Scribe logoScribeNotes & practice
Sign inStart free
  1. Notes
  2. ›IB Mathematics: Applications and Interpretation
  3. ›4 Statistics and probability
  4. ›4.13 Non-linear regression

IB Mathematics: Applications and Interpretation · Topic 4 · Statistics and probability

4.13 Non-linear regression

Fit non-linear models to data with technology, interpret their parameters, and compare their fit with a linear model.

  • 5 min read
  • 6 parts
  • 10 questions
  • Draft
  • HL only
Start a Quick 5 on this topic→Jump to the 10 questions
1/6A straight line can fail to describe curved data
  1. 1A straight line can fail to describe curved data
  2. 2Each model type has its own equation and shape
  3. 3Technology finds the curve with the smallest residuals
  4. 4Parameters describe real features of the situation
  5. 5Fit statistics let you compare competing models
  6. 6A fitted model is reliable only inside its data range
  7. Check your recall
  8. Practice: 10 questions
1

Part 1 of 6: A straight line can fail to describe curved data

A cart rolls down a slope and speeds up. Its distance against time forms a curve, not a line. A straight line fitted to such data gives a poor description.

Fitting a curve of a chosen type to paired data is called non-linear regression. The aim is the same as for a line: a model that predicts one variable from another.

To judge any fit, look at each residual. A line fitted to curved data leaves residuals that form a pattern: positive, then negative, then positive again.

Figure 1. A straight line leaves systematic gaps, while the quadratic curve passes close to every point.

Key terms · tap one for its definition

non-linear regression

Non-linear regression is the fitting of a curved model, rather than a straight line, to paired data in order to predict one variable from another.

residual

A residual is the observed value of yyy minus the value of yyy predicted by the model at the same xxx.

Part 1 done

2

Part 2 of 6: Each model type has its own equation and shape

Now the task is choosing which curve to try. The scatter diagram and the context guide this choice.

Every model contains constants, such as aaa, bbb and ccc. Each constant is a parameter. Technology finds the parameter values that fit the data best, and you then interpret them.

Common regression models

ModelEquationTypical pattern
Quadraticy=ax2+bx+cy=ax^2+bx+cy=ax2+bx+cOne turning point
Cubicy=ax3+bx2+cx+dy=ax^3+bx^2+cx+dy=ax3+bx2+cx+dUp to two turning points
Exponentialy=abx+cy=ab^x+cy=abx+cConstant multiplier per unit of xxx
Powery=axby=ax^by=axbyyy changes with a power of xxx
Siney=asin⁡(bx+c)+dy=a\sin(bx+c)+dy=asin(bx+c)+dRepeating wave
Model
Quadratic
Equation
y=ax2+bx+cy=ax^2+bx+cy=ax2+bx+c
Typical pattern
One turning point
Model
Cubic
Equation
y=ax3+bx2+cx+dy=ax^3+bx^2+cx+dy=ax3+bx2+cx+d
Typical pattern
Up to two turning points
Model
Exponential
Equation
y=abx+cy=ab^x+cy=abx+c
Typical pattern
Constant multiplier per unit of xxx
Model
Power
Equation
y=axby=ax^by=axb
Typical pattern
yyy changes with a power of xxx
Model
Sine
Equation
y=asin⁡(bx+c)+dy=a\sin(bx+c)+dy=asin(bx+c)+d
Typical pattern
Repeating wave

A model should match the context as well as the points. A quadratic suits a thrown object, whereas a sine model suits a repeating cycle such as daily temperature. An exponential model suits growth by a constant percentage.

Key term

parameter

A parameter is a constant in a model’s equation whose value is chosen to fit the data.

Part 2 done

Quick check

One question on the part you just read

Easy

Identify the quantity that least squares regression makes as small as possible.
Show answer and explanation

Answer: A · Worked explanation

Least squares chooses parameters that minimise the sum of the squared residuals. The plain sum of residuals is not minimised, because positive and negative residuals cancel each other. A model can have residuals summing to zero and still fit badly.
3

Part 3 of 6: Technology finds the curve with the smallest residuals

The calculator needs a rule for deciding which curve is best. The usual rule is least squares. It chooses the parameters that make the sum of squares of residuals, written SSresSS_\mathrm{res}SSres​, as small as possible.

  • Enter the xxx values and yyy values in two lists.
  • Draw a scatter diagram and note its shape.
  • Choose the regression type that suits the shape.
  • Read the parameters and the fit statistics from the output.

Squaring stops positive and negative residuals from cancelling. It also gives large misses extra weight.

Key term

sum of squares of residuals

The sum of squares of residuals, SSresSS_\mathrm{res}SSres​, is the total of the squared residuals; a smaller value means the model lies closer to the data.

