Bioassay — Quantal Response Method

Bioassay — Quantal Response Method analyzes all-or-none outcomes, such as died or survived, from the number responding out of the number tested at each dose. It fits a logit, probit, gompit or log-log model on log dose with a common slope, tests regression, parallelism and linearity with chi-square tests, and reports effective doses and relative potency with confidence limits. Bioassay — Quantal Response Method is one of the bioassay tools in Zometric Statistical Software, run from the Statistical Workspace or its own tool page.

Last reviewed

Key takeaway

The quantal response method fits a binary dose-response model on log dose with a common slope, from the number of subjects responding at each dose, to estimate effective doses such as ED50 and the relative potency of test preparations.

On this page
  1. What is Bioassay — Quantal Response Method?
  2. When to use it
  3. Data you need
  4. How to run Bioassay — Quantal Response Method in Zometric Statistical Software
    1. In the Statistical Workspace
    2. On the tool page
  5. Options
  6. Output
  7. Worked example

What is Bioassay — Quantal Response Method?

A quantal response assay measures an all-or-none outcome in each subject, such as died or survived, or responded or did not respond. At each dose, you count how many of the subjects tested responded. The tool estimates the effective dose, for example the ED50, which is the dose that makes 50% of subjects respond. It also estimates the relative potency of one or more test preparations against a standard.

The model is a generalized linear model for binomial counts. For a preparation, the probability p of a response at a dose is described on the log-dose scale:

g(p) = a + b × log(dose)

Here g is the link function: logit, probit, gompit (complementary log-log) or log-log. The slope b is common to all preparations. Each preparation has its own intercept a. The model is fitted by maximum likelihood. The effective dose for a response proportion p is:

ED(p) = dose at which g(p) = a + b × log(dose), that is log ED = (g(p) − a) / b

The relative potency of a test preparation is the ratio of the standard's ED50 to the test preparation's ED50, which equals exp((a(test) − a(standard)) / b) on the natural log scale. The confidence limits for effective doses and potency come from Fieller's theorem, and the tool reports Fieller's g. A g below 1 means the limits are reliable.

By default, validity is checked with likelihood-ratio chi-square tests, and you can switch to Finney's classical minimum chi-square instead. The tests are the quantal counterparts of the validity checks in the Bioassay — Parallel Line Method: significance of the regression, non-parallelism and non-linearity. The approach follows the statistical analysis of biological assays in the European Pharmacopoeia (chapter 5.3) and in the United States Pharmacopeia (general chapter on the design and analysis of biological assays).

When to use it

Use it when:

  • Each subject either responds or does not, and you know how many responded out of how many tested at each dose.
  • You want an effective dose such as ED50 or LD50, with confidence limits.
  • You want to compare a test preparation with a standard, in a toxicology, vaccine or other potency test.

When not to use it, and what to use instead:

Data you need

Use one row per dose group of each preparation. The first (grey) row holds the column names.

  • A numeric column with the number of subjects that responded. It must be between 0 and the number tested.
  • A numeric column with the number of subjects tested in that group.
  • A numeric column with the dose. Doses must be strictly positive, because the tool takes their logarithm.
  • A column that names the preparation of each row, for example Standard and TestA. The text for the standard must be the same in every row.

Include the standard and at least one test preparation. Use at least 3 to 4 dose levels per preparation. The doses should span a good part of the response curve, ideally from a low response (near 0%) to a high response (near 100%).

Number Responding

Number of Subjects

Dose (mg/kg)

Preparation

5

40

2

Standard

21

40

4

Standard

32

40

8

Standard

9

40

2

TestA

…

…

…

…

How to run Bioassay — Quantal Response Method in Zometric Statistical Software

Sign in to Zometric Statistical Software. You can run Bioassay — Quantal Response Method in the Statistical Workspace, where most people work, or on its own tool page.

In the Statistical Workspace

  1. Open Statistical Workspace.
  2. In the Data grid, type the column names in the first (grey) row, then type or paste (Ctrl+V) your data below. Put the number responding, the number of subjects, the dose and the preparation name in four columns, with one row per dose group.
  3. Click + Add Tool, search for Bioassay — Quantal Response Method and open it. It is listed under Bioassay. You can also find it under Browse › Regression.
  4. In Options: Bioassay — Quantal Response Method, choose your columns and settings (see Options below).
  5. Click Calculate. The results appear above the data.
  6. To keep them, use Download › (PDF, Word or Excel), or Export › to put several analyses into one file.

Only columns with a value in the first data row are sent to the tool, so don't leave the first cell of a column blank.

