MerQuR+Statistics Glossary

Statistics Glossary → Effective dose (ED50, LD50) and dose-response

Effective dose (ED50, LD50) and dose-responseconcept

In a dose-response relationship, the dose estimated to produce a response in a given proportion (e.g. 50%) of subjects; when the response is death it is called the lethal dose (LD).

In dose-response experiments, different doses are given to different groups of subjects and the proportion responding (e.g. recovering or dying) in each group is recorded. The S-shaped relationship between this proportion and the dose (often the log of the dose) is usually described with a probit or logit model: in the probit model, Φ⁻¹(p) = a + b·dose. Once the model is fitted, the dose giving a desired response proportion p is ED_p = (Φ⁻¹(p) - a) / b. ED50 is the dose that produces the effect in half the subjects and LD50 the dose that kills half of them; since Φ⁻¹(0.5) = 0, ED50 = -a / b.

Because ED_p is a ratio of two coefficients, Fieller's method is used for its standard error and confidence interval; this method can give asymmetric and even unbounded intervals when the uncertainty in the slope b is large. If the dose was modelled on a log scale, the result is back-transformed to the original dose units. Doses at extreme proportions (ED10, ED90) are estimated much less precisely than ED50, because the data usually concentrate in the middle range. If there is a natural response (one seen in the control group too) or overdispersion, the model must be adjusted accordingly.

An everyday example

A crop pesticide is being tested: 50 insects are exposed at each of 5 doses, and the proportion killed rises from 4% at the lowest dose to 96% at the highest. The probit curve fitted to these points might give the dose that kills half the insects (LD50) as, say, 2.4 mg per litre. To decide the dose to recommend to farmers, however, higher proportions such as LD90 are usually of more interest.

In MerQur

In Probit Regression, the 'Dose-response (ED/LD)' extra output gives ED_p with Fieller confidence limits, and 'Dose-response options: log10(dose)' models the dose on a logarithmic scale.

Analyses where this concept appears

Probit Regression

Related concepts

Abbott correction (natural response)Link function