Everywhere else on this site works with a single dataset. This page asks if GRIN can recover the true representation from many simulated participants, and do its uncertainty intervals mean what they say?
Because the data are simulated, the true answer is known for every participant, so the estimates can be checked directly against it. For the same kind of comparison on your own or published data, visit the Analyse page.
On 24,000 held-out simulated matrices, mean recovery correlations were 0.91 for the eight location parameters and 0.59 for the four within-stimulus correlations. The nominal 90% marginal intervals covered the truth 92.8% of the time.
These are pooled results from the fixed research checkpoint on test matrices spanning 5–500 trials per stimulus. The small live exercise below is for intuition and will vary from run to run; it uses the same browser weights as Analyse.
Sensitivities (z-scores)
Correlations (ρ)
Each point is one parameter of one simulated participant. Its true value across and GRIN's estimate up. Points on the dashed diagonal were recovered exactly.
An interval is only useful if its stated coverage is right. For each stated level of confidence, a 50% interval, an 80% interval, and so on, this plots how often the true value actually fell inside. A calibrated method sits on the dashed diagonal: its 90% intervals contain the truth about 90% of the time. Bowing above the line is under-confidence and below it is over-confidence.
The sensitivities land tightly on the diagonal. The correlations are looser, because a weak within-stimulus correlation leaves only a faint trace in a single confusion matrix. That is a limit of what the data carry rather than of the estimator, and a maximum-likelihood fit shows the same split. An independence problem walks through why. As you lower the trials per stimulus the points spread out, but the calibration curve should stay near the diagonal: the estimates get less certain and report it.