Build an optimal design step by step: describe your model (or models),
set the assumed parameter values and the criterion, then compute.
No R code required.
Step: what kind of design problem is this?
Which one do I want?
Classical
— "I know the model, and I want the
most precise estimates of its parameters."
Compound
— "I am not sure whether the
interaction is real, and I want a design that is good
either way"; or "I care about the slopes of one model
and the overall fit of another"; or "my collaborators
disagree about the model."
The compound branch reports how efficient the design is
for each objective, and what a design optimised for any
one objective alone would have cost under the others.
Note.
The compound branch builds approximate
(continuous-weight) and exact (integer-run) designs, and runs the
same simulation study. Fitting the parameters from a data set is
available in the classical branch only — fit them there, or
elsewhere, and type the estimates in as each objective's assumed
values.
Step: the model
This will be a locally optimal design.
For a
non-linear model the information matrix depends on the
parameters themselves, so the design is optimal only at an
assumed parameter value. You will be asked for those values
(or you can draw them at random) in a later step.
Step: what do you already have?
With an existing design or data set the new design is chosen to
complement what you already have (a second-stage design).
Step: your existing design
Covariate columns plus a 'count' column (integer runs) or a
'weight' column (proportions) — the format the app downloads.
Factor covariates use integer levels 1..L.
With a 'count' column, n0 is filled in from the counts.
With weights you must supply it: proportions carry no sample size.
Step: your existing data set
Covariate columns plus a response column, one row per run.
Factor covariates use integer levels 1..L.
Estimating the parameters fits the model you described in step 1
to these data (maximum likelihood) and uses the estimates as the
assumed values — you can still edit them afterwards. The data set
is read as soon as you upload or paste it; the button below just
reads it again.
Step: assumed parameter values (theta)
A locally optimal design depends on these values. Type them in,
or draw them at random and edit.
One experiment, so one set of covariates and one design
region — shared by every objective. The models come next.
Step: the objectives
Each objective is a model plus a criterion. They may be
entirely different models — different families, different
terms, different numbers of parameters.
Step: weighting and any existing design
The approximate optimum is computed first, rounded to whole
runs, then improved by random exchanges.
Same format as the download: one column per covariate plus a
'count' or 'weight' column.