of the student handout. These questions ask them to sketch the shapes of simple polynomial equations
(𝑦 = 𝑥, 𝑦 = 𝑥 , and 𝑦 = 𝑥 ) and think of a rectangular prism as a simple approximation for the shape
of a fish (to motivate the polynomial relationships used later in the modeling).

There are two distinct paths forward for students that are both reasonable: transforming the response
variable or including polynomial terms as predictor variables to address the curvature (polynomial
regression). There are subtle differences in the approaches, but both can result in an appropriate model.
In the following pages, transforming the response variable will be discussed first followed by
polynomial regression. Depending on the amount of time available, you may wish to guide the students
through only one of these possibilities – transformation of the response variable is the clear candidate
for this as it may seem simpler and more accessible to students.

Statistics Teacher/STatistics Education Web: Online Journal of K-12 Statistics Lesson Plans 4
[Link] or [Link]
Contact Author for permission to use materials from this lesson in a publication
When looking at the shape of the scatterplot, it is reasonable for students to think the relationship is
described by 𝑦 = 𝑥 as it resembles half of a parabola. It is also reasonable for students to think that
𝑦 = 𝑥 is appropriate because it resembles part of the graph of a cubic function. These forms might be
more familiar to students and suggest polynomial regression. However, students may also recognize
that 𝑦 = 𝑥 or 𝑦 = 𝑥 might similarly be reasonable ways of expressing the relationship they see –
these forms suggest a transformation of the response variable. Students should be encouraged to pursue
either a 2nd- or 3rd-degree polynomial model or a transformation using square roots or cube roots. If
there isn’t sufficient variety in the models that students select on their own, consider assigning certain
models to groups of students. Ensuring that diverse models are covered (e.g., Models A, B, C, and D
shown below) will lead to a richer discussion at the final stage of the lesson when students compare
their results with classmates.

Students should record the model they plan to fit in the New Model section of their student handout. In
summary, the following models are all reasonable depending on the patterns that students identify.

Label Model Description

Simple Linear Regression Model
A 𝑊𝑒𝚤𝑔ℎ𝑡 = 𝑏 + 𝑏 (𝐿𝑒𝑛𝑔𝑡ℎ) with Square Root Transformation
Simple Linear Regression Model
B 𝑊𝑒𝚤𝑔ℎ𝑡 = 𝑏 + 𝑏 (𝐿𝑒𝑛𝑔𝑡ℎ) with Cube Root Transformation
Quadratic Regression Model,
without lower-order terms
or
C 𝑊𝑒𝚤𝑔ℎ𝑡 = 𝑏 + 𝑏 (𝐿𝑒𝑛𝑔𝑡ℎ)
Simple Linear Regression Model
with Square Transformation for
the Predictor
Cubic Regression Model, without
lower-order terms
or
D 𝑊𝑒𝚤𝑔ℎ𝑡 = 𝑏 + 𝑏 (𝐿𝑒𝑛𝑔𝑡ℎ)
Simple Linear Regression Model
with Cube Transformation for the
Predictor
E 𝑊𝑒𝚤𝑔ℎ𝑡 = 𝑏 + 𝑏 (𝐿𝑒𝑛𝑔𝑡ℎ) + 𝑏 (𝐿𝑒𝑛𝑔𝑡ℎ) Quadratic Regression Model
𝑊𝑒𝚤𝑔ℎ𝑡 = 𝑏 + 𝑏 (𝐿𝑒𝑛𝑔𝑡ℎ) + 𝑏 (𝐿𝑒𝑛𝑔𝑡ℎ)
F Cubic Regression Model
+ 𝑏 (𝐿𝑒𝑛𝑔𝑡ℎ)

For the purposes of this lesson, fitting models of the form 𝑊𝑒𝚤𝑔ℎ𝑡 = 𝑏 + 𝑏 (𝐿𝑒𝑛𝑔𝑡ℎ) is reasonable.
However, in polynomial regression, it is typical to include all lower-order terms in a model. That is, if
a 3rd-degree term is in a model, it should also include the 2 nd-order and 1st-order terms, so: 𝑊𝑒𝚤𝑔ℎ𝑡 =
𝑏 + 𝑏 (𝐿𝑒𝑛𝑔𝑡ℎ) + 𝑏 (𝐿𝑒𝑛𝑔𝑡ℎ) + 𝑏 (𝐿𝑒𝑛𝑔𝑡ℎ) . There are different schools of thought around this
practice, but nuanced interpretations of the coefficients are easier when the lower-order terms are
included; when they are not included, the interpretation of the coefficient of the highest-order term also
must account for the lower-order components. If the course you are teaching does not cover multiple
linear regression – or does not cover multiple linear regression until much later – this likely does not
need to be mentioned.

