A function model connects an input quantity to an output quantity. Choosing a model requires more than finding a formula through a few points: the pattern, units, domain, and behavior should all make sense for the situation.
Begin with the kind of change
A linear model describes constant additive change per input unit. An exponential model describes constant multiplicative change over equal input increments. A quadratic can describe a changing additive rate, while a sinusoidal model describes a repeating oscillation.
For equally spaced inputs, constant first differences suggest linear behavior; constant ratios suggest exponential behavior; constant second differences suggest quadratic behavior. These are diagnostic patterns, not proof that a real process follows the model forever.
Always check that the input spacing is equal before comparing raw differences or ratios. A doubling over two hours is different from a doubling over one hour.
Example: distinguish linear and exponential data
At times 0, 1, 2, and 3, a culture has counts 100, 150, 225, and 337.5 in a continuous approximation. The differences are 50, 75, and 112.5, so they are not constant. The ratios are all 1.5, suggesting
The initial factor 100 is the value at equal to zero. The multiplier 1.5 means a 50% increase per time unit, not an increase of 1.5 organisms. At equal to 4, the model predicts
A fractional modeled count is possible even when actual organisms are whole objects; the formula is an approximation or expected trend. Report precision appropriate to the context.
Equal time steps multiply the quantity by 1.5, so the graph curves upward instead of adding a constant amount.
Equal time steps multiply the quantity by 1.5, so the graph curves upward instead of adding a constant amount.
Describe this figure
Equal time steps multiply the quantity by 1.5, so the graph curves upward instead of adding a constant amount.
Marked points: filled marker at (0, 100); filled marker at (1, 150); filled marker at (2, 225); filled marker at (3, 337.5); filled marker at (4, 506.25).
Example: build a linear model from context
A delivery company charges a fixed fee plus a rate per kilometer. A 5 km delivery costs 18 dollars and a 9 km delivery costs 26 dollars. The rate is
dollars per kilometer. Substituting the first point into gives so equals 8. The model is
The intercept represents the fixed fee, not the price per kilometer. The natural domain includes nonnegative distances, perhaps restricted further by the service area. A negative-distance prediction has algebraic meaning but no meaning for this delivery setting.
Estimate parameters from more than two points
Real measurements usually do not lie exactly on a simple curve. Regression chooses parameters according to a fitting criterion, often minimizing squared vertical residuals. A good fit should leave residuals without a clear systematic pattern.
A high fit statistic alone does not guarantee an appropriate model. A curve can fit a small interval well yet behave implausibly outside it. Use the context to choose a family and the data to estimate its parameters, then assess discrepancies.
For example, a population cannot grow exponentially without limit in a resource-limited environment. A short-term exponential approximation may still be useful, but long-term extrapolation needs a different justification.
Interpolation and extrapolation carry different risks
Interpolation predicts within the observed input range. Extrapolation predicts beyond it. Both use the model, but extrapolation is usually more vulnerable because the pattern may change outside the data.
Suppose delivery prices were observed from 5 to 9 km. Predicting 7 km uses interpolation; predicting 100 km assumes the same pricing policy continues far beyond the observed range. A surcharge or maximum service distance could invalidate that assumption.
Also distinguish an explanatory relationship from a causal one. A fitted association between two measurements does not itself establish that changing one causes the other.
Try it yourself
Values at equally spaced inputs 0, 1, 2, 3 are 40, 34, 28, 22. Choose a simple model, interpret its parameters, and predict the value at input 5.
Answer: Constant differences of suggest The initial value is 40 and the rate is a decrease of 6 per input unit. The prediction is 10. If the output represents a nonnegative quantity, the model must stop or change before it predicts negative values.
What to remember
Choose a family from the pattern and context, determine parameters with units, restrict the domain sensibly, and check residual behavior and extrapolation limits.