Written and maintained by CASRAI Editorial Board
Published · Last updated
On this page
The glm function in R is the foundational routine for estimating generalized linear models (GLMs) across quantitative empirical research. While classical ordinary least squares via R’s lm() assumes a continuous, normally distributed response with constant variance, empirical investigations routinely encounter binary outcomes (clinical trial remissions, program completions), discrete event counts (hospital admissions, citations, defects), rates, and strictly positive skewed durations. The glm() function, part of the standard stats package in R, unifies these non-normal response distributions within the mathematical framework formulated by John Nelder and Robert Wedderburn (1972) and expanded by Peter McCullagh and John Nelder (1989).
By connecting an exponential family error distribution to a systematic linear predictor through a monotonic link function, glm() enables researchers to model non-linear phenomena while retaining the inferential properties and interpretability of linear regression. Mastering glm()—from specifying appropriate link functions and offset terms to interpreting deviance residuals and troubleshooting numerical separation—is critical for rigorous quantitative research in biostatistics, health economics, sociology, and computational biology.
Ask CASRAI · free to tryQuestion about this guide? Ask CASRAI answers with linked sources. 2 free questions a day, no account.
What it is used for in research
In quantitative academic and institutional research, glm() serves as the primary computational instrument when the foundational assumptions of ordinary linear regression fail. Standard ordinary least squares (OLS) regression requires that error terms follow a normal distribution with homogeneous variance (homoscedasticity) and that the expected relationship between covariates and the outcome is strictly additive on the measurement scale. For an overview of traditional Gaussian modeling, see our guide on linear regression and multiple regression.
When response variables are bounded—such as binary probabilities restricted to the [0, 1] interval, or count data strictly restricted to non-negative integers—OLS models frequently produce invalid out-of-bounds predictions and distorted standard errors. Generalized linear models resolve these issues through three structural components:
- The Random Component: The response variable Y follows a probability distribution from the exponential dispersion family, which includes Gaussian, Binomial, Poisson, Gamma, and Inverse Gaussian distributions. Its conditional distribution has expected mean μ = E(Y) and variance Var(Y) = φ · V(μ), where V(μ) is the variance function and φ is the dispersion parameter.
- The Systematic Component: A linear combination of explanatory covariates forming the linear predictor η = Xβ, where X is the design matrix and β is the vector of unknown regression parameters.
- The Link Function: A monotonic, differentiable function g(·) that maps the expected mean μ to the linear predictor scale: g(μ) = η = Xβ. The inverse link function g⁻¹(η) = μ transforms linear predictor values back to the natural scale of the observed response.
Across empirical research disciplines, glm() is applied to diverse analytical challenges:
- Epidemiology and Clinical Trials: Estimating adjusted odds ratios (OR) for binary patient outcomes (treatment success vs. failure, disease recurrence) via logistic regression (
family = binomial(link = "logit")) or relative risks via log-binomial models. - Health Services and Public Health: Analyzing patient visits, hospital readmission frequencies, or symptom incidence using Poisson regression (
family = poisson(link = "log")) or quasi-Poisson models to accommodate overdispersion. For detailed count data methodology, see our guide on Poisson regression analysis. - Health Economics and Cost Studies: Modeling highly skewed, strictly positive expenditure data and length-of-stay metrics using Gamma regression with a log link (
family = Gamma(link = "log")), avoiding retransformation bias associated with log-transformed ordinary linear models. - Bioassay and Toxicology: Calculating median lethal doses (LD50) and concentration-response curves through probit regression (
family = binomial(link = "probit")), which maps probabilities via the cumulative standard normal distribution. - Demography and Actuarial Science: Modeling cohort survival proportions and hazard rates using the complementary log-log link (
family = binomial(link = "cloglog")), which directly corresponds to grouped continuous-time proportional hazards.