Part 3 done

4

Part 4 of 6: Parameters describe real features of the situation

A fitted equation helps only if you can explain what it says. Parameters usually carry meaning.

In the model y=abxy=ab^xy=abx, aaa is the value of yyy when x=0x=0x=0. The base bbb is the multiplier for each unit increase in xxx. A base of 1.4 means growth of 40% per unit.

In a quadratic, ccc is the value of yyy when x=0x=0x=0. The sign of aaa decides whether the curve opens up or down. In the sine model, ∣a∣|a|∣a∣ is the amplitude and ddd is the midline value.

Worked example

Interpreting and using an exponential model

A bacterial culture is modelled by N=250×1.4tN=250\times1.4^tN=250×1.4t, where NNN is the number of bacteria and ttt is the time in hours. At t=3t=3t=3 the observed count is 700. Interpret 250 and 1.4, then find the predicted count and the residual at t=3t=3t=3.
  1. 1
    At t=0t=0t=0, N=250×1.40=250N=250\times1.4^0=250N=250×1.40=250, so the culture starts with 250 bacteria.
  2. 2
    The base 1.4 multiplies the count each hour, a 40% increase per hour.
  3. 3
    Predicted count: N=250×1.43=250×2.744=686N=250\times1.4^3=250\times2.744=686N=250×1.43=250×2.744=686.
  4. 4
    Residual === observed −-− predicted =700−686=+14=700-686=+14=700−686=+14.

Answer: The predicted count is 686 bacteria and the residual is +14+14+14, so the model slightly underestimates the observed count.

Exam tip

When interpreting a parameter, state its meaning in context and include its units, not only its value.

Part 4 done

Quick check

One question on the part you just read

Easy

A data point has a residual of -3. State what this means.
Show answer and explanation

Answer: C · Worked explanation

A residual is observed minus predicted. A residual of -3 means the observed value is 3 below the prediction, so the model overestimates at that point. The statement that the model predicts a value 3 less than observed would correspond to a residual of +3+3+3.
5

Part 5 of 6: Fit statistics let you compare competing models

Suppose a line and a quadratic are fitted to the same data. A number is needed to compare them. The first is SSresSS_\mathrm{res}SSres​: the smaller it is, the closer the curve lies to the points.

The second is the coefficient of determination, R2R^2R2. A value near 1 shows a close fit. For a line, R2R^2R2 equals the square of the correlation coefficient.

Also inspect the residuals. Random scatter around zero suggests a suitable model. A visible pattern suggests that a better model exists.

Common mistake

A high R2R^2R2 does not prove a model is correct. A model can fit the data well yet make no sense in context.

Key term

coefficient of determination

The coefficient of determination, R2R^2R2, is the proportion of the variation in yyy explained by the model, between 0 and 1.

Part 5 done

6

Part 6 of 6: A fitted model is reliable only inside its data range

Predicting inside the range of the data is generally safe. Predicting beyond it is extrapolation, and it is far less reliable.

Beyond the data, a model may behave unrealistically. A quadratic eventually turns and changes direction, and an exponential model keeps growing without bound, though real resources run out.

Check that each prediction is sensible in context. A negative number of people or a negative length shows that the model has been pushed too far. If a line fits almost as well as a curve, the simpler line may be the better choice.

Go deeperWhy extra parameters make R2R^2R2 rise

Adding a term to a polynomial model can never increase SSresSS_\mathrm{res}SSres​, because the old model is a special case of the new one. So R2R^2R2 can only rise as terms are added.

This is why a higher R2R^2R2 alone is weak evidence. Prefer the simplest model that fits well and makes sense in context.

Key term

extrapolation

Extrapolation is the use of a model to predict values outside the range of the data used to fit it.

Note read

Check your recall[14 marks]

6 written questions. Answer each one in your head or on paper, then open the mark scheme.

  1. 1.Define[2 marks]

    Define the term residual.

    Show mark schemeHide mark scheme

    Model answer

    A residual is the observed value of yyy minus the value predicted by the model at the same xxx.

    • 1 mark for stating that it is a difference between an observed and a predicted value of yyy
    • 1 mark for giving the order observed minus predicted
  2. 2.Describe[2 marks]

    Describe how the pattern of residuals can show that a straight line is unsuitable for a data set.

    Show mark schemeHide mark scheme

    Model answer

    If the residuals are positive, then negative, then positive again, rather than randomly scattered about zero, the data are curved. A straight line is then unsuitable.