Statistics Teacher/STatistics Education Web: Online Journal of K-12 Statistics Lesson Plans 5
[Link] or [Link]
Contact Author for permission to use materials from this lesson in a publication
Any polynomials of higher degree than this are probably too complicated to consider in this analysis.
There is some evidence that a 4th-degree model would not be unreasonable, but the value added to this
model by the additional complexity is likely not practically useful. A 3rd-degree model works well and
is more interpretable than a 4th-degree model. Similarly, a quartic root transformation is likely
unnecessarily complex given the real-world relationship motivating this dataset. If, say, one group
wants to pursue something of this complexity, that should be fine – the model can then be compared in
the next phase of the lesson and will likely not be preferred to one of the other models.

Note: if the software you are using does not support multiple linear regression or polynomial
regression, then models E and F cannot be fit. Instead, students pursuing an approach consistent with
polynomial regression (i.e., focused on making adjustments to the predictors) should be encouraged to
fit models C or D, respectively. If using TI-series calculators (or a similar technology), note that the
order of the fish observations in the datasets has been randomized: it would be reasonable to have
students manually enter the first, say, 10, 20, or 30 observations into lists and then use calculator’s
functions to make appropriate transformations and perform the regressions. Models A, B, C, and D can
all be fit in this way. (While 10 observations may minimally work, a curved pattern is apparent when
juxtaposed with a fitted regression line using the first 30 observations.)

Fitting and Evaluating Nonlinear Models

Students now put their plan (developed in the second stage) into action. The logistics of how students
enact their plan depends on the software tool used and the technical abilities of the students. A few
possibilities are described below. During this stage, you will support the students as they use software
to model and as they answer the questions in the New Model section of the student handout. Depending
on the software available, your students’ abilities, and your pedagogical goals, you may consider
providing students with copies of computer output that has already been created (included as a
supplement to this lesson).

For students comfortable with both linear regression and a full-featured statistical software package
(e.g. Minitab, SPSS, R), students may be expected to fit their model directly from the raw data using
the regression modeling and/or transformation tools built-in to the software. For students with less
experience with fitting models, you may wish to suggest these tasks to them (presented in the list
below in the approximate order they might be completed):
 Create new columns (new variables) corresponding to the transformed variables they plan to
analyze (e.g., a column for 𝑊𝑒𝑖𝑔ℎ𝑡 or 𝐿𝑒𝑛𝑔𝑡ℎ ).
 Make a scatterplot to explore bivariate relationships with these newly created variables.
 Perform a linear regression using these newly created variables.
 To make a residual plot:
o Create a column of predicted fish lengths based on the regression equation for the new
model.
o Create a column of residuals by subtracting the predicted fish lengths from the actual
fish lengths.
o Make a scatterplot with residuals on the y-axis and predicted values on the x-axis.

If students have less experience with statistical software and/or linear modeling, you may wish to
provide them with a dataset that already includes the appropriate calculated variables (included with

Statistics Teacher/STatistics Education Web: Online Journal of K-12 Statistics Lesson Plans 6
[Link] or [Link]
Contact Author for permission to use materials from this lesson in a publication
this lesson). If providing students with these pre-calculated variables, this file should not be shared
with them until after they have determined which model they want to fit. Otherwise, seeing the
calculated variables included in the dataset may suggest to them a path forward – the path forward
should be suggested by the patterns they observe rather than the ease of fitting the model.

Once students have fit their new regression model, they answer questions about the new analysis in the
New Model section of the student handout. The questions are largely the same as the ones they were
asked to answer about the original simple linear regression model.

Note: Depending on which model students chose to fit, interpreting the value of the fitted slope might
be more difficult or impossible; depending on your pedagogical goals, you may wish to omit the
question asking students to interpret the slope of their alternative model (perhaps just sharing correct
interpretations with them instead to illustrate the complexity).

 Propose a new model to explain Weight using Length as a predictor. Why did you choose this?
o This was answered in the previous stage of the lesson.
 What is the regression equation?
 What is the predicted (fitted) value for the fish with ID 4394? (See data excerpt on pg. 1)
 What is the residual for the fish with ID 4394?
 Interpret the slope.
 Interpret the 𝑅 value.
 Examine the residual plots and comment on any potential problems.
 Overall, does this model appear to be a good fit for the data?

Then students compare the new model to the simple linear regression model in the Model Comparison
section of the student handout.

 Comparison of their fitted model with the original model:

o Which model has a higher 𝑅 value? What does this mean?
o For which model is the fitted slope easier to interpret?
o Compare the residual plots for the two models and comment on the similarities and
differences.

