Question

Comparing and Contrasting Predictions
Compare and contrast interpolations and extrapolations based on a scatterplot.

Answer 1

Definitions
- Interpolation: predicting values of the response for x-values that lie inside (between) the range of observed predictor values shown on the scatterplot.
- Extrapolation: predicting values for x-values that lie outside the range of observed predictor values (beyond the leftmost or rightmost points on the scatterplot).

Similarities
- Both use a model or pattern inferred from the scatterplot (e.g., a fitted line or curve).
- Both assume the relationship indicated by the plotted data continues in the region where you predict.
- Both produce a single-point prediction plus uncertainty; the uncertainty depends on model fit and data variability.

Key differences
- Domain: interpolation is inside the observed x-range; extrapolation is outside it.
- Reliability: interpolation is usually much more reliable. Extrapolation is more uncertain because it depends entirely on the assumed continuation of the pattern beyond known data.
- Sensitivity to model form: extrapolated predictions are far more sensitive to the chosen model (linear vs. quadratic vs. something else). Small changes in model form or parameter estimates can produce large differences outside the data range.
- Uncertainty behavior: standard prediction intervals typically widen quickly outside the observed x-range, reflecting increased uncertainty for extrapolation.
- Influence of extreme points: boundary and extreme observations have outsized influence on extrapolations (high leverage), so prediction can be dominated by a few points.

When each is appropriate
- Interpolation: preferred when you need predictions inside the observed range and the scatterplot shows a reasonably consistent pattern (no obvious model violations).
- Extrapolation: only done cautiously — useful for short-range extrapolation if you have good theoretical reason the relationship continues, or if domain knowledge/mechanistic models support it. Avoid long-range extrapolations from purely empirical scatterplots.

Practical checks on a scatterplot before predicting
- Look for a clear trend (linear, curvilinear) and homoscedastic residuals.
- Check how densely the x-range is sampled near the prediction point: sparse coverage near the boundary increases risk.
- Watch for nonlinearity or change in slope that suggests the pattern will not continue beyond observed values.
- Consider influential points at the edges: they can distort extrapolated trend.

Examples
- Interpolation: You have heights 150–190 cm and predict weight for 170 cm — reasonable if the trend is stable.
- Extrapolation: Using the same data to predict weight for 230 cm — risky because there are no observed examples that large; biological limits and nonlinearity likely make predictions unreliable.

Best practices
- Prefer interpolation when possible.
- If you must extrapolate, justify the model with subject-matter theory, quantify and report wider uncertainty, and seek more data outside the current range.
- Use diagnostic tools (residual plots, influence measures) and consider alternative models or mechanistic approaches when extrapolating.

Short takeaway
Interpolation uses the pattern inside your data and is usually safe; extrapolation reaches beyond observed data and is much more uncertain and model-dependent — do it only with strong justification and careful uncertainty assessment.