The table below summarizes the exponential dispersion families, canonical links, alternative link functions, variance structures, and standard research applications supported natively in base R via stats::family:
| Distribution Family | Default (Canonical) Link | Alternative Supported Links | Variance Function V(μ) | Standard Research Applications |
|---|---|---|---|---|
gaussian |
identity [μ] |
log, inverse |
1 | Continuous responses; verifies numerical equivalence with OLS while allowing non-identity links. |
binomial |
logit [ln(μ / (1 – μ))] |
probit, cauchit, log, cloglog |
μ(1 – μ) | Binary outcomes, success/failure proportions, case-control studies, dose-response bioassays. |
poisson |
log [ln(μ)] |
identity, sqrt |
μ | Discrete counts, incidence rates over time or space, log-linear contingency table analysis. |
Gamma |
inverse [1 / μ] |
identity, log |
μ² | Positive continuous right-skewed variables (clinical costs, physical waiting times, reaction times). |
inverse.gaussian |
1/mu^2 [1 / μ²] |
inverse, identity, log |
μ³ | First-passage times, highly skewed physical durations, accelerated life testing. |
quasi |
identity |
User-specified link & variance | Custom V(μ) | Semi-parametric quasi-likelihood estimation specifying only mean-variance relationships. |
quasibinomial |
logit |
probit, cloglog, identity |
φ · μ(1 – μ) | Overdispersed binomial proportions where residual variance exceeds nominal binomial variance. |
quasipoisson |
log |
identity, sqrt |
φ · μ | Overdispersed count data where sample variance exceeds conditional mean (dispersion φ > 1). |
How to install or access it
The glm() function is a core component of the stats package, which is built directly into every standard distribution of R maintained by the R Core Team. Because stats is part of base R and attaches automatically whenever an R interactive session or script initializes, researchers do not need to install any external package from CRAN or call library(stats) before using glm().
To verify the installation and access the official documentation inside your active R environment, run the following commands in the R terminal:
# Inspect active R version and environment details
R.version.string
# Confirm stats package namespace contains glm
find("glm")
# Expected output: [1] "package:stats"
# Access the official R help manual page for glm
?stats::glm
help(glm)
When specialized model families or diagnostic extensions are needed—such as penalized lasso/ridge regression via glmnet, negative binomial count modeling via MASS::glm.nb, or mixed-effects GLMs via lme4—researchers can install those packages from CRAN using standard library commands. For guidance on package installation, see our article on how to install libraries in R.
Step-by-step instructions
The following walkthrough provides a complete, reproducible step-by-step task fitting a generalized linear model in R. We analyze vehicle transmission type (0 = automatic, 1 = manual) from the standard built-in mtcars dataset using binary logistic regression, calculate adjusted odds ratios with confidence intervals, and generate response-scale predictions with valid error bounds.
Step 1: Inspect the dataset and verify variable types
Load the built-in dataset and review summary statistics. Ensure the binary dependent variable am is numeric (0/1) or a two-level factor, and check the candidate predictors wt (weight in 1,000 lbs) and hp (gross horsepower):
# Load built-in data and inspect variable structure
data("mtcars")
str(mtcars[, c("am", "wt", "hp")])
table(mtcars$am)
Step 2: Execute glm() with formula and family specification
Fit the logistic regression model by supplying the symbolic formula am ~ wt + hp, setting data = mtcars, and specifying family = binomial(link = "logit"). The fitting method uses Iteratively Reweighted Least Squares (IWLS):
# Fit binary logistic regression model
logit_model <- glm(
formula = am ~ wt + hp,
data = mtcars,
family = binomial(link = "logit")
)
# Verify fitted object classes
class(logit_model)
# Expected output: [1] "glm" "lm"
Step 3: Extract and review the complete model summary
Call the standard extractor method summary() to inspect maximum likelihood parameter estimates, standard errors, Wald z-statistics, and deviance goodness-of-fit metrics:
# Display the complete model summary
model_summary <- summary(logit_model)
print(model_summary)
Step 4: Compute adjusted odds ratios and 95% profile likelihood confidence intervals
Because the logit link models the log-odds of the outcome, ln(μ / (1 – μ)) = β₀ + β₁X₁ + β₂X₂, raw coefficient estimates reflect additive changes on the log-odds scale. Exponentiate the coefficients using exp(coef()) to obtain odds ratios (OR). Compute profile likelihood confidence intervals via confint(), which are more reliable than Wald intervals in finite samples:
# Exponentiate coefficients to calculate Odds Ratios (OR)
point_estimates <- coef(logit_model)
odds_ratios <- exp(point_estimates)
# Calculate 95% profile likelihood confidence intervals
ci_profile <- exp(confint(logit_model))
# Combine into a clear results table
or_table <- cbind(
Estimate_LogOdds = point_estimates,
Odds_Ratio = odds_ratios,
CI_2.5 = ci_profile[, 1],
CI_97.5 = ci_profile[, 2]
)
print(round(or_table, 4))
Step 5: Generate predicted probabilities and confidence bounds on the response scale
When applying a fitted GLM to new data via predict(), the type parameter controls the scale of estimation:
type = "link"(default): Produces predictions on the linear predictor scale η = Xβ (log-odds for logistic regression).type = "response": Applies the inverse link function g⁻¹(η) to produce predictions on the natural response scale μ (probabilities between 0 and 1).
