Statistics Glossary → Probit Regression
Probit Regressionanalysis
A regression model that links the probability of a binary outcome (died / survived, passed / failed) to predictors through the cumulative standard normal distribution, widely used in dose-response studies to estimate effective doses (ED50, LD50).
Concept
When the outcome is yes/no, fitting linear regression directly can produce probabilities below 0 or above 1. The probability is therefore linked to a linear combination of the predictors through an S-shaped curve. In the probit model this curve is the cumulative standard normal distribution (Φ): P(Y = 1) = Φ(b₀ + b₁x₁ + …). The model rests on the idea that each individual has an unobserved 'tolerance' threshold and that these thresholds are normally distributed: the response occurs when the dose exceeds the individual's threshold. This reading is especially natural in dose-response studies in toxicology and agriculture.
Probit coefficients are read on the latent scale, in z-score units: a positive coefficient raises the probability of the event, but by how much depends on the starting level of the probability. For this reason the average marginal effect (how many points, on average, a one-unit increase in the predictor changes the probability) is often reported instead of the coefficients. In dose-response analysis ED_p is the dose at which the event occurs with probability p; ED50 is the dose that produces a response in half the individuals, and its confidence limits are built with Fieller's method.
Probit and logit models give almost the same probabilities for most data; logit coefficients are roughly 1.6-1.8 times the probit ones. Logit is more common in general-purpose analyses because it offers the odds-ratio interpretation; probit is preferred where the assumption of a latent normal tolerance is meaningful and in the dose-response tradition. Model fit is assessed with the Hosmer-Lemeshow test and the classification table; overdispersion in grouped data should be checked separately, since it makes standard errors look too small.
When to use it
- When relating the probability of a binary outcome to predictors, especially in dose-response experiments where an effective or lethal dose (ED50, LD50) must be estimated and the latent normal threshold reading makes sense.
- If results are to be expressed as odds ratios, binary logistic regression is the more common and easily interpreted option.
- If the outcome has more than two ordered categories, use ordered probit or ordinal logistic models; if events are counted per unit of time or area, use Poisson regression.
An everyday example
A greenhouse grower wants to find the concentration at which a new natural spray is effective against aphids. At each of five concentrations 50 aphids are tested, and the dead are counted after 24 hours: 4 die at the lowest concentration and 47 at the highest. Probit regression fits an S-shaped curve between the logarithm of concentration and the probability of death and gives the concentration that kills half the insects (LD50) with a confidence interval. The grower can also work out the concentration needed for, say, 90% effect (LD90) and avoid using more spray than necessary.
Parameters in MerQur
Predictors (X) and Target (Y, binary)
Concept: The dependent variable is a 0/1 variable showing whether the event occurred; the predictors are variables thought to affect its probability. If the data are summarised in groups, the dependent variable can be the number of events in each group.
Everyday example: If each row is a single aphid, the dependent variable is 'died = 1, survived = 0' and the predictor is the spray concentration. If each row is a concentration group, the dependent variable is the number of dead aphids in that group.
What it does: Selects the binary (0/1) dependent variable and the predictors; if the trials column is selected, Y is the number of events. A Y with more than two values raises an error.
Effect on the result: All outputs.
On the validation data: UCLA admit ~ gre + gpa + rank: gre coefficient 0.0013756.
Categorical (factor) variables
Concept: Variables coded with numbers that merely label groups should be declared categorical. Each level then gets its own effect relative to a reference level; otherwise the model assumes a meaningless linear relation across the codes.
Everyday example: If the experiment ran in three greenhouses coded 1, 2 and 3, greenhouse should be declared categorical. Assuming greenhouse 3 has 'twice' the effect of greenhouse 1 makes no sense.
What it does: Makes the numerically coded columns written separated by commas categorical (dummies with the first level as the baseline level); if left empty, they stay numeric.
Effect on the result: Coefficients and AME.
On the validation data: rank -> the rank_2..4 dummies are created; when left empty, rank becomes a single numeric predictor.
