Home › Glossary
Glossary
Plain definitions of the design and statistics terms you meet across OptiGein Design. Each entry points to the page that explains it in context, so you can read a full definition here and the worked detail there.
How to read the letters below.
Terms are grouped by first letter. Standard field vocabulary (ANOVA, p-value, desirability) is used
as is; where the software gives a term a particular meaning on screen, the definition matches what
you actually see.
A
| Term | What it means |
|---|---|
| Adequate precision | A signal to noise measure. It compares the range of predictions across the design points against the average prediction error. A value above about 4 means the model is good enough to navigate the design space. See Fit summary, model and ANOVA. |
| Adjusted R squared | R squared corrected for the number of terms in the model, so adding a term that does not really help no longer inflates the score. See Fit summary, model and ANOVA. |
| Alias | When a design cannot tell two effects apart, they are aliased: their estimates add together and the software reports them as one. The Alias Matrix lists them. See Evaluating a design. |
| ANOVA | Analysis of variance. The table that splits the variation in a response into the part each model term explains and the part left over, then tests whether each part is large enough to matter. See Fit summary, model and ANOVA. |
| Axial point | In a central composite design, a run placed out along one factor axis, on the face of the factorial box or beyond it, used to estimate curvature. Also called a star point. See Response surface designs. |
B
| Term | What it means |
|---|---|
| Box-Behnken | A response surface design that places runs at the midpoints of the edges of the factor space plus centre points, so it fits a quadratic model without any extreme corner settings. See Response surface designs. |
| Box-Cox | A plot that suggests whether a power transform of the response would steady the spread of the residuals, and which power to use. See Diagnostics. |
C
| Term | What it means |
|---|---|
| Centre point | A run with every numeric factor set to the middle of its range. Repeated centre points measure pure error and reveal curvature. The example design carries five of them. See Response surface designs. |
| Coded factor | A factor rescaled so its low level is minus one and its high level is plus one, with the centre at zero. Coefficients on coded factors can be compared directly because they share a scale. See Fit summary, model and ANOVA. |
| Confidence interval | A range that is likely to contain the true mean response at a given set of factor settings. The Point Prediction page reports it. See Prediction and confirmation. |
| Cook's distance | An influence measure for a single run. A large value means dropping that run would noticeably change the fitted model. See Diagnostics. |
| Corrected total | The total variation in the response about its mean, before the model explains any of it. It is the bottom line of the ANOVA table and equals the model plus residual sums of squares. See Fit summary, model and ANOVA. |
| C.V. percent | The coefficient of variation: the residual standard deviation written as a percentage of the mean response, a quick read on the relative noise in the experiment. See Fit summary, model and ANOVA. |
D
| Term | What it means |
|---|---|
| Degrees of freedom | The number of independent pieces of information behind a quantity, roughly the number of runs minus the number of things estimated from them. See Fit summary, model and ANOVA. |
| Desirability | A score from 0 to 1 that says how well one set of factor settings meets a goal. The optimizer combines the desirability of every response into one overall score and searches for its highest value. See Numerical optimization. |
| DFBETAS | For one run and one model term, how far that run shifts the term's coefficient, measured in standard errors. It shows which run drives which coefficient. See Diagnostics. |
| DFFITS | How far a single run shifts its own fitted value, in standard errors. Another way to spot a run with heavy influence. See Diagnostics. |
F
| Term | What it means |
|---|---|
| Factor | An input you set on purpose and vary across the experiment, such as a concentration or a temperature. Factor A carries the letter; its label is the name you type, so a column reads like A:HPMC E15 (% w/v). See The design workspace. |
| Fold-over | A second block of runs built by reversing the signs of a fractional factorial, added to break aliases and raise the resolution of the combined design. See Factorial designs. |
| Fraction of design space | The FDS graph reads the fraction of the design region along the horizontal axis against the largest prediction error over that fraction, so you can see how evenly the design predicts. See Evaluating a design. |
H
| Term | What it means |
|---|---|
| Hierarchy | The rule that if a model keeps a higher order term it should also keep the lower order terms inside it, for example keeping A and B whenever AB is present. The software preserves hierarchy when it builds an equation. See Fit summary, model and ANOVA. |
I
| Term | What it means |
|---|---|
| Interaction | When the effect of one factor depends on the level of another. An interaction term such as AB captures that dependence. See Fit summary, model and ANOVA. |
L
| Term | What it means |
|---|---|
| Lack of fit | A test that compares the model's error against the pure error from replicated runs. A significant lack of fit means the chosen model shape misses something real in the data. See Fit summary, model and ANOVA. |
| Lambda | The power in a Box-Cox transform. Lambda of 1 is no transform, 0 is a log and 0.5 is a square root. See Diagnostics. |