In statistical practice, model fitting is often an iterative process where exploratory data analysis and
theoretical knowledge motivate models. These models are fit, examined, and compared, which in turn
motivates new models that may be considered. In this lesson, students engage in a limited facsimile of
model building: an initial model is considered, an alternative model is proposed based on an
examination of the initial model fit and theoretical considerations (i.e., the relationship between length,
volume, and weight), and that new model is fit and examined. As a class, several models will be
considered, but it would be reasonable to expect a single analyst to generate all of these models – and
more – as part of the model fitting process.

Group Discussion to Compare Models

After students have finished fitting their model, evaluating it, and comparing it to the original model,
you should facilitate a group discussion where each group shares its model and their findings. For each
of the models that students might fit, the 𝑅 value will be higher than the original model. This suggests

Statistics Teacher/STatistics Education Web: Online Journal of K-12 Statistics Lesson Plans 7
[Link] or [Link]
Contact Author for permission to use materials from this lesson in a publication
that pursuing a nonlinear modeling approach improves the model fit. However, improvement in 𝑅
alone is not enough to adjudicate which model is best. 7 Students should also look to the residual plots
for improvement in the curvature (and non-constant variance if this was noted earlier). Depending on
which model students fit, they may see an improvement in the residual plots but still note that
problems exist. (Sometimes this is the best we can hope for!)

Students should then discuss which model (or models) among all the fitted models is best. This section
of the lesson plan offers a brief overview of the model comparisons. The following sections provide
the fitted values and residuals, the interpretation of the slopes, and evaluation of the models in much
more detail.

Table with the regression equation and 𝑅 value for all models considered.
Label Description Regression Equation 𝑅 Residuals vs. Fits Plot
Simple Linear Clear curvature,
Weight = - 8.330
0 Regression (original .9231 potential non-constant
+ 0.1766 Length
model) variance
Simple Linear
Mild potential
Regression Model Weight^(1/2) = - 0.9896
A .9601 curvature, no apparent
with Square Root + 0.04222 Length
non-constant variance
Transformation
Simple Linear
No apparent
Regression Model Weight^(1/3) = 0.007328
B .9637 curvature, no apparent
with Cube Root + 0.02221 Length
non-constant variance
Transformation
Clear curvature
Quadratic Regression (though improved
Weight = - 2.151
C Model, without lower- .9506 over the original
+ 0.001239 Length^2
order terms mode), potential non-
constant variance
Cubic Regression No apparent
Weight = - 0.02667
D Model, without lower- .9626 curvature, potential
+ 0.000011 Length^3
order terms non-constant variance
Weight = 5.766 - 0.22211 No apparent
Quadratic Regression
E Length + 0.002768 .9628 curvature, potential
Model
Length^2 non-constant variance
Weight = 0.80 - 0.0133 No apparent
Cubic Regression
F Length - 0.000108 Length^2 .9630 curvature, potential
Model
+ 0.000013 Length^3 non-constant variance
Note that all models exhibit non-normality of the residuals if one examines the Normal Probability
Plot, though this is not central to the lesson.

The cube root transformation model (model B) is likely the best model of the ones considered. Not
only does it have the highest 𝑅 value, it more importantly addresses the curvature in the residual plot

7
As more predictors are added to a linear regression model, the 𝑅 value will always increase or stay the same – never
decrease. This is because 𝑅 = 1 − : 𝑆𝑆𝑇𝑜𝑡𝑎𝑙 is the same for all models that use the same set of 𝑌 values, and
𝑆𝑆𝐸𝑟𝑟𝑜𝑟 can only become smaller (or stay the same) as additional predictors are included in a model.

Statistics Teacher/STatistics Education Web: Online Journal of K-12 Statistics Lesson Plans 8
[Link] or [Link]
Contact Author for permission to use materials from this lesson in a publication
noted in the original model, the non-constant variance noted in the original model, and has a relatively
straightforward interpretation. Other comparisons of similar pairs of models are:

 Between models A and B, model B would be preferred because of its improvement in the
residual plots.

 Model C is likely not defensible relative to any of the others: while it represents and
improvement over the original model, there are still clear problems in the residual plots.
Model D would be preferred to C because of the improvement in the residual plots.

 Models E and F are quite similar, and – based on the model comparisons examined in this
lesson – neither is definitively better than the other. Still, if a single model must be chosen,
model F might be preferred because it is consistent with the real-world relationship that weight
is closely related to volume, which is three-dimensional property.

Students should summarize the group discussion in the Overall Comparison section of the student
handout.

 Overall comparison:
o Out of all of the models, which model do you prefer?
o Why?