Confidence Level (%90, %95, %99, Other (custom))
Concept: Sets how much assurance the confidence intervals for coefficients, marginal effects and effective-dose estimates carry. A higher level gives a wider interval.
Everyday example: If the LD50 lies between 2.1 and 2.8 mg/L with 95% confidence, asking for 99% widens this range and increases the uncertainty the grower must allow for when choosing a dose.
What it does: Sets the level of the Wald confidence interval (b +- z SE), the AME interval and the Fieller ED limits.
Effect on the result: Intervals only.
On the validation data: gre 95% [0.0001038, 0.002647]; 90% [0.0003083, 0.002443]; 99% [-0.000296, 0.003047]. Example 99.1 ED50 (log10 dose) 3.391: 95% [2.612, 4.271], 90% [2.749, 4.100].
Hosmer-Lemeshow
Concept: The Hosmer-Lemeshow test groups observations by predicted probability and compares the observed number of events in each group with what the model expects. A significant result shows that the model systematically mispredicts probabilities in some range.
Everyday example: The model expects about 20 deaths among 100 aphids with a predicted death probability of around 20%. If 35 actually died, the model underestimates the effect at low doses, and the test catches this.
What it does: Tests the agreement between observed and expected counts in 10 quantile groups of the predicted probabilities (Hosmer-Lemeshow test). It suits individual binary data; for grouped data the Pearson or deviance goodness-of-fit test is more appropriate.
Effect on the result: An additional table only.
On the validation data: UCLA binary: chi-square 12.609, df 8, p 0.126. In events/trials data there are 7 distinct probabilities, so MerQur drops the duplicate cut points and computes with 5 groups (df 4).
Event category (Larger value (default), Smaller value)
Concept: The event category determines which of the two outcomes has its probability modelled. Choosing the other category flips the signs of the coefficients; what the model says is unchanged.
Everyday example: With 'died' as the event, the concentration coefficient is positive: the probability of death rises with concentration. With 'survived' as the event, the same coefficient is negative: the probability of survival falls as concentration rises.
What it does: P(Y = larger value) or P(Y = smaller value) is modeled; the coefficient signs flip.
Effect on the result: Coefficient signs, sensitivity and specificity.
On the validation data: gre coefficient 0.0013756 and -0.0013756; sensitivity 0.2362 and 0.9304.
Classification cut-off
Concept: The classification cut-off sets the predicted probability above which an observation is classified as 'event'. Lowering it catches more of the real events (higher sensitivity) but also produces more false alarms (lower specificity).
Everyday example: With a cut-off of 0.5, aphids with a predicted death probability above 50% are counted as 'will die'. If every surviving aphid means a new colony for the grower, lowering the cut-off to 0.3 tracks the risk of survival more closely.
What it does: If the predicted probability is equal to or greater than this value, it is counted as an event; it determines accuracy, sensitivity, specificity and the 2x2 table.
Effect on the result: Classification outputs only.
On the validation data: UCLA binary accuracy: 0.610 for cut 0.3, 0.710 for 0.5, 0.685 for 0.7.
Extra output: AME (average marginal effect)
Concept: The average marginal effect gives, in probability units, how much a small increase in a predictor changes the event probability on average. Because a probit coefficient is on the z scale it is hard to interpret directly; the marginal effect conveys the same information in 'percentage points'. For categorical predictors the difference in probability between levels is computed.
Everyday example: An average marginal effect of 0.18 means a one-unit increase in log concentration raises the probability of death by 18 percentage points on average. A result such as 'the probability of death is on average 6 points lower in greenhouse 2 than in greenhouse 1' is the marginal effect of a categorical predictor.
What it does: Gives the average marginal effect (dy/dx) with the delta-method SE. For a categorical predictor the discrete change is computed (level - baseline level, other dummies 0).
Effect on the result: An additional table only.
On the validation data: For continuous predictors gre 0.000561, gpa 0.155; rank2 -0.1564, rank3 -0.2868, rank4 -0.3213.
