probit analysis

flo_kussener

Hey,

I have the following table :

Dose is an amount, Test is the number of test and detected the number of detected test. Ratio is equal at detected/Test.

So it is clear that all tests are detected before a Dose of 2.

If I do a probit model:

I have a model which is giving me the following prediction (which seems not so good...):

so if I ask for a 0.95 prediction, I have the following result:

Inverse Prediction

Specified Probability(ratio)

Predicted Dose

Lower 95%

Upper 95%

Type of CI

0,9500000

6,631872

-5,76180

19,02554

Wald

But clearly, a prediction of 6.6 seems really far away of what the data seems to say us.

What did I do wrong???

Thank for your suggestion,

Florence

image002.png

image004.png

image003.png

image005.png

image001.png

malcolm_moore1

Is this real data Florence as it is unusual to see a zero dose being tested? Dose-Response experiments are often done on a log scale so it would be more typical to see a dose column e.g. 0.01, 0.1, 1, 10, 100 (or smaller increments on a log scale) and then transformation to log(dose) in the analysis which aids relative potency estimation when two or more compounds are tested, it also prevents negative estimates when using inverse prediction.

Given that context, best action in this case would be to redo the experiment with better doses as the estimate of dose required to achieve 95% response rate is 6.6 with 95% confidence interval -5.8 to 19.0. We really need some doses to be tested that do not result in all failures other than the zero dose - and the zero dose should not have been used in any case as it prevents use of log(dose) which in return gives negative inverse prediction intervals.

Malcolm