Add Tool search in the Statistical Workspace showing Bioassay — Quantal Response Method in the Bioassay category

On the tool page

  1. Open Statistical Tools and click Bioassay — Quantal Response Method (category: Bioassay).
  2. Enter data in the Data Pane the same way, use Load File to open an Excel or CSV file, or pick a dataset from — Load Example —, which also fills in the options. The example is listed as toxicology-assay.
  3. Set the Options and click Calculate.
  4. Use Download › to save the results as PDF, Word or Excel.
Bioassay — Quantal Response Method tool page with the toxicology data in the Data Pane and Number Responding, Number of Subjects, Dose (mg/kg), Preparation, Standard and Logit selected

No account yet? Talk to us for a demo or a trial.

Options

The tool page and the Workspace show the same 22 options, in this order.

Columns and model

  • Response: the column with the number of subjects that responded in each group. Required.
  • Subject: the column with the total number of subjects in each group. Required. The response must be between 0 and this number.
  • Dose: the column with the dose. Required. Values must be positive.
  • Preparation: the column that says which preparation each row belongs to. Required.
  • Standard preparation label: the exact text used for the standard in the Preparation column. Required. Default Standard. It must match exactly.
  • Link function: Logit (default), Probit, Gompit (complementary log-log) or Log-log. It sets how the response proportion relates to log dose. Logit and probit are the usual choices and give similar answers.
  • Logarithm base: Natural log (ln) (default) or Log base 10. It is applied to the dose. The potency and the effective doses do not depend on the base.
  • Effective dose percentiles: the response percentages to report effective doses for, separated by commas, for example 10, 50, 90 for ED10, ED50 and ED90. Default 50. Each value must be strictly between 0 and 100.
  • Also compute ED50 via Spearman-Karber method: unticked by default. Adds a distribution-free estimate of the ED50 next to the model-based one. It is a cross-check, and it needs responses that run from about 0% to about 100%.

Potency

  • Assumed standard potency: the nominal (labeled) potency of the standard. Default 100.
  • Assumed test preparation potency: the nominal potency shared by all test preparations, unless you override it below. Default 100.
  • Per-preparation potency overrides (optional): nominal potencies for particular test preparations, written as TestA=110, TestB=95. They replace the shared default for those preparations.

Confidence and compliance

  • Confidence level (%): the level for the confidence limits of the effective doses and the potency. Default 95. Allowed values are 50 to 99.999.
  • Compliance lower limit (%): the lowest acceptable relative potency. Default 75.
  • Compliance upper limit (%): the highest acceptable relative potency. Default 125. The lower limit must be less than the upper limit.

Validity tests and fitting

  • Validity test method: how the regression, non-parallelism and non-linearity tests are computed. The potency, the effective doses, their confidence limits and the regression coefficients are the same under both choices.
    • Likelihood-ratio deviance (default) fits nested models by maximum likelihood, from one intercept per preparation up to a saturated model with one parameter per dose group. Each test is a deviance difference, −2 × (log-likelihood of the simpler model − log-likelihood of the larger model), compared with a chi-square distribution.
    • Finney's classical minimum chi-square is the classical pharmacopoeial method. It starts from the common-slope fit and, for each dose group, computes a working logit or probit: the fitted linear predictor plus (observed proportion − fitted proportion) / (slope of the link curve at that point). A weighted least-squares analysis of these values, with weight = group size × (slope of the link curve)² / (fitted proportion × (1 − fitted proportion)), gives the chi-square for each test. It never fits a separate parameter for a single dose group, so it stays stable when a group has a 0% or 100% observed response.
  • Significance level for validity tests (α): the threshold for the validity tests. Default 0.05. Allowed values are 0.001 to 0.5.
  • Maximum iterations (MLE convergence): the most iterations allowed for the maximum likelihood fit. Default 100. Allowed values are 10 to 1000.

Output sections

  • Show case diagnostics table: ticked by default. Lists each dose group with its observed and estimated proportion, standardized residual and outlier flag.
  • Show dose-response graph: ticked by default.
  • Show bar plot of proportion responding by dose-preparation group: unticked by default.
  • Show normal probability plot of residuals: unticked by default.
Options with Number Responding, Number of Subjects, Dose (mg/kg), Preparation, Standard preparation label Standard, Link function Logit and Logarithm base Natural log (ln)

Output

The results appear in this order.