Fitted values and residuals

The student handout asks students to make a prediction and calculate a residual only for fish ID 4394
(the first row in the included image of the 6 rows of the dataset). You could modify this value easily to
be any of the other IDs included in the image of the first 6 rows (or even a different value from the
dataset not pictured). With how many possible answers there are for this (considering the models
students may fit), this may not be expedient for grading. To facilitate grading, the table below gives the
fitted value and residual for the first six observations for the original model and models A-F. (These
values were obtained using a computer: depending on how and when students round, the values they
obtain may be slightly different.)

Table showing the fitted value and residual for the first six observations in the dataset for each model
considered in this lesson.

FishID 4394 3302 4780 13184 6 13321

Length 72 54 61 58 73 93
Weight 4.349 1.916 2.592 2.115 4.118 9.469
Model 0 Fitted Value 4.384 1.205 2.441 1.912 4.560 8.092
(Original) Residual -0.035 0.711 0.151 0.203 -0.442 1.377
Fitted Value 2.050 1.290 1.586 1.459 2.092 2.937
Model A
Residual 0.035 0.094 0.024 -0.005 -0.063 0.140
Fitted Value 1.606 1.207 1.362 1.295 1.629 2.073
Model B
Residual 0.026 0.035 0.012 -0.012 -0.026 0.043

Statistics Teacher/STatistics Education Web: Online Journal of K-12 Statistics Lesson Plans 9
[Link] or [Link]
Contact Author for permission to use materials from this lesson in a publication
 Fitted Value 4.272 1.462 2.460 2.017 4.452 8.566
Model C
Residual 0.077 0.454 0.132 0.098 -0.334 0.903
Fitted Value 4.155 1.738 2.517 2.159 4.332 8.986
Model D
Residual 0.194 0.178 0.075 -0.044 -0.214 0.483
Fitted Value 4.126 1.845 2.519 2.197 4.305 9.054
Model E
Residual 0.223 0.071 0.073 -0.082 -0.187 0.415
Fitted Value 4.127 1.810 2.533 2.197 4.303 9.071
Model F
Residual 0.222 0.106 0.059 -0.082 -0.185 0.398

Comparing the analyses can take several forms: the appropriateness of the model fit and/or how well
the explanatory variable accounts for variability in the response variable (i.e. comparing 𝑅 values).
Interpretation of slope
On their worksheet, students are asked to interpret the fitted slope for the original model and their new
model. The interpretation of the slope for the original model may not be fully appropriate (considering
problems were identified), but a standard interpretation 8 might be:

For every 1-millimeter increase in fish length, we expect a 0.1766-gram increase in fish weight,
on average.

Because these units are relatively small, it might make sense to adjust the interpretation, such as:

For every 10-millimeter increase in fish length, we expect a 1.766-gram increase in fish weight,
on average.

Depending on which model students chose to fit, interpreting the value of the fitted slope might be
more difficult or impossible; depending on your pedagogical goals, you may wish to omit the question
asking students to interpret the slope of their alternative model (perhaps just sharing correct
interpretations with them instead to illustrate the complexity). A full interpretation for the fitted slope
for each alternative model follows.

A. Because this is a simple linear regression model, the interpretation is more straightforward –
but there is still a challenge. A reasonable interpretation might be:

For every 1-millimeter increase in fish length, we expect a 0.04222- 𝑔𝑟𝑎𝑚 increase in fish
𝑤𝑒𝑖𝑔ℎ𝑡, on average.

In this model, the response variable is not 𝑊𝑒𝑖𝑔ℎ𝑡; the response variable is 𝑊𝑒𝑖𝑔ℎ𝑡 and the
interpretation should reflect that. The interpretation is more straightforward than for models A
and C because the explanatory variable is still simply 𝐿𝑒𝑛𝑔𝑡ℎ, but this model does not establish

8
Adjust the interpretation as needed to include aspects you value, such as restricting the interpretation to only apply to trout
perch collected from the Athabasca River and the Peace River.

Statistics Teacher/STatistics Education Web: Online Journal of K-12 Statistics Lesson Plans 10
[Link] or [Link]
Contact Author for permission to use materials from this lesson in a publication
 a linear relationship between 𝐿𝑒𝑛𝑔𝑡ℎ and 𝑊𝑒𝑖𝑔ℎ𝑡 – it instead establishes a linear relationship
between 𝐿𝑒𝑛𝑔𝑡ℎ and 𝑊𝑒𝑖𝑔ℎ𝑡, which is a nonlinear relationship.

