8 Interpreting and Communicating Quantitative Results

A framework for quantitative interpretation

A quantitative result is meaningful only when readers know what quantity it describes, how it was obtained, and where it can reasonably be used. Interpret every result as evidence produced under particular assumptions rather than as an unquestionable fact.

A useful report connects five elements:

1. The quantity estimated, measured, or predicted.

2. Its numerical value and units.

3. Its uncertainty or interval.

4. Evidence supporting the calculation or model.

5. Limitations that restrict appropriate use.

A concise conclusion can follow this pattern:

Under the stated assumptions, the model predicts [quantity] = [value] [units], with [uncertainty or interval]. The result was [verified, calibrated, or validated] using [evidence]. It is most reliable for [domain] and should not be extrapolated to [limitation].

Takeaway: A number becomes useful information only when its meaning, reliability, and limitations are communicated together.

Quantity, units, and dimensional consistency

Start by naming what was measured or calculated, then report the value with its units. Units identify the meaning of a number and provide a rapid check for invalid calculations.

For example, if a cyclist travels 30 miles in 2 hours, the average rate is

For a relationship $y=f(x)$, the average rate of change is $\frac{\Delta y}{\Delta x}$, and its units are the units of $y$ divided by the units of $x$. The instantaneous rate $\frac{dy}{dx}$ has the same unit structure. Thus, if temperature $T$ is measured in degrees Celsius and time $t$ in minutes, $\frac{dT}{dt}$ has units of degrees Celsius per minute.

Dimensional analysis checks whether terms in an equation are compatible. For

$d$ has units of length, $v_0t$ has units of length, and $at^2$ has units of length. The equation is dimensionally consistent. By contrast, $d=v_0+at^2$ cannot be valid as written if $v_0$ has units of length per time, because the terms being added do not have matching units.

Retain units throughout calculations and attach them to the final value. A slope of 4 in a graph of mass versus time means a rate such as 4 kilograms per day, not merely “a slope of 4.”

What kind of result is it?

Not all numerical statements have the same status. Distinguish among these categories:

• An exact value follows from a definition, exact conversion, or algebraic identity.

• An estimate is obtained through approximation, interpolation, or numerical computation.

• A measurement is based on observation and is affected by instrument limitations, sampling, and natural variability.

• A model prediction is generated from equations or rules and depends on assumptions about how a system behaves.

The wording should make the category visible. “The model predicts a concentration of $12.4\ \text{mg/L}$” communicates more than “The concentration is $12.4\ \text{mg/L}$,” because it identifies the value as a prediction rather than a direct observation.

The number of reported digits should also match the quality of the evidence. Extra digits can falsely imply precision that the measurement or model does not support.

Takeaway: State whether a result is exact, estimated, measured, or predicted before asking readers to interpret it.

Uncertainty and propagation

Uncertainty describes incomplete knowledge or variability associated with a result. It can arise even when a measurement is performed carefully. Common sources include random variation among repeated observations, systematic effects that shift results in one direction, uncertain model parameters or initial conditions, simplifying assumptions, and rounding or other numerical approximations.

A basic statement has the form

where $\hat{x}$ is the reported estimate and $u$ is an uncertainty measure. Explain what $u$ means: it might be a standard uncertainty, a confidence interval, a prediction interval, or another specified range. The symbol $\pm$ alone does not state the probability or confidence associated with the interval.

When an output depends on uncertain inputs,

input uncertainty can propagate into uncertainty in $Y$. For relatively small, independent uncertainties, a first-order approximation is

The partial derivatives show how strongly the inputs affect the output. If inputs are correlated, covariance terms may also be needed.

For a distance $d=(100.0\pm0.5)\ \text{m}$ and time $t=(20.0\pm0.2)\ \text{s}$,

The approximate relative uncertainty is

which gives approximately $u_v=0.06\ \text{m/s}$. A suitable report is therefore

provided the uncertainty definition and assumptions are stated. Reporting $5.000000$ would suggest more precision than the data justify.

Takeaway: Report uncertainty with a clear definition, not merely with a symbol or extra digits.

Sensitivity and influential inputs

Sensitivity analysis asks which inputs influence an output most. For a model $y=f(x_1,x_2,\ldots,x_n)$, local sensitivity to $x_i$ is often represented by

A large magnitude means that a small absolute change in $x_i$ can produce a relatively large change in $y$ near the evaluated point. Because variables may have different units and scales, a dimensionless measure can be more informative:

This quantity is an elasticity. It approximates the percentage change in $y$ caused by a 1% change in $x_i$. For $y=Ax^2$,

Thus, a 1% increase in $x$ produces approximately a 2% increase in $y$. This can be more useful than the predicted value alone because it indicates which measurements or assumptions deserve attention.

Sensitivity can be studied by changing one input at a time, calculating derivatives or elasticities, varying several inputs through a designed experiment, or using simulations such as Monte Carlo sampling. Sensitivity and uncertainty answer different questions: sensitivity identifies influential inputs, whereas uncertainty analysis describes how uncertain the output is given uncertain inputs and model assumptions.