    • 1 mark for stating that the residuals show a clear pattern rather than random scatter
    • 1 mark for linking the pattern to curved data, so a line is unsuitable
  3. 3.Calculate[3 marks]

    A population is modelled by N=80×1.25tN=80\times1.25^tN=80×1.25t. At t=2t=2t=2 the observed value is 130. Calculate the predicted value and the residual.

    Show mark schemeHide mark scheme

    Model answer

    Predicted: N=80×1.252=80×1.5625=125N=80\times1.25^2=80\times1.5625=125N=80×1.252=80×1.5625=125. Residual =130−125=+5=130-125=+5=130−125=+5.

    • 1 mark for substituting t=2t=2t=2 into the model
    • 1 mark for obtaining a predicted value of 125
    • 1 mark for obtaining a residual of +5+5+5
  4. 4.Deduce[2 marks]

    In the model y=abxy=ab^xy=abx, a=40a=40a=40 and b=0.9b=0.9b=0.9. Deduce the value of yyy when x=0x=0x=0 and the percentage change in yyy for each unit increase in xxx.

    Show mark schemeHide mark scheme

    Model answer

    When x=0x=0x=0, y=40×0.90=40y=40\times0.9^0=40y=40×0.90=40. The multiplier 0.9 means yyy falls by 10% per unit increase in xxx.

    • 1 mark for deducing that y=40y=40y=40 when x=0x=0x=0
    • 1 mark for deducing a 10% decrease per unit of xxx
  5. 5.Explain[2 marks]

    Explain why a high R2R^2R2 alone does not show that a model should be used to predict values.

    Show mark schemeHide mark scheme

    Model answer

    A high R2R^2R2 shows a close fit only within the data. The model may still be unrealistic in context or outside the data range, where extrapolation is unreliable.

    • 1 mark for stating that a high R2R^2R2 describes the fit only to the data given
    • 1 mark for explaining that the model may be unrealistic in context or unreliable when extrapolated
  6. 6.Discuss[3 marks]

    Discuss the risks of using a fitted quadratic model to predict values well outside the range of the data.

    Show mark schemeHide mark scheme

    Model answer

    Predicting this far outside the data is extrapolation, so the pattern may not continue. A quadratic eventually turns, so it could predict values that make no sense, such as negative counts. Predictions should be checked in context and treated with caution.

    • 1 mark for identifying the prediction as extrapolation, which is unreliable
    • 1 mark for stating that a quadratic turns and may give unrealistic values
    • 1 mark for concluding that predictions must be checked against the context

Practice this topic

10 quick-fire multiple-choice questions, easy to hard. Tap an answer to see if you're right and the worked explanation. (The written questions with mark schemes are under “Check your recall” above.)

0 right0 of 10 answered

Includes the 2 quick checks in the note above.

3.

Easy·Identify[1 mark]

Identify the model best suited to data that repeat in a regular cycle, such as tide height over several days.
Show answer and explanation

Answer: D · Worked explanation

A sine model has a repeating wave shape with an amplitude, a midline and a period. A quadratic has a single turning point, so it cannot repeat. An exponential model keeps rising or falling, so it also cannot describe a regular cycle.
4.

Medium·Calculate[1 mark]

A model is y=500×0.8xy=500\times0.8^xy=500×0.8x. Calculate the predicted value of yyy when x=3x=3x=3.
Show answer and explanation

Answer: A · Worked explanation

0.83=0.5120.8^3=0.5120.83=0.512, so y=500×0.512=256y=500\times0.512=256y=500×0.512=256. The value 1200 comes from multiplying 500 by 0.8×30.8\times30.8×3, which treats the exponent as a multiplier. The value 200 comes from reducing 500 by 20% of the original amount three times in a straight line, which is not how a constant multiplier works.
5.

Medium·Identify[1 mark]

A population is modelled by N=1200×1.06tN=1200\times1.06^tN=1200×1.06t, where ttt is in years. Identify what the base 1.06 shows.
Show answer and explanation

Answer: B · Worked explanation

The base is the multiplier per unit of ttt. Multiplying by 1.06 is a 6% increase. The statement that the population increases by 1.06 each year confuses a multiplier with an addition, and 106% would be a multiplier of 2.06.
6.

Medium·Compare[2 marks]

A linear model for a data set has R2=0.64R^2=0.64R2=0.64 and SSres=180SS_\mathrm{res}=180SSres​=180. An exponential model has R2=0.97R^2=0.97R2=0.97 and SSres=18SS_\mathrm{res}=18SSres​=18. Compare the fits.
Show answer and explanation

Answer: C · Worked explanation

A smaller SSresSS_\mathrm{res}SSres​ means the curve lies closer to the points, and a larger R2R^2R2 means more variation is explained. The exponential model explains 97% of the variation, so 3% is the unexplained part. A large SSresSS_\mathrm{res}SSres​ indicates a worse fit, not a better one.
7.