To avoid generating confidence intervals that exceed the valid [0, 1] probability range, calculate upper and lower bounds on the linear predictor scale where errors are approximately normal, then transform the bounds using the model’s inverse link function (family(model)$linkinv):
# Construct hypothetical vehicle profiles
new_data <- data.frame(
wt = c(2.2, 3.2, 4.2),
hp = c(110, 150, 200)
)
# Predict on the linear predictor scale with standard errors
pred_link <- predict(logit_model, newdata = new_data, type = "link", se.fit = TRUE)
# Calculate 95% critical thresholds on the link scale
z_crit <- qnorm(0.975)
link_lower <- pred_link$fit - z_crit * pred_link$se.fit
link_upper <- pred_link$fit + z_crit * pred_link$se.fit
# Transform estimates and bounds to probabilities via the inverse link
inv_link <- family(logit_model)$linkinv
predicted_probabilities <- data.frame(
wt = new_data$wt,
hp = new_data$hp,
pred_prob = inv_link(pred_link$fit),
ci_lower = inv_link(link_lower),
ci_upper = inv_link(link_upper)
)
print(round(predicted_probabilities, 4))
Specialized Response Formats in glm()
The glm() function accommodates three distinct response specifications for binomial models and rate modeling:
- Individual Binary Vector: A numeric vector of 0s and 1s, or a two-level factor variable (the first factor level denotes failure, the second denotes success).
- Two-Column Matrix of Successes and Failures: For aggregated trial data, provide a matrix combining counts:
glm(cbind(successes, failures) ~ treatment, family = binomial, data = df). - Proportions with Sample Weights: When the response represents the proportion of successes (values between 0.0 and 1.0), specify total sample counts via the
weightsargument:glm(success_prop ~ treatment, weights = sample_size, family = binomial, data = df). - Poisson Exposure Offsets: When event counts are observed over varying exposure periods or population sizes, include an offset to model the incidence rate:
glm(events ~ exposure_group + offset(log(person_years)), family = poisson, data = df).
Output and how to read it
Executing summary(model) produces a comprehensive analytical report. A representative summary from our logistic regression fit appears below:
Call:
glm(formula = am ~ wt + hp, family = binomial(link = "logit"),
data = mtcars)
Deviance Residuals:
Min 1Q Median 3Q Max
-2.2537 -0.1568 -0.0168 0.1543 1.3436
Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) 18.86630 7.44356 2.535 0.0113 *
wt -8.08348 3.06868 -2.634 0.0084 **
hp 0.03626 0.01773 2.045 0.0409 *
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
(Dispersion parameter for binomial family taken to be 1)
Null deviance: 43.230 on 31 degrees of freedom
Residual deviance: 9.841 on 29 degrees of freedom
AIC: 15.841
Number of Fisher Scoring iterations: 8
Each section of this output delivers specific diagnostic and inferential insights:
1. Call
Documents the formula, family specification, and dataset supplied to glm(), ensuring complete transparency and reproducibility across analytical pipelines.
2. Deviance Residuals
Deviance residuals measure the individual contribution of each observation to the overall model deviance. For observation i, the signed deviance residual dᵢ satisfies ∑ dᵢ² = D, where D is the total residual deviance. In a well-specified model, deviance residuals should be approximately symmetric around zero, with median close to zero and quartiles evenly spaced.
3. Coefficients Table
- Estimate: Maximum likelihood parameter estimates on the link function scale. Here, each additional 1,000 lbs of vehicle weight reduces the log-odds of a manual transmission by 8.083 (holding horsepower constant), corresponding to an adjusted odds ratio of exp(-8.083) ≈ 0.00031. Conversely, each additional unit of horsepower increases the log-odds by 0.036 (adjusted OR = exp(0.036) ≈ 1.037).