  1. Summary Statistics: for each preparation and dose, the number of subjects (N Subjects), the number responding (N Responding) and the Proportion Responding.
  2. Model Summary: the Link Function, the Deviance, the Residual DF, the Log-Likelihood and whether the fit Converged. A small deviance relative to the residual DF means the model fits the group proportions well.
  3. Regression Coefficients: for each preparation the Intercept, the Common Slope and the equation, for example logit(p) = intercept + slope × log(Dose).
  4. Validity of Assay — Likelihood-Ratio Chi-Square Tests: Significance of Regression, Non-parallelism and Non-linearity, with DF, Chi-Square and P-Value. With Finney's classical minimum chi-square, this table is titled Validity of Assay — Finney's Classical Minimum Chi-Square, adds a Preparations row, and is followed by Non-linearity — Breakdown by Preparation, one non-linearity chi-square for each preparation (a preparation with fewer than 3 doses gets a note instead).
    • Notes / Warnings appears with the deviance method when a dose group has exactly 0% or 100% responding. It names the group, for example Standard/2, and warns that the regression and non-linearity chi-square values may be unstable or inflated. Switch to Finney's method in that case. The note does not appear with Finney's method.
  5. Validity Tests Summary: one line for each test with a plain-language result. All three should read as good: a significant regression, and a non-significant non-parallelism and non-linearity.
  6. Effective Dose (ED) Results: for each preparation and each percentile you asked for, the Effective Dose with Lower CL, Upper CL and CI Reliable (G<1). With the Spearman-Karber option ticked, its ED50 is shown too.
  7. Potency Results: for each test preparation the Estimated Potency, Lower CL, Upper CL, Relative Potency (%), % Precision (the width of the confidence interval as a percentage of the lower limit), Fieller g, Compliance (Pass or Fail against your limits) and CI Reliable (G<1).
  8. Case Diagnostics, if ticked: for each dose group the observed and Estimated Proportion, the Standardized Residual and an Outlier flag for an absolute value above 2.
  9. Dose-Response Plot: the observed and fitted values on the link scale (for example the logit of the proportion) against log dose, with one line for each preparation. The lines are parallel. Any other plot you ticked follows.

Check the three validity tests first. A significant non-parallelism means the preparations do not behave as dilutions of each other, so a relative potency is not meaningful. A significant non-linearity means the model does not describe the dose-response on this scale, and another link function or dose range may fit better. Then read the effective doses and the potency. Compliance checks only that the estimated potency lies between your two limits. It does not look at the confidence limits. A wide potency interval means the design had too little information, and more subjects or more doses would help.

Worked example

This example uses the tool's own example dataset, listed as toxicology-assay in — Load Example —. On the tool page, pick it and click Calculate. The example fills in Response = Number Responding, Subject = Number of Subjects, Dose = Dose (mg/kg), Preparation = Preparation, Standard preparation label = Standard, Link function = Logit and Effective dose percentiles = 50. All other options keep their defaults.

A toxicology lab compares a test compound (TestA) with a Standard in an all-or-none assay. It doses 40 subjects at each of 2, 4, 8 and 16 mg/kg for each preparation and counts how many respond. What are the ED50 values, and what is the relative potency?

Data summary.

Summary statistics: Standard responding 5, 21, 32 and 35 of 40 at 2, 4, 8 and 16 mg/kg; TestA responding 9, 20, 32 and 37 of 40

Preparation

Dose (mg/kg)

N Subjects

N Responding

Proportion Responding

Standard

2

40

5

0.125

Standard

4

40

21

0.525

Standard

8

40

32

0.8

Standard

16

40

35

0.875

TestA

2

40

9

0.225

TestA

4

40

20

0.5

TestA

8

40

32

0.8

TestA

16

40

37

0.925

The proportion responding rises from 0.125 to 0.875 for the Standard and from 0.225 to 0.925 for TestA. The doses cover a wide part of the curve.

Model and validity.

Validity chi-square tests: regression 112.82 (p below 0.001), non-parallelism 0.0282 (p 0.8667), non-linearity 3.2374 (p 0.5189)

Source

DF

Chi-Square

P-Value

Significance of Regression

1

112.82

<0.001

Non-parallelism

1

0.0282

0.8667

Non-linearity

4

3.2374

0.5189

The model summary shows a deviance of 3.2655 on 5 residual degrees of freedom (8 groups minus 3 parameters) and a converged fit. With a common slope of 1.881 on ln(dose), the intercepts are -2.7703 for the Standard and -2.5789 for TestA. The regression is significant (chi-square 112.82, P < 0.001). The non-parallelism test is not significant (chi-square 0.0282, P = 0.8667), so a common slope fits. The non-linearity test is not significant either (chi-square 3.2374, P = 0.5189). The assay is valid.