Medium·Determine[1 mark]

The model y=4sin⁡(0.5x)+10y=4\sin(0.5x)+10y=4sin(0.5x)+10 has xxx in radians. Determine its period.
Show answer and explanation

Answer: B · Worked explanation

The period is 2πb\frac{2\pi}{b}b2π​, where b=0.5b=0.5b=0.5, so it is 2π0.5=4π\frac{2\pi}{0.5}=4\pi0.52π​=4π. The value 2π2\pi2π is the period when b=1b=1b=1, which ignores the factor 0.5. The value 8π8\pi8π is double the correct period, and π4\frac{\pi}{4}4π​ is far too short; neither follows from dividing 2π2\pi2π by bbb.
8.

Hard·Identify[2 marks]

A quadratic model fitted to data for xxx from 0 to 10 has R2=0.995R^2=0.995R2=0.995. It predicts a negative number of customers at x=40x=40x=40. Identify the best statement.
Show answer and explanation

Answer: D · Worked explanation

A value of R2R^2R2 near 1 describes the fit only inside the data range. Beyond it, extrapolation is unreliable, and a negative count is clearly unrealistic. Also, R2R^2R2 near 1 does not make SSresSS_\mathrm{res}SSres​ zero. SSresSS_\mathrm{res}SSres​ is zero, and R2R^2R2 exactly 1, only for a perfect fit.
9.

Hard·Deduce[2 marks]

A straight line is fitted to seven data points. Its residuals, in order of xxx, are +5+5+5, +1+1+1, -3, -4, -3, +1+1+1, +3+3+3. Deduce the best conclusion.
Show answer and explanation

Answer: A · Worked explanation

The residuals add to zero, as they do for any least squares line, so that fact cannot show a good fit. They are positive, then negative, then positive, a U-shaped pattern. This indicates the data curve upwards, so a quadratic opening upwards is worth trying.
10.

Hard·Calculate[2 marks]

A power model is y=1.5x3y=1.5x^3y=1.5x3. Calculate how many times larger yyy is when x=8x=8x=8 than when x=4x=4x=4.
Show answer and explanation

Answer: C · Worked explanation

At x=4x=4x=4, y=1.5×64=96y=1.5\times64=96y=1.5×64=96. At x=8x=8x=8, y=1.5×512=768y=1.5\times512=768y=1.5×512=768. The ratio is 76896=8\frac{768}{96}=896768​=8, which is 232^323 since xxx doubles. The value 6 comes from multiplying 2 by the exponent 3 instead of raising 2 to that power. In a power model, a change in xxx by a factor kkk changes yyy by a factor k3k^3k3 here.

Keep going

Next topic4.14 Linear combinations of random variablesPrevious topic4.12 Designing data collection

More in Topic 4 · Statistics and probability

  1. 4.9Normal distributionNotes · 10 questions
  2. 4.10Spearman's rank correlationNotes · 10 questions
  3. 4.11Chi-squared and t-testsNotes · 10 questions
  4. 4.12Designing data collectionNotes · 10 questions· HL only
  5. 4.13Non-linear regressionNotes · 10 questions· HL only
  6. 4.14Linear combinations of random variablesNotes · 10 questions· HL only
  7. 4.15Central limit theoremNotes · 10 questions· HL only
  8. 4.16Confidence intervalsNotes · 10 questions· HL only
  9. 4.17Poisson distributionNotes · 10 questions· HL only

All IB Mathematics: Applications and Interpretation topics →

Draft: written by Scribe and checked automatically, not yet by a subject reviewer.

Practice questions are original, written by Scribe and checked automatically for correctness.

Topic codes follow the official numbering. Not endorsed by IB Diploma Programme or any exam board.

On this page

0/6 read

  1. 1A straight line can fail to describe curved data
  2. 2Each model type has its own equation and shape
  3. 3Technology finds the curve with the smallest residuals
  4. 4Parameters describe real features of the situation
  5. 5Fit statistics let you compare competing models
  6. 6A fitted model is reliable only inside its data range
  7. Check your recall
  8. Practice: 10 questions
All coursesScribeHow it worksPricingPrivacyTerms

Notes and questions are written by Scribe in our own words. Topic codes follow the official course numbering so you can find your place. Not produced or endorsed by the IB, College Board, AQA, OCR, Pearson Edexcel or any exam board.

© 2026 Scribe

Line and curve fitted to the same data

Distance (m)0123456051015Time (s)
  • Straight line
  • Quadratic curve
  • Data