This 1935 landmark paper by C. I. Bliss revolutionized toxicology by introducing the probit transformation. By converting sigmoid (S-shaped) dosage-mortality curves into straight lines, Bliss provided a mathematically rigorous way to calculate lethal dosages and their associated errors, a method that remains a cornerstone of bioassay analysis today.
TL;DR
Bliss introduces the "probit" (probability unit) to linearize the sigmoid relationship between toxic dosage and organism mortality. This allows researchers to apply standard linear regression and R.A. Fisher’s maximum likelihood methods to determine toxicological potency with high statistical precision.
What Problem Does This Paper Address?
In toxicological studies, plotting the percentage of organisms killed against the dose typically results in a sigmoid curve. These "S-curves" are notoriously difficult to analyze: they are non-linear, and the change in mortality per unit of dose is inconsistent (smallest at the extremes and largest near 50% kill).
Before this paper, biologists struggled to accurately estimate the "median lethal dose" (LD50) or compare the effectiveness of different poisons. Bliss sought to move beyond "fitting by eye" to a standardized statistical framework that accounts for the varying reliability of data points across the dosage range.
Methods
The core of the paper is a two-step transformation:
- Logarithmic Scaling: The dosage is converted to its logarithm. Bliss argues that susceptibility often follows a geometric rather than arithmetic distribution, aligning with the Weber-Fechner law.
- Probit Transformation: Mortality percentages are converted into "probits." A probit is essentially the normal deviate ($z$-score) of the cumulative normal distribution, shifted by +5 to eliminate negative numbers.

Once linearized, Bliss applies a weighted linear regression. Because a 50% kill rate is statistically more "stable" than a 0% or 100% kill rate, he assigns higher weights to observations near the median. He also includes a specialized appendix by R.A. Fisher to handle the "zero survivor" problem, allowing 100% mortality data to be incorporated into the model via maximum likelihood estimation.
Datasets
Bliss demonstrates his method using experimental data from Strand (1930). The dataset tracks the mortality of the adult flour beetle (Tribolium confusum) after five hours of exposure to varying concentrations of gaseous carbon disulphide ($CS_2$).
Key Insights and Ideas
- Susceptibility as a Distribution: Bliss treats the dosage-mortality curve as a cumulative normal frequency distribution. Each point on the curve represents the proportion of the population whose "individual lethal dose" is less than or equal to the applied dose.
- Linearization is Power: By turning the curve into a line ($Y = a + b(X - \bar{x})$), researchers can easily calculate the slope ($b$), which measures how uniformly a population responds to a toxin.
- The Weighting Coefficient: Bliss provides tables for a weighting coefficient ($z^2/pq$). This ensures that the regression line is influenced most by the most reliable data points, preventing outliers at the extreme ends of the scale from skewing the results.
- Handling the Extremes: The paper provides a rigorous way to use 0% and 100% kill data, which were previously often discarded or handled inconsistently.
Key Results
The paper provides the famous Table I, a "cheat sheet" for researchers to convert mortality percentages directly into probits without performing complex calculus.
Using the Tribolium data, Bliss shows that the probit-log(dose) plot yields a remarkably straight line. He calculates the regression equation and defines the "zone of error"—a hyperbolic confidence interval around the regression line that provides a 19-in-20 (95%) chance of enclosing the true mortality curve.
Limitations and Future Work
- Assumption of Normality: The method assumes that the distribution of susceptibility is normal (after log transformation). Bliss acknowledges that if the underlying distribution is skewed or follows a different law (like the $z$-distribution), the probit model might require adjustment.
- Biological "Breaks": Bliss notes that some datasets show a "break" or change in slope at low dosages. He suggests this might not be a mathematical artifact but a biological reality, such as "protective stupefaction" where low doses trigger a physiological defense that changes the organism's susceptibility.
- Future Directions: The author indicates that subsequent papers will apply these statistical methods to compare different sets of dosage-mortality data and analyze time-survival curves.
Summary updated