Effective doses.

ED50 results: Standard 4.3613 mg/kg (95% limits 3.5177 to 5.3506), TestA 3.9394 mg/kg (3.1616 to 4.8362)

Preparation

Percentile

Effective Dose

Lower CL

Upper CL

CI Reliable (G<1)

Standard

ED50.0

4.3613

3.5177

5.3506

Yes

TestA

ED50.0

3.9394

3.1616

4.8362

Yes

The ED50 of the Standard is 4.3613 mg/kg (95% limits 3.5177 to 5.3506). The ED50 of TestA is 3.9394 mg/kg (3.1616 to 4.8362). It takes a lower dose of TestA to get a 50% response, so TestA is the more potent one.

Potency.

Potency results for TestA: relative potency 110.71%, 95% limits 82.53 to 149.154, % precision 80.7268, Fieller g 0.0493, compliance Pass

Preparation

Estimated Potency

Lower CL

Upper CL

Relative Potency (%)

% Precision

Fieller G

Compliance (75.0–125.0%)

CI Reliable (G<1)

TestA

110.71

82.53

149.154

110.71

80.7268

0.0493

Pass

Yes

The relative potency of TestA is 110.71%. As a check, 4.3613 / 3.9394 = 1.1071. The 95% limits are 82.53% to 149.154%. Fieller g is 0.0493, so the limits are reliable. The result is Pass against 75% to 125%, but the interval is wide (% precision 80.7268). With 40 subjects per dose, the true potency could be anywhere from about 83% to 149%. A tighter limit would need more subjects.

Dose-response plot.

Dose-response plot of logit of proportion responding against log(Dose) with two parallel fitted lines for Standard and TestA

The TestA line lies slightly above the Standard line, which is what a potency above 100% looks like. The points follow the lines closely. The largest standardized residuals in the case diagnostics are -1.0547 and -1.0125, both for the Standard, so no group is an outlier.

Try this.

  • Change Link function to Probit. In the app, the potency is 111.8% (limits 83.7% to 150.1%), almost the same.
  • Tick Also compute ED50 via Spearman-Karber method. In the app, the page warns that the responses do not run from about 0% to about 100%, so that estimate is not reliable for this data.
  • Change Validity test method to Finney's classical minimum chi-square. The potency stays 110.71% with the same limits. The regression chi-square is 77.8418 (P < 0.001), the non-parallelism chi-square stays 0.0282 (P = 0.8667) and the non-linearity chi-square is 3.2682 (P = 0.514), so the assay is still valid. The breakdown gives a non-linearity chi-square of 3.2027 (P = 0.2016) for the Standard and 0.0654 (P = 0.9678) for TestA.
  • Change Effective dose percentiles to 10, 50, 90. The table then lists ED10, ED50 and ED90 for each preparation. Expect wider limits at ED10 and ED90 than at the ED50, because they lie further from the middle of the data.

Frequently asked questions

What is a quantal response bioassay?

It is an assay where each subject's outcome is all-or-none, such as died or survived. Each dose group reports the number responding out of the number tested. A binomial model on log dose with a common slope gives effective doses such as ED50 and a relative potency.

What is the ED50?

It is the dose expected to make 50% of the subjects respond. You can ask for other percentiles, for example ED10 and ED90, in Effective dose percentiles. Each comes with confidence limits. Doses far from the middle of the curve are less well determined, so their limits are wider.

Which link function should I use?

Logit and probit are the most common choices and usually agree closely. Gompit (complementary log-log) and log-log are asymmetric, and suit responses that approach 0% and 100% at different speeds. If you are unsure, compare the fit: look at the deviance in the model summary and at the non-linearity test.

Which validity test method should I choose?

Likelihood-ratio deviance is the default. Finney's classical minimum chi-square is the classical pharmacopoeial method, and it is more stable when a dose group has 0% or 100% response. Use it when your data has such groups, or when your procedure requires it. With the deviance method, a Notes / Warnings table tells you when a group is at 0% or 100%. The choice does not change the potency or its confidence limits.

What do I do if the non-parallelism test is significant?

A significant test means the dose-response curves of the preparations have different slopes, so a single relative potency does not describe them. Check the data for errors, try another link function, or investigate whether the test preparation really behaves like the standard.

Can the response be a proportion instead of a count?

No. The tool needs the number that responded and the number tested in each group, so it can weight each group by its size. Enter the counts, not the percentages.