B. As in model A, this is a relatively straightforward interpretation but still challenging because of

its nonlinear nature. A reasonable interpretation might be:

For every 1-millimeter increase in fish length, we expect a 0.02221- 𝑔𝑟𝑎𝑚 increase in fish
𝑤𝑒𝑖𝑔ℎ𝑡, on average.

In this model, the response variable is not 𝑊𝑒𝑖𝑔ℎ𝑡; the response variable is 𝑊𝑒𝑖𝑔ℎ𝑡 and the
interpretation should reflect that. The interpretation is more straightforward than for models A
and C because the explanatory variable is still simply 𝐿𝑒𝑛𝑔𝑡ℎ, but this model does not establish
a linear relationship between 𝐿𝑒𝑛𝑔𝑡ℎ and 𝑊𝑒𝑖𝑔ℎ𝑡 – it instead establishes a linear relationship
between 𝐿𝑒𝑛𝑔𝑡ℎ and 𝑊𝑒𝑖𝑔ℎ𝑡, which is a nonlinear relationship.

C. Because this model uses 𝐿𝑒𝑛𝑔𝑡ℎ as the only predictor, the fitted slope value of 0.001239 is
the expected increase in 𝑊𝑒𝑖𝑔ℎ𝑡 (in grams) for a 1-unit increase in 𝐿𝑒𝑛𝑔𝑡ℎ . Of course,
(1 𝑚𝑚) = 1 𝑚𝑚 , so the interpretation can be somewhat simplified. However, if students
which to modify the interpretation as was done above (i.e. describing a 10-millimeter increase
in 𝐿𝑒𝑛𝑔𝑡ℎ), this becomes much more challenging. It is also worth noting that this does not refer
to a 1-millimeter2 increase in the surface area of the fish: only the length is being measured, and
surface area is a distinct property which was not measured. The slope also incorporates the
change due to the linear term that was omitted, further complicating the interpretation.

D. As in model C, the interpretation is complicated but possible. Because this model uses
𝐿𝑒𝑛𝑔𝑡ℎ as the only predictor, the fitted slope value of 0.000011 is the expected increase in
𝑊𝑒𝑖𝑔ℎ𝑡 (in grams) for a 1-unit increase in 𝐿𝑒𝑛𝑔𝑡ℎ . Of course, (1 𝑚𝑚) = 1 𝑚𝑚 , so the
interpretation can be somewhat simplified. However, if students which to modify the
interpretation as was done above (i.e. describing a 10-millimeter increase in 𝐿𝑒𝑛𝑔𝑡ℎ), this
becomes much more challenging. It is also worth noting that this does not refer to a 1-
millimeter3 increase in the volume of the fish: only the length is being measured, and volume is
a distinct property which was not measured. The slope also incorporates the change due to the
linear and quadratic terms that were omitted, further complicating the interpretation.

E. Because this model includes both 𝐿𝑒𝑛𝑔𝑡ℎ and 𝐿𝑒𝑛𝑔𝑡ℎ as predictors, there is not a single slope
to interpret. This makes sense: interpreting the slope in linear regression describes the linear
change we expect in the response for a change in the explanatory. In this model, we are fitting a
nonlinear model and cannot simply interpret it as we would a linear model.

F. As in model E, the slope cannot be interpreted for this model because there is not a single
slope: there are three slopes for this nonlinear model. Interpreting the slope in linear regression
describes the linear change we expect in the response for a change in the explanatory. In this
model, we are fitting a nonlinear model and cannot simply interpret it as we would a linear
model.

Statistics Teacher/STatistics Education Web: Online Journal of K-12 Statistics Lesson Plans 11
[Link] or [Link]
Contact Author for permission to use materials from this lesson in a publication
If students fit a model using either a 4th-degree polynomial or a quartic root transformation, the
challenges in interpretation are similar to the above.

Evaluation of models

The scatterplot with the regression fit and the Residuals vs. Fits graph for each model are shown
below. While each model (A-F) is an improvement over Model 0, Model B is perhaps the best among
the models considered based on the Residuals vs. Fits graph (though not overwhelmingly better than
any other model). A note at the end of this document includes additional comments about residual
plots.

