Must a growth curve increase?

No. The term includes decline and other patterns of change.

Is a straight line allowed?

Yes. A linear growth curve assumes a constant rate over the modeled period.

Does a fitted slope establish an intervention effect?

No. Causal interpretation requires additional design and assumptions.

AI Glossary / Statistics and probability / Intermediate

Growth curve (statistics)

Definition

A growth curve in statistics describes how an outcome changes over time, often using repeated measurements from the same individuals or other units. Growth-curve models estimate features such as starting level, rate of change, and curvature, and can describe differences in those features across units. The name does not require the outcome to increase: a trajectory can rise, fall, level off, or change direction. This entry concerns statistical longitudinal trajectories, not the growth function of a hypothesis class in learning theory.

What it means

A single before-and-after difference compresses a process into two observations. A growth curve uses the sequence of measurements to describe the pattern of change. Two units can reach the same endpoint after starting at different levels and changing at different rates.

Repeated measurements also create dependence: observations from one person, store, or machine are usually related. A suitable statistical model accounts for that structure instead of treating every row as an unrelated case.

Growth-curve analysis separates the average trajectory from variation around it. Units may differ mainly in starting level, rate of change, or more complex features. The model does not automatically explain why they differ.

How it works

A simple continuous-outcome model is y_it = (b0 + u0_i) + (b1 + u1_i)*t + e_it. Here b0 and b1 describe the average intercept and slope, u0_i and u1_i describe unit-specific deviations, and e_it represents remaining measurement-level variation.

Time zero determines the intercept's interpretation, and time units determine the slope's units. Mixed-effects and latent-growth formulations provide related ways to express these models. Covariates can explain some differences in trajectories, while residual correlation may require additional modeling.

Choose a shape suited to the process and observation window: linear, quadratic, piecewise, or a substantively justified nonlinear function. Check observed trajectories, model fit, measurement comparability, and missing-data assumptions. More complicated curves require adequate observations and information; software convergence alone does not establish that all parameters are meaningfully identified.

Three practical examples

Learning assessment: describe changes in repeated performance scores, distinguishing average progress from differences in starting level and improvement rate.
Business operations: model a store-level outcome over time while accounting for repeated observations within stores and relevant operating conditions.
AI-supported monitoring: summarize trajectories of equipment measurements, checking whether a simple trend adequately describes observed change before relying on its predictions.

Think of it this way

Think of several travelers on different paths. A photograph of their final positions cannot reveal where each began or how quickly each moved. A growth curve uses successive observations to describe their paths. The average path is informative, but no traveler is required to follow it exactly.

Worked example

Consider an original descriptive example with three units measured at t = 0, 1, and 2 months. Unit A has values 10, 12, 14; unit B has 12, 13, 14; and unit C has 8, 11, 14.

Their exact linear trajectories are A: 10 + 2*t, B: 12 + 1*t, and C: 8 + 3*t. Average intercept is (10 + 12 + 8)/3 = 10, and average slope is (2 + 1 + 3)/3 = 2 units per month. The average trajectory is therefore 10 + 2*t, with values 10, 12, and 14, matching the directly calculated mean at each occasion.

All three units finish at fourteen, but their initial levels and slopes differ. Relative to the average curve, the intercept deviations are 0, 2, and -2, while slope deviations are 0, -1, and 1. Higher starts accompany slower growth in this constructed example.

Recenter time as t_new = t - 1. The intercepts become 12, 13, and 11, representing month-one levels, while slopes remain 2, 1, and 3. Independent least-squares fits confirm the original coefficients. These noise-free lines illustrate trajectory parameters and time centering; three units are not a sufficient basis for confident population inference, and this calculation is not a fitted random-effects likelihood analysis.

Illustrative case study

In a fictional employee-training project, analysts measure the same skill outcome at several planned occasions. They examine individual trajectories before fitting an average growth model. Some participants start higher but improve more slowly, while others show rapid early gains followed by a plateau.

The team compares plausible curve shapes and checks whether the assessment remains comparable across occasions. It also investigates whether participants with weaker results disproportionately stop responding. Any claim that training caused improvement requires an appropriate comparison design beyond the growth model itself. This fictional case illustrates longitudinal description and model checking rather than an established program effect.

Common misconceptions and mistakes

Common mistakes include treating repeated observations as independent, interpreting an average curve as every individual's path, and using equally spaced time codes when actual intervals differ.

A positive average slope does not mean every unit improved. Differences between age groups measured once are not automatically equivalent to within-person change. Likewise, a nonlinear curve can fit the observed period while producing implausible long-range extrapolations.

Missing observations can be informative. Methods that use available records still depend on assumptions about the missing-data process; they do not automatically remove dropout bias. Changing instruments or definitions can also mimic growth. Separate statistical trajectory modeling from causal attribution, measurement validation, and decisions about intervention.

Business use

Use growth curves when the timing and pattern of repeated outcomes matter to a decision. Specify the time origin, measurement schedule, unit of analysis, plausible trajectory shape, and dependence structure. Report uncertainty and meaningful variation across units. For business use, translate slopes into the relevant outcome per time unit and state the period over which the model has supporting evidence.

Check your understanding

Answer: Two units both reach a score of twenty after two months.

Can their growth curves differ?

Answer: Yes. One might start at sixteen and gain two points per month; another might start at ten and gain five. Endpoint equality does not imply equal starting levels or rates of change.

Related terms

Longitudinal data
Nonlinear regression
repeated measures
random intercept
random slope
mixed-effects model
latent growth model
time centering

Learn next

Learn repeated-measure dependence and intercept-slope interpretation first. Then study nonlinear trajectories, measurement invariance, missing-data assumptions, and the distinction between longitudinal association and causal effects.

Missing data

Sources

Muthén and Muthén, Mplus User's Guide, Chapter 6: https://www.statmodel.com/HTML_UG/chapter6V8.htm — primary examples of linear, quadratic, piecewise, and correlated-residual growth models. Grimm and Ram, Non-linear Growth Models in Mplus and SAS: https://pmc.ncbi.nlm.nih.gov/articles/PMC3717396/ — original authors' indexed abstract and methodological passages on nonlinear trajectories; direct page retrieval encountered a browser check. The three-unit calculation is original and independently verified.

Previous: Growing self-organizing map Next: Growth function