- Std. Error: Asymptotic standard errors derived from the inverse Fisher information matrix [Xᵀ W X]⁻¹.
- z value: The Wald statistic z = β̂ / SE(β̂). In Gaussian and quasi-families where dispersion is estimated, R reports a Student’s t value instead.
- Pr(>|z|): Two-sided p-value evaluating the null hypothesis H₀: βⱼ = 0 against the standard normal distribution.
4. Dispersion Parameter
Reports the model scale parameter φ. In binomial and Poisson models, theoretical dispersion is fixed at 1.0 by definition. In Gaussian models, φ represents the residual variance σ̂². In quasi-models (e.g., quasipoisson), R estimates φ empirically from Pearson residuals: φ̂ = [1 / (n – p)] ∑ [(yᵢ – μ̂ᵢ)² / V(μ̂ᵢ)].
5. Null Deviance and Degrees of Freedom
Measures the deviance of an intercept-only baseline model: D_null = 2[ℓ(saturated) – ℓ(null)] on n – 1 degrees of freedom (here, 43.230 on 31 df). It represents total variation prior to entering explanatory variables.
6. Residual Deviance and Degrees of Freedom
Measures the unexplained deviance remaining after including model covariates: D_res = 2[ℓ(saturated) – ℓ(model)] on n – p degrees of freedom (here, 9.841 on 29 df). A large reduction from Null Deviance indicates strong predictive performance. To evaluate overall model significance, calculate the drop-in-deviance test statistic: χ² = D_null – D_res = 43.230 – 9.841 = 33.389 on 31 – 29 = 2 df (p = 5.6 × 10⁻⁸).
7. Akaike Information Criterion (AIC)
Calculated as AIC = -2 ln(L) + 2k, where L is the maximized model likelihood and k is the number of estimated parameters (including intercept). When evaluating non-nested models on identical datasets, smaller AIC values denote superior parsimony and out-of-sample predictive capability.
8. Number of Fisher Scoring Iterations
The number of IWLS iterations required to achieve convergence within the tolerance threshold set by glm.control(epsilon = 1e-8). Well-behaved models converge in 4 to 8 iterations. Values approaching the default iteration ceiling (25) often indicate numerical instability, separation, or severe multicollinearity.
Troubleshooting
When executing glm() on observational and experimental research data, analysts frequently encounter numerical convergence warnings and specification flags. The table below outlines common issues, their mathematical root causes, and verified remediations:
| Warning or Error Message | Mathematical Root Cause | Verified Remediation |
|---|---|---|
Warning: glm.fit: algorithm did not converge |
The IWLS algorithm reached the maximum iteration limit (maxit = 25) without satisfying convergence tolerance (epsilon = 1e-8). Often caused by unscaled continuous variables or flat likelihood profiles. |
Standardize continuous covariates using scale(); increase maximum iterations via control = glm.control(maxit = 100); inspect data for extreme outliers or severe collinearity. |
Warning: glm.fit: fitted probabilities numerically 0 or 1 occurred |
Complete separation or quasi-complete separation. A covariate or linear combination perfectly partitions the binary response, causing maximum likelihood estimates and standard errors to diverge toward infinity. | Do not interpret Wald z-tests (the Hauck-Donner effect causes Wald statistics to collapse toward zero). Refit using Firth’s penalized likelihood via the CRAN package logistf, Bayesian priors via rstanarm, or exact logistic regression. |
| Overdispersion in Poisson regression (D_res / df_res >> 1) | The empirical variance of count data exceeds the conditional mean (Var(Y) > μ), violating the Poisson equi-dispersion property. Causes deflated standard errors and inflated Type I error rates. | Refit model using family = quasipoisson to adjust standard errors via empirical dispersion parameter φ, or fit negative binomial regression with dispersion parameter θ using MASS::glm.nb(). |
Coefficients table displays NA for some predictors |
Exact linear dependence (perfect multicollinearity) among covariates, causing a rank-deficient design matrix X. | Inspect pairwise correlations and compute Generalized Variance Inflation Factors (GVIF) via car::vif(); remove linearly redundant variables. In glm(), rank-deficient fits drop collinear terms automatically when singular.ok = TRUE. |
Error: y values must be 0 <= y <= 1 |
Supplying values outside the [0, 1] range to family = binomial. |
Ensure binary outcomes are coded strictly as numeric 0/1 or two-level factors; for aggregated trials, pass cbind(successes, failures) or supply proportions with total sample counts in weights. |
The Hauck-Donner Effect in Separation
When complete separation occurs in logistic regression, the profile likelihood curve approaches an asymptote as the parameter estimate β̂ diverges to infinity. As β̂ grows large, its estimated standard error SE(β̂) increases at an even faster rate. As a consequence, the Wald statistic z = β̂ / SE(β̂) collapses toward zero, generating paradoxically large p-values (e.g., p = 0.998) for predictors that possess perfect explanatory power. This mathematical breakdown, known as the Hauck-Donner effect, underscores why researchers must employ profile likelihood ratio tests or penalized likelihood estimation (Firth’s method) whenever separation warnings appear in R.
