longitudinal-growth-curves

Fit a piecewise linear model using exhaustive breakpoint search with BIC to select the number of segments, or use the ruptures library's Pelt algorithm with a BIC-derived penalty. Validate detected breakpoints against observable changes in content types and context diversity.

Fit all three candidate models and compare them using AIC and BIC, preferring BIC for explanatory analysis. A delta BIC above 10 indicates the lower-BIC model is clearly superior; values between 2 and 6 show only positive evidence.

AIC penalizes complexity by 2 per parameter and suits predictive model selection, while BIC penalizes by log(n) per parameter and favors parsimonious explanatory models. When they disagree, this workflow prefers BIC because the goal is explanation rather than prediction.

No, timelines shorter than 3 months or fewer than 12 time-windowed data points are insufficient for trajectory modeling. Short timelines cannot distinguish between linear, logistic, and piecewise model shapes reliably.

Gaps in activity are meaningful data indicating lurking or absence, so windows with zero activity must be filled rather than dropped. Removing them biases the fitted trajectory upward and distorts stage classification.

No, this approach models historical trajectories only and does not forecast future activity. It classifies past and current Digital Maturity stages and detects when transitions occurred, not what happens next.