Scatterplot with fitted models and residual Residuals vs. Fits graph for Model 0 (Original)

Statistics Teacher/STatistics Education Web: Online Journal of K-12 Statistics Lesson Plans 12
[Link] or [Link]
Contact Author for permission to use materials from this lesson in a publication
 Scatterplots with fitted models and residual Residuals vs. Fits graphs for Models A, B, C, and D

Scatterplots with fitted models and residual Residuals vs. Fits graphs for Models E and F

Statistics Teacher/STatistics Education Web: Online Journal of K-12 Statistics Lesson Plans 13
[Link] or [Link]
Contact Author for permission to use materials from this lesson in a publication
 Attached Materials
The following files are included:
 Handouts
o Student handout (Word file)
o Selected computer output from the analyses for the Trout Perch data that can be used if
students do not have access to appropriate technology (Word file; included in the .zip
file containing the data files)
 Data files
o Trout Perch data, cleaned and appropriate for analysis (Minitab and .csv)
o Trout Perch data, cleaned and appropriate for analysis, with calculated variables
(Minitab and .csv)
o An additional fish dataset from the same source (Slimy Sculpin) that exhibits the same
relationships and can be used to create exam questions, assignments, etc.
o The raw data files for both Trout Perch and Slimy Sculpin (as they were downloaded
from the Environment and Climate Change Canada website).
 Technology guides
o Minitab
 A brief overview for using Minitab to perform the analyses in this lesson
 A Minitab workbook file containing various analyses already performed for the
Trout Perch Data.
o CODAP
 A brief overview for using CODAP to perform some of the analyses in this
lesson
 A CODAP file containing two of the models and residual plots as an example.
o TI-83/84
 A brief overview for using TI-83/84 calculators to perform the analyses in this
lesson (with a subset of the data)

Reflections and Additional Recommendations

Thoughts on extensions

Students are asked to make a prediction and calculate a residual for a single observation using both
Model 0 and the model that they propose. By itself, a single residual often has little value – even
answering questions such as “Is this residual particularly large (or small)?” require consideration of the
residual relative to others. This discussion of residuals can be extended to more precisely articulate the
shortcomings of Model 0: you can ask students to explain if the original model will systematically
overestimate or underestimate the weights of any types of fish. As indicated by the rectangles below,
the shortest and longest fish will tend to have their weights underestimated by the original regression
model. There is a corresponding tendency for fish with lengths closer to the average length to have
their weights overestimated, though this is less clear from the graph due to the abundance of
observations obscuring the line. Such a discussion might be prompted by assigning each group of

Statistics Teacher/STatistics Education Web: Online Journal of K-12 Statistics Lesson Plans 14
[Link] or [Link]
Contact Author for permission to use materials from this lesson in a publication
students a different observation (intentionally chosen because the model overestimates or
underestimates it9) and identifying patterns among the set of residuals across groups.

This discussion could also be related to extrapolation: to what extent would Model 0 or any of the
alternative models be appropriate for fish shorter than the 46 mm or longer than 108 mm (the
minimum and maximum, respectively)? While extrapolation can be dangerous and should not be
undertaken lightly, Model 0 would perform even worse for extrapolations, but the alternative models
might be more cautiously applied for fish close to the lengths in our dataset (e.g., 40 mm and 110 mm)
because we seem to have accounted for the salient curve in the data and we have a model that has some
basis in a physical relationship. Still, consulting with someone who has expertise in fish biology would
be advisable before using any of these models to make moderate extrapolations outside of a classroom
setting.

When using polynomial regression (and multiple linear regression in general), model over-fit is a
substantial concern. There can be a tendency for some students to want to fit very complex models
(e.g. degree n-1 polynomials) that can result in perfect fit for the dataset they are using but would be
terrible models when applied to any other dataset. This suggests two extensions:
1. Discuss the behaviour of the fitted models just outside the values that are included in the
dataset. Often, polynomial models will appear quite unrealistic for values even slightly outside
the dataset.
2. Explore model validation techniques such as cross-validation. Each group could be given a
slightly different dataset of randomly chosen fish. The groups could develop their models and
then apply them to the remaining fish and calculate appropriate fit statistics. Many textbooks on
linear regression include such a topic, including STAT2 by Cannon et al. (2018).

9
Some candidates for these observations are the fish with the following FishID: 3319, 3331, 3343, 3344 (long fish that are
all underestimated); 3327, 4052, 4151, 13320 (mid-length fish that are all overestimated); 4, 3259, 4091, 4198 (short fish
that are all underestimated). Note that because the FishID values are in a random order, it may be easiest to sort the dataset
by FishID in a spreadsheet program (such as Microsoft Excel) to identify these observations. Again, these observations
were intentionally chosen to illustrate the underestimation/overestimation pattern – there are, of course, observations that do
not fit this pattern, such as 1196, 3165, 3179, 4282 (mid-length fish that are all underestimated).