Free alternatives
While base R’s stats::glm() provides the universal reference implementation for generalized linear models, researchers utilize various open-source and free computational libraries depending on computational scale, hierarchical data structures, and programming environments:
| Software / Package | Ecosystem & Interface | Optimization Algorithm | Key Research Advantages | Primary Limitations |
|---|---|---|---|---|
stats::glm() |
R (Base stats package) | Iteratively Reweighted Least Squares (IWLS) | Native base R installation; zero external dependencies; universally documented reference standard; comprehensive extractor methods. | Single-threaded, in-memory processing; lacks native support for random effects or regularized penalty estimation. |
statsmodels.formula.api.glm |
Python (statsmodels) | Iteratively Reweighted Least Squares / Newton-Raphson | Seamless integration with Python scientific stack (pandas, NumPy, SciPy); native robust sandwich covariance estimators (HC0-HC3, cluster-robust). | Slower execution on massive categorical factor expansions; formula parsing requires external patsy dependency. |
GLM.jl |
Julia | Native Julia high-performance IRLS | Exceptional computational speed; orders of magnitude faster execution on large datasets; fully typed scientific computing architecture. | Smaller ecosystem of secondary diagnostic packages and automated publication-table generators compared to R. |
lme4::glmer() / glmmTMB |
R (CRAN) | Adaptive Gauss-Hermite Quadrature / Laplace Approximation (TMB) | Accommodates clustered, longitudinal, and multilevel data through generalized linear mixed-effects models (GLMMs); supports zero-inflation. | Substantially greater computational overhead; potential convergence issues with complex random-effects correlation structures. |
glmnet |
R (CRAN) | Cyclical coordinate descent | Estimates Lasso (L1), Ridge (L2), and Elastic Net regularized generalized linear models; excels in high-dimensional p >> n settings. | Does not calculate classical asymptotic standard errors or p-values; requires cross-validation tuning for penalty parameter λ. |
| GNU PSPP | Standalone GUI & CLI | Maximum Likelihood Estimation | Free, open-source alternative to IBM SPSS Statistics; accessible graphical interface for non-programmers and teaching labs. | Limited support for advanced exponential dispersion families, custom link functions, and modern regularized modeling. |
For research involving high-dimensional covariate selection or multicollinearity penalties, consult our comparison of Ridge vs LASSO regression. If your study design incorporates repeated measures, longitudinal clustering, or hierarchical sampling, explore our tutorial on linear mixed-effects models in R. For modern predictive modeling workflows featuring cross-validation and resampling, refer to our guide on the Tidymodels machine learning workflow in R.
FAQ
How do I choose between lm() and glm() in R?
Choose lm() when your dependent variable is continuous, normally distributed, and exhibits homogeneous variance across predictor values (homoscedasticity). Choose glm() when your outcome variable is non-normal—such as binary responses (0 or 1), discrete counts, rates, or skewed positive durations—and requires an exponential family distribution and link function. Fitting glm(y ~ x, family = gaussian(link = "identity")) yields parameter estimates identical to lm(y ~ x), although lm() is computationally optimized specifically for ordinary least squares QR decomposition.
How do I convert logistic regression coefficients into odds ratios?
To convert logistic regression coefficients into odds ratios, apply the exponential function to point estimates: exp(coef(model)). Because the logit link models the log-odds of the outcome, exponentiating parameter estimate β̂ⱼ yields the multiplicative change in the odds of the event for each one-unit increase in covariate Xⱼ, holding other model covariates constant. Always report corresponding 95% profile likelihood confidence intervals via exp(confint(model)).