Extra output: Classification table
Concept: The classification table compares the model's predictions at the chosen cut-off with the actual outcomes in four cells: correct events, correct non-events, false alarms and missed events. Accuracy, sensitivity and specificity are computed from this table.
Everyday example: If in a 250-aphid trial the model correctly predicts 105 of 120 deaths and 112 of 130 survivors, sensitivity is 105/120 ≈ 0.88 and specificity 112/130 ≈ 0.86.
What it does: Gives the 2x2 classification table according to the cut value.
Effect on the result: An additional table only.
On the validation data: At cut 0.5, tp 30, tn 254, fp 19, fn 97.
Extra output: Dose-response (ED/LD)
Concept: The effective dose (ED_p) is the dose level at which the event occurs with probability p; when the event is death it is called the lethal dose (LD_p). ED50 is the dose producing a response in half the individuals. Its confidence limits are computed with Fieller's method, which correctly reflects the uncertainty of this quantity, a ratio of two estimates.
Everyday example: If the model gives an LD50 of 2.4 mg/L and an LD90 of 5.1 mg/L, the grower knows that a 2.4 mg/L solution is needed to kill half the aphids and about 5.1 mg/L to kill nine in ten.
What it does: Gives ED_p = (Phi^-1(p) - a)/b (the first predictor is the dose, the others at their mean) and the Fieller confidence limits (z; no heterogeneity factor).
Effect on the result: An additional table only.
On the validation data: Example 99.1 dose data: ED50 (dose) 3.853 [3.100, 4.596]; with log10 dose 3.391 [2.612, 4.271].
Dose-response options: log10(dose)
Concept: The dose-response relation is often linear not in the dose itself but in its logarithm, because doubling the dose has a similar effect at low and high doses. A log dose also fits the common practice of increasing doses in multiples in experiments.
Everyday example: If the grower chose concentrations of 1, 2, 4, 8 and 16 mg/L, the log10 transform makes these doses equally spaced. If the probability of death rises by a similar amount with each doubling, the log scale improves the model markedly.
What it does: Takes the log10 of the first predictor (rows <= 0 are dropped with a warning).
Effect on the result: Dose scale and ED.
On the validation data: Example 99.1 log10(Dose) coefficient 3.4181 (SE 0.7455).
Dose-response options: Natural response
Concept: Natural response is the event occurring at some rate regardless of treatment (e.g. some individuals dying even in the untreated group). Abbott's correction adds this background rate to the model, so the effective dose is computed from the treatment-related response only.
Everyday example: If 5 of 50 aphids in an unsprayed control group also die within 24 hours from natural causes, this 10% background mortality should not be credited to the spray. The natural-response option separates it out and gives the LD50 based on the spray's real effect.
What it does: Abbott correction P = C + (1 - C) Phi(xb); C is estimated by maximum likelihood; ED is computed from this model.
Effect on the result: Coefficients, C and the ED table.
On the validation data: Example 99.4: C 0.27347 (SE 0.09459), log-likelihood -387.2467. In the Example 99.1 log10 model, C 0.1186. In the binary UCLA data C is at the boundary (about 0) and a boundary warning is given.
Number of trials (n) column
Concept: If the data are in events/trials form, each row represents a group: the dependent variable is the number of events and the trials column the number of individuals in that group. The model uses these counts to work with each group's event rate.
Everyday example: There is a single row per concentration: '2 mg/L, 50 aphids, 18 dead'. The trials column holds 50 and the dependent variable 18, so there is no need to write 250 separate rows for 250 aphids.
What it does: If selected, the data are in events/trials form: Y is the number of responders and this column is the number of trials; the data are expanded to Bernoulli rows (BIC with this n).
Effect on the result: n and BIC in grouped data.
On the validation data: Example 99.1: 7 rows, 74 trials.