Statistics Teacher/STatistics Education Web: Online Journal of K-12 Statistics Lesson Plans 15
[Link] or [Link]
Contact Author for permission to use materials from this lesson in a publication
This dataset was collected as part of a study to monitor fish health in the context of development in
Canada’s Oil Sands Region. Specifically, fish were sampled upstream from developed sites to serve as
a reference for fish sampled downstream from developed sites that could exhibit environmental effects
from the developed site. Facilitating a discussion with students about how a model developed in this
lesson plan (e.g., Model B) could be used to determine if environmental impacts were occurring could
connect this lesson back to the research context and motivate further statistical considerations. In the
raw data included with this lesson, information about where fish were sampled is included (US =
upstream; DS = downstream); the dataset used in this lesson includes fish from both upstream and
downstream locations. Possible approaches students may consider are:
 Fitting a model and determining if new fish are much smaller (or larger) than predicted, leading
to a discussion about residuals and unusual points
 Including the upstream/downstream information in the model as a categorical variable. For less
experienced students, this might take the form of using colours to indicate upstream or
downstream group membership on a graph and/or fitting two regression models and visually
inspecting them for differences. See the graph below for an example. You could ask, “how
much difference in regression coefficients is enough difference to determine if the developed
sites seem to have an effect on fish?” to motivate inferential techniques with regression. More
advanced students could use more formal approaches and possibly consider interaction terms.

Note that the there is a rather small coefficient for the categorical predictor variable, but the p-value is
also rather small (about 0.001). This could motivate a discussion about statistical significance versus
practical significance.10

10
The term statistical significance is no longer recommended by many (e.g., Wasserstein et al., 2019). Such a small
difference coefficient could be used as an example of when the term significance might not convey the intended message.

Statistics Teacher/STatistics Education Web: Online Journal of K-12 Statistics Lesson Plans 16
[Link] or [Link]
Contact Author for permission to use materials from this lesson in a publication
Thoughts on differentiation

Depending on where students are at in their learning experience, you may wish to provide some of the
students with the dataset that contains the calculated variables already while other students may be
asked to use software to calculate the variables themselves. Moreover, some groups could be focused
on just one type of model (i.e. polynomial or transformation of the response) from the beginning,
which would provide more structure. Due to the complexity of the interpretations of the slope for
various models, providing some structure to this (e.g., providing students with a template and have
them fill in blanks) would be another way to support students with different learning needs and could
be combined with other modifications.

Thoughts on finding more datasets

Datasets appropriate for nonlinear regression modeling abound. If searching for more nonlinear
datasets for which polynomial regression is appropriate, focus on physical relationships (e.g., the
period of a pendulum or Boyle’s Law), scientific formulae, etc. as the inspiration: many nonlinear
relationships are well-established and appropriate for students learning regression. I chose to use fish
data in this lesson because the overall three-dimensional size (volume) of fish is often approximated
using a one-dimensional measure (length).11

Notes on Other Residual Plots

Students with more regression modeling experience may be expected to make residual diagnostic plots,
such as those shown in the graph below. (This graph is called a 4-in-1 Plot and is produced by Minitab.
It simply shows four residual diagnostic plots in a single graph. All four of these diagnostic plots were
produced using the (regular) residuals, though students may use standardized residuals, too. The graphs
in the 4-in-1 plot are: a normal probability plot of the residuals (top left), a histogram of the residuals
(bottom left), a scatterplot of the residuals (Y-axis) versus the fitted values (X-axis), and a line plot
with residuals on the Y-axis and the order of the observation in the dataset on the X-axis.)

11
Using fish as a context for nonlinear modeling has been done before in textbooks including STAT2 (Cannon et al., 2013,
2018) and The Practice of Statistics (Starnes et al., 2012). The dataset used in this lesson has not previously been used.

Statistics Teacher/STatistics Education Web: Online Journal of K-12 Statistics Lesson Plans 17
[Link] or [Link]
Contact Author for permission to use materials from this lesson in a publication
The most important of these graphs for our initial modeling purpose is Residuals vs. Fits: a clear
curved pattern is present, suggesting a nonlinear relationship between the variables. (Non-constant
variance may also be an issue, though addressing the nonlinearity would be the first priority. To see the
non-constant variance, note that the vertical distance between the largest and smallest residual at each
point on the x-axis systematically changes, indicated by the red lines. Among the fish with the smallest
length, there is less variability in weights; among fish that are of average length for the dataset there is
considerable variability in weights. This observation about non-constant variance is a separate problem
from the curvature.) The Residuals vs. Fits graph shows the same overestimation/underestimation
mentioned before, but the addition of the horizontal dashed line at 0 might make this clearer.