What is the difference between predict(model, type = “link”) and predict(model, type = “response”)?
The argument type = "link" (the default) calculates predictions on the scale of the linear predictor η = Xβ. For a logistic regression model, this corresponds to predicted log-odds; for a Poisson model, it represents predicted log-counts. In contrast, type = "response" applies the inverse link function g⁻¹(η) to return predictions on the original scale of the response variable—yielding predicted probabilities bounded between 0 and 1 for binomial models, and expected event counts for Poisson models.
How do I include an offset in a Poisson GLM for exposure or population differences?
In count rate models, the expected event count is proportional to exposure time or population size T, such that ln(μ / T) = Xβ, which rearranges algebraically to ln(μ) = ln(T) + Xβ. To constrain the coefficient of ln(T) to exactly 1.0, include it as an offset term. In R, specify this either directly within the model formula using offset(log(exposure)) or via the dedicated argument: glm(y ~ x1 + x2, offset = log(exposure), family = poisson, data = df).
How do I diagnose and handle overdispersion in Poisson models?
Overdispersion occurs when the empirical variance of count data exceeds the conditional mean (Var(Y) > μ). In the summary() output, evaluate the ratio of Residual Deviance to Residual Degrees of Freedom: if D_res / df_res substantially exceeds 1.0 (e.g., > 1.5 or 2.0), overdispersion is present. Remediate this by refitting with family = quasipoisson(link = "log"), which adjusts standard errors using the empirical dispersion parameter φ, or fit a negative binomial model via MASS::glm.nb(y ~ x, data = df), which incorporates an explicit dispersion parameter θ.
What causes the warning “glm.fit: fitted probabilities numerically 0 or 1 occurred”?
This warning signals complete or quasi-complete separation in binary logistic regression: one or more predictors perfectly classify the outcome into 0s and 1s. Consequently, maximum likelihood parameter estimates diverge toward infinity (β̂ → ±∞), standard errors explode, and Wald z-test statistics paradoxically collapse toward zero due to the Hauck-Donner effect. Resolve separation by fitting penalized likelihood models using the logistf package or applying weakly informative priors via rstanarm.
How do I test overall model significance against the intercept-only null model?
Evaluate overall model significance by performing a Likelihood Ratio Test comparing Null Deviance (D_null) and Residual Deviance (D_res). In R, compute the p-value against the chi-square distribution: pchisq(model$null.deviance - model$deviance, df = model$df.null - model$df.residual, lower.tail = FALSE). Alternatively, fit the null model null_mod <- glm(y ~ 1, family = binomial, data = df) and call anova(null_mod, model, test = "Chisq").
Sources
The technical specifications, mathematical definitions, and algorithmic descriptions in this guide are derived from official R documentation and foundational statistical literature:
- R Core Team. (2026). R: A Language and Environment for Statistical Computing. R Foundation for Statistical Computing, Vienna, Austria. Official documentation for package stats:
stats::glm: Fitting Generalized Linear Models. Available at: https://stat.ethz.ch/R-manual/R-devel/library/stats/html/glm.html (Accessed: 2026-11-12).stats::family: Family Objects for Solving GLMs. Available at: https://stat.ethz.ch/R-manual/R-devel/library/stats/html/family.html (Accessed: 2026-11-12).stats::glm.control: Auxiliary for Controlling GLM Fitting. Available at: https://stat.ethz.ch/R-manual/R-devel/library/stats/html/glm.control.html (Accessed: 2026-11-12).stats::predict.glm: Predict Method for GLM Fits. Available at: https://stat.ethz.ch/R-manual/R-devel/library/stats/html/predict.glm.html (Accessed: 2026-11-12).
- Nelder, J. A., & Wedderburn, R. W. M. (1972). Generalized Linear Models. Journal of the Royal Statistical Society: Series A (General), 135(3), 370–384. DOI: 10.2307/2344614.
- McCullagh, P., & Nelder, J. A. (1989). Generalized Linear Models (2nd ed.). Chapman & Hall/CRC, Boca Raton, FL. Monograph defining deviance, exponential families, quasi-likelihood, and IWLS convergence theory.