| Leverage | How much a single run's own settings pull the fitted model toward its response. Leverage runs from 0 to 1, and a value near 1 means the run largely fits itself. See Diagnostics. |
M
| Term | What it means |
|---|---|
| Main effect | The average change in a response as one factor moves from its low to its high level, on its own. See Factorial designs. |
| Mixture | An experiment where the factors are the proportions of the parts of a blend, so they must add up to a fixed total, usually 100 percent. See Mixture designs. |
| Model order | How complex the fitted equation is: linear (main effects only), two factor interaction, quadratic, or cubic. See Fit summary, model and ANOVA. |
O
| Term | What it means |
|---|---|
| Orthogonal | A design property where the factor columns are uncorrelated, so each effect is estimated independently of the others without inflating its error. See Evaluating a design. |
P
| Term | What it means |
|---|---|
| Predicted R squared | A measure of how well the model predicts a run that was left out, built from PRESS. It should stay close to the adjusted R squared; a wide gap warns of overfitting. See Fit summary, model and ANOVA. |
| Prediction interval | A range likely to contain a single new measurement, not the mean, at a given set of settings. It is wider than the confidence interval. See Prediction and confirmation. |
| PRESS | Predicted residual error sum of squares. Each run is dropped in turn, the model is refit, and the miss on the dropped run is squared and summed. Smaller is better. See Fit summary, model and ANOVA. |
| Pure error | The variation between replicated runs at identical settings. It measures the noise of the experiment on its own, with no help from any model. See Fit summary, model and ANOVA. |
| p-value | The chance of seeing an effect this large if the term truly did nothing. A small p-value, commonly below 0.05, is read as evidence the term matters. See Fit summary, model and ANOVA. |
R
| Term | What it means |
|---|---|
| Randomisation | Performing the runs in a shuffled order so that anything drifting over time is not mistaken for a factor effect. The Run column holds this order. See The design workspace. |
| Replicate | A run repeated at exactly the same factor settings. Replicates supply the pure error estimate. See Which design do I need?. |
| Residual | The gap between a measured response and the value the model predicts for that run. The diagnostics work almost entirely with residuals. See Diagnostics. |
| Resolution | A rating of a fractional factorial that says how badly effects are aliased. Resolution V keeps main effects and two factor interactions clear of each other; lower numbers mix them. See Factorial designs. |
| Response | An output you measure on each run, such as disintegration time or hardness. You analyse one response at a time. See The design workspace. |
| Response surface | A smooth fitted model, usually quadratic, that maps the responses across the factor space so you can find peaks, valleys and good operating regions. See Response surface designs. |
| Run | One experiment: a single row of the run sheet, with its factor settings and the responses you record. The example design has thirteen. See The design workspace. |
S
| Term | What it means |
|---|---|
| Scheffe model | The classical polynomial form for mixture data, which drops the intercept and keeps one linear term per component. OptiGein Design fits the equivalent slack variable form instead: it keeps the intercept and aliases away one component term. Both give identical predictions. See Mixture designs. |
| Screening | An early stage that uses a small design to sift many factors and find the few that actually matter, before studying them in detail. See Factorial designs. |
| Sequential sum of squares | The Type I breakdown that adds model terms in order and credits each with the extra variation it explains on top of the terms already in. It drives the model order suggestion on Fit Summary. See Fit summary, model and ANOVA. |
| Simplex | The triangular, or higher dimensional, region of allowed blends in a mixture problem, where the components add up to the total. Simplex-lattice and simplex-centroid designs place runs across it. See Mixture designs. |
| Standard order | The tidy, unrandomised order in which the design is written out on paper, before the run order shuffles it. See The design workspace. |
| Studentized residual | A residual divided by its own standard error so all residuals share one scale. The internally studentized version uses every run; the externally studentized version leaves the run out first, which makes outliers stand out. See Diagnostics. |
T
| Term | What it means |
|---|---|
| Tolerance interval | A range expected to contain a stated fraction of the whole population of future results, at a stated confidence. The Confirmation page reports it. See Prediction and confirmation. |
| Transform | A mathematical reshaping of the response, such as a log, square root or inverse, applied before fitting to steady the variance or straighten a relationship. See Diagnostics. |
| Two factor interaction | An interaction between exactly two factors, written like AB. It is the simplest kind of interaction and the first curvature a factorial can pick up. See Fit summary, model and ANOVA. |
V
| Term | What it means |
|---|---|
| VIF | Variance inflation factor. It says how much a term's variance is enlarged because that term overlaps with the others. A VIF of 1 means the term is estimated independently. See Evaluating a design. |
Cannot find a term.
If a word on screen is not listed here, it usually appears with its own short explanation right under
the table or graph that uses it. The analysis and
diagnostics pages carry the most of these notes.