For Model B (perhaps the most appropriate of the alternative models), the 4-in-1 Plot is shown below.
Notice that both the curvature and non-constant variance apparent in the Residuals vs. Fits graph have
been considerably improved relative to Model 0. In the Residuals vs. Fits graph – a plot of the
residuals against the row number for each observation – there is an improvement in Model B relative
to Model 0, too: in Model 0, there seems to be an asymmetry in the magnitude of the positive residuals
when compared to the negative residuals (note that 0.0 is not in the middle of the Y-axis scale, and
there is no need for -5.0 to be labeled on the Y-axis scale). However, with Model B this asymmetry
appears to be improved: the magnitude of the positive residuals and negative residuals seem to be
about the same (note that 0.0 is in the middle of the Y-axis scale and there seems to be roughly equal
scattering above and below 𝑦 = 0). Lastly, the distribution of the residuals for Model B more closely
resembles the normal distribution. While not critical to the analysis in this lesson, the distribution of
the residuals becomes important when inferential methods are used. For Model 0, the distribution of
the residuals appears to be somewhat right skewed, but the distribution of the residuals for Model B
appears to be approximately bell-shaped. While the distribution of the residuals does not appear to be
exactly normal (as can be seen with the departures from the straight-line pattern in the Normal
Probability Plot), the skew apparent in Model 0 is no longer present – yet another improvement. The
Versus Order line plots for both Model 0 and Model B do not reveal potential patterns or relationships

Statistics Teacher/STatistics Education Web: Online Journal of K-12 Statistics Lesson Plans 18
[Link] or [Link]
Contact Author for permission to use materials from this lesson in a publication
between the residuals and the order the observations appeared in the dataset; if such a pattern existed, it
would warrant further investigation. Note that the asymmetry in the magnitude of the residuals
apparent in the Residuals vs. Fits graph for Model 0 and the improvement in this area for Model B can
also be seen in the Versus Order graphs.

Examining the 4-in-1 Plots for the other models will reveal similar results: all six models represent an
improvement relative to Model 0, but Models B, D, and F tend to represent greater improvement than
Models A, C, and E.

Further Reading

This lesson touches on many regression topics that are emphasized (or covered) in introductory
statistics courses: you may find it helpful to use a regression-focused textbook as a reference. While
there are many such books available, I reference four here to point you in the right direction. Cannon et
al. (2018) and Mendenhall and Sincich (2012) are regression books that focus on applications and have
minimal prerequisites. Kutner et al. (2005) and Bingham and Fry (2010) include both applications and
theory; these books routinely draw on knowledge of calculus and linear algebra. Every topic covered in
this lesson is included in each of these books. While not as comprehensive as a textbook, another
resource to consider are the course notes for Penn State’s STAT 501: Regression Methods course; these
notes are freely available online (see Lesson 9: Data Transformations 12).

12
The Pennsylvania State University, STAT 501: Regression Methods, Lesson 9: Data Transformations
[Link]

Statistics Teacher/STatistics Education Web: Online Journal of K-12 Statistics Lesson Plans 19
[Link] or [Link]
Contact Author for permission to use materials from this lesson in a publication
 References
Bingham, N. H., & Fry, J. M. (2010). Regression: Linear models in statistics. Springer.
Cannon, A. R., Cobb, G. W., Hartlaub, B. A., Legler, J. M., Lock, R. H., Moore, T. L., Rossman, A. J.,
& Witmer, J. A. (2013). STAT2: Building Models for a World of Data. W. H. Freeman and
Company.
Cannon, A. R., Cobb, G. W., Hartlaub, B. A., Legler, J. M., Lock, R. H., Moore, T. L., Rossman, A. J.,
& Witmer, J. A. (2018). STAT2: Modeling with Regression and ANOVA (2nd ed.). W. H.
Freeman and Company.
Environment and Climate Change Canada. (2019). Wild Fish Health, Oil Sands Region.
[Link]
region/wild-fish-health-oil-sands-region/
Kutner, M. H., Nachtsheim, C. J., Neter, J., & Li, W. (2005). Applied linear statistical models (5th
ed.). McGraw-Hill Irwin.
Mendenhall, W., & Sincich, T. (2012). A second course in statistics: Regression analysis (7th ed.).
Prentice Hall.
The Pennsylvania State University (n.d.). STAT 501: Regression Methods [online course].
[Link]@PSU. [Link]
Starnes, D. S., Yates, D. S., & Moore, D. S. (2012). The practice of statistics (4th ed.). W. H. Freeman.
Wasserstein, R. L., Schirm, A. L., & Lazar, N. A. (2019). Moving to a World Beyond “p < 0.05.” The
American Statistician, 73(sup1), 1–19. [Link]

Statistics Teacher/STatistics Education Web: Online Journal of K-12 Statistics Lesson Plans 20
[Link] or [Link]
Contact Author for permission to use materials from this lesson in a publication