Overdispersion scale (None (scale = 1), Pearson, Deviance)
Concept: Overdispersion is variability in grouped data greater than the model's binomial assumption predicts; it arises, for example, when individuals in the same group are not independent. If ignored, standard errors come out too small and confidence intervals too narrow; a scale correction inflates the standard errors accordingly.
Everyday example: If the 50 aphids at one concentration were all on a single leaf and that leaf held the spray less well, the deaths are not independent. The fluctuation between groups turns out larger than expected, and the Pearson scale correction widens the LD50's confidence interval to match.
What it does: Scales the standard error by sqrt(phi) (phi = Pearson chi-square or deviance divided by degrees of freedom); in events/trials form, phi is computed from the grouped rows. The scale applies only to the coefficient table (LR, Type III, AME and ED are not scaled).
Effect on the result: Coefficient SE, z, p and intervals.
On the validation data: Example 99.1 dose data: phi_Pearson 3.483280904, phi_deviance 4.049526004.
Frequency weight
Concept: A frequency weight shows how many observations with identical values a row represents. The analysis treats the row as if it were repeated that many times; this differs from a sampling weight and changes the actual number of observations.
Everyday example: If the data are summarised as '2 mg/L, died, 18' and '2 mg/L, survived, 32', the frequency weights are 18 and 32. The model treats these two rows as 50 separate aphids.
What it does: Counts each row as an observation repeated as many times as its frequency.
Effect on the result: Coefficients, n and SE.
On the validation data: With frequency-weighted data, the coefficients and n equal those of the expanded data in which each row is repeated as many times as its frequency.
Link (distribution) (Probit (normal), Logit (logistic))
Concept: The link function sets the S-curve that turns the linear combination into a probability. Probit uses the normal distribution and logit the logistic distribution; the two curves are very similar in the middle and differ most in the tails, the logistic distribution having slightly heavier tails.
Everyday example: When both probit and logit are fitted to the same aphid data, the LD50 values usually come out very close. The difference shows up mostly at extreme doses such as LD99; if a very high effect is the goal, the choice of link can affect the dose decision.
What it does: Switches the link in the probit form to logit; the marginal effect, AME, natural response and Fieller ED/LD use the distribution of the selected link.
Effect on the result: Coefficients and ED.
On the validation data: In the log10 dose-response example, when the logit link is selected the coefficients, SEs and ED10/50/90 are computed from the logistic distribution.
Example on real data
Data: UCLA IDRE binary (graduate admission)
Variables: target admit (admitted 0/1), predictors gre, gpa and rank_t (school rank, categorical)
- Question: How is the probability of admission to a graduate program related to the GRE score, the grade point average (GPA) and the rank of the undergraduate school (r1-r4)? (Binary dependent variable; probit link.)
- Model test: LR χ²(5) = 41.56, p < .001; the predictors jointly explain the admission probability significantly.
- Result: GRE (b = 0.0014, z = 2.12, p = .034) and GPA (b = 0.478, z = 2.44, p = .015) increase the admission probability; compared with r1, applicants from r2 (b = -0.415, p = .034), r3 (b = -0.812, p < .001) and r4 (b = -0.936, p < .001) schools have a lower admission probability.
- Effect size / fit: McFadden R² = .083 (weak explanatory power); AUC = 0.693, at cut 0.5 accuracy 71%, sensitivity .236, specificity .930: the model is weak at catching admitted applicants.
- Additional finding: Marginal effect at mean X: a 1-point increase in GPA raises the admission probability by about 0.167; an r4 school lowers the probability by 0.327 compared with r1.
- APA: Probit regression, LR χ²(5) = 41.56, p < .001, McFadden R² = .08; GPA b = 0.48, z = 2.44, p = .015, N = 400.
Related concepts
Effective dose (ED50, LD50) and dose-responseAverage marginal effect (AME)Abbott correction (natural response)Hosmer-Lemeshow goodness-of-fit testLink functionOverdispersion and the scale parameterClassification table (confusion matrix)Frequency weight (summary-table data)
The numbers in the examples are MerQur's output on independent, publicly available validation data; see the MerQur Technical Note for details.