Takeaway: Use sensitivity to prioritize measurement, monitoring, and robustness checks.

Evaluating models with evidence

Model evaluation has three related but distinct parts.

Verification asks whether the stated mathematical problem was implemented and solved correctly. Useful checks include testing code or spreadsheet formulas, checking units and signs, comparing with an exact solution or simple limiting case, checking conservation laws or known identities, and examining numerical convergence as a step size or grid spacing changes.

Calibration adjusts selected model parameters so that model outputs agree with selected observations. A transparent calibration records which data were used, which parameters were adjusted or held fixed, the error measure or objective, and the conditions represented by the data. Excessive flexibility can cause a model to fit noise rather than the underlying process.

Validation asks whether the model represents the real system well enough for its intended purpose. Compare predictions with independent observations or trusted reference results over relevant conditions. Agreement with calibration data alone is not strong evidence of predictive ability. No finite validation study proves that a model is correct in every circumstance; it provides evidence within a stated domain.

Useful comparison measures include absolute error, relative error, mean absolute error, root mean square error, bias, graphical residual analysis, and prediction intervals. A single statistic is rarely enough. A model can have a small average error while failing consistently at extreme values or for a particular subgroup.

A residual is the difference between an observed value and its fitted or predicted value. Residual plots can reveal patterns that an overall error statistic hides.

Takeaway: Verification checks implementation, calibration fits selected data, and validation tests suitability for intended use.

Fitted models and data range

A fitted function summarizes patterns in data, but it does not automatically establish causation. For a linear model

interpret $b_0$ as the predicted value of $y$ when $x=0$ only when zero is meaningful and lies within the relevant range. Interpret $b_1$ as the predicted change in $y$ for a one-unit increase in $x$, including appropriate units. Examine residuals and restrict interpretation to the range supported by the data.

Interpolation estimates within the observed data range and is usually less risky than Extrapolation, which predicts beyond that range. Extrapolation may fail when a mechanism changes, a constraint is reached, or a relationship is valid only locally.

A high correlation or strong goodness-of-fit statistic does not by itself prove causation, correct model structure, or reliable future prediction. Consider study design, omitted variables, measurement quality, dependence among observations, selection bias, and whether the data range matches the intended use.

Takeaway: Interpret fitted models as conditional summaries of data, not as automatic explanations of cause or guarantees of future behavior.

Decisions, optimization, and change over time

An optimization result is meaningful only when its objective and constraints are explicit. A complete interpretation states:

1. the quantity being maximized or minimized;

2. the decision variables;

3. the constraints;

4. whether the solution is local or global, when known;

5. how parameter changes affect the solution; and

6. whether the solution is practically feasible.

A design that minimizes cost while ignoring reliability may not be the best design for real use. If a small change in an uncertain parameter moves the optimum substantially, the recommended decision may not be robust. If several alternatives perform nearly equally well, the supposedly optimal choice may not be decisively better in practice.

For differential-equation models, interpret both the rate law and its parameters. For example,

has solution

where $P_0$ is the initial amount and $r$ has units of inverse time. If $r=0.03\ \text{year}^{-1}$, the model's instantaneous proportional growth rate is 3% per year near the modeled conditions. This does not mean that the quantity increases by exactly 3 percentage points each year, because exponential growth compounds continuously.

Differential-equation models can depend strongly on initial conditions, parameters, and boundary conditions. A model that fits a short interval may become unreliable over a longer interval if constant growth, constant forcing, or unchanged environmental conditions no longer hold.

Takeaway: Interpret optimization and dynamic-model results in relation to objectives, constraints, parameters, conditions, and feasibility.

Limitations and clear communication

Limitations define where a result can reasonably be used; they are part of the result rather than an optional disclaimer. Important limitations may involve restricted data ranges, small samples, measurement error, missing observations, assumptions about independence or linearity, constant parameters, omitted mechanisms, uncertain initial or boundary conditions, dependence on a particular calibration set, extrapolation beyond validated conditions, numerical approximation, or mismatch between a model output and the decision being made.

A specific limitation is more useful than a vague warning. For example:

The model was calibrated using summer observations from 2018–2022 and has not been tested during winter conditions; predictions outside that temperature range should therefore be treated as uncertain.

Tables are useful when exact values, units, and uncertainty estimates must be compared. Graphs are useful for showing trends, variability, residuals, and departures from a model. Effective graphs include labeled axes and units, a clear title or caption, a legend when needed, uncertainty bars or bands when relevant, a scale that does not exaggerate or conceal meaningful differences, and a clear distinction between observations and predictions.

When a difference is small relative to uncertainty, say so directly rather than treating point estimates as decisively different. The goal is to match the strength of the claim to the strength of the evidence.

Takeaway: State the domain of reliability and the conditions under which the result should not be used.