Warning: non-integer #successes in a binomial glm! (survey packages)

There is nothing wrong, computationally, but statistically you may not be doing something that makes much sense. In such a case, it is probably better to use a robust regression method, which is generally a good idea for proportional response data if your data include units with exactly 1 or exactly 0.

Like @HoongOoi said, glm.fit with binomial family expects integer counts and throws a warning otherwise; if you want non-integer counts, use quasi-binomial. The rest of my answer compares these.

Quasi-binomial in R for glm.fit is exactly the same as binomial for the coefficient estimates (as mentioned in comments by @HongOoi) but not for the standard errors (as mentioned in the comment by @nograpes).

Comparison of source code

A diff on the source code of stats::binomial and stats::quasibinomial shows the following changes:

• the text "binomial" becomes "quasibinomial"
• the aic function returns NA instead of the calculated AIC

and the following removals:

• setting outcomes to 0 when weights = 0
• check on integrality of weights
• simfun function to simulate data

Only simfun could make a difference, but the source code of glm.fit shows no use of that function, unlike other fields in the object returned by stats::binomial such as mu.eta and link.

Minimal working example

The results from using quasibinomial or binomial are the same for the coefficients in this minimal working example:

library('MASS')
library('stats')

gen_data <- function(n=100, p=3) {

  set.seed(1)  
  weights <- stats::rgamma(n=n, shape=rep(1, n), rate=rep(1, n))
  y <- stats::rbinom(n=n, size=1, prob=0.5)
  theta <- stats::rnorm(n=p, mean=0, sd=1)
  means <- colMeans(as.matrix(y) %*% theta)
  x <- MASS::mvrnorm(n=n, means, diag(1, p, p))

  return(list(x=x, y=y, weights=weights, theta=theta))  
}

fit_glm <- function(family) {
  data <- gen_data()
  fit <- stats::glm.fit(x = data$x,
                        y = data$y,
                        weights = data$weights,
                        family = family)
  return(fit)
}

fit1 <- fit_glm(family=stats::binomial(link = "logit"))
fit2 <- fit_glm(family=stats::quasibinomial(link = "logit"))

all(fit1$coefficients == fit2$coefficients)

Comparison with the quasibinomial probability distribution

This thread suggests that the quasibinomial distribution is different from the binomial distribution with an additional parameter phi. But they mean different things in statistics and in R.

First, no place in the source code of quasibinomial mentions that additional phi parameter.

Second, a quasiprobability is similar to a probability, but not a proper one. In this case, one cannot compute the term (n \choose k) when the numbers are non-integers, although one could with the Gamma function. This may be a problem for the definition of the probability distribution but is irrelevant for estimation, as the term (n choose k) do not depend on the parameter and fall out of optimisation.

The log-likelihood estimator is:

The log-likelihood as a function of theta with the binomial family is:

where the constant is independent of the parameter theta, so it falls out of optimisation.

Comparison of standard errors

The standard errors calculated by stats::summary.glm use a different dispersion value for the binomial and quasibinomial families, as mentioned in stats::summary.glm:

The dispersion of a GLM is not used in the fitting process, but it is needed to find standard errors. If dispersion is not supplied or NULL, the dispersion is taken as 1 for the binomial and Poisson families, and otherwise estimated by the residual Chisquared statistic (calculated from cases with non-zero weights) divided by the residual degrees of freedom.

...

cov.unscaled: the unscaled (dispersion = 1) estimated covariance matrix of the estimated coefficients.

cov.scaled: ditto, scaled by dispersion.

Using the the above minimal working example:

summary1 <- stats::summary.glm(fit1)
summary2 <- stats::summary.glm(fit2)

print("Equality of unscaled variance-covariance-matrix:")
all(summary1$cov.unscaled == summary2$cov.unscaled)

print("Equality of variance-covariance matrix scaled by `dispersion`:")
all(summary1$cov.scaled == summary2$cov.scaled)

print(summary1$coefficients)
print(summary2$coefficients)

shows the same coefficients, same unscaled variance-covariance matrix, and different scaled variance-covariance matrices:

[1] "Equality of unscaled variance-covariance-matrix:"
[1] TRUE
[1] "Equality of variance-covariance matrix scaled by `dispersion`:"
[1] FALSE
       Estimate Std. Error   z value   Pr(>|z|)
[1,] -0.3726848  0.1959110 -1.902317 0.05712978
[2,]  0.5887384  0.2721666  2.163155 0.03052930
[3,]  0.3161643  0.2352180  1.344133 0.17890528
       Estimate Std. Error   t value   Pr(>|t|)
[1,] -0.3726848  0.1886017 -1.976042 0.05099072
[2,]  0.5887384  0.2620122  2.246988 0.02690735
[3,]  0.3161643  0.2264421  1.396226 0.16583365

There's nothing wrong, glm is just picky when it comes to specifying binomial (and Poisson) models. It warns if it detects that the no. of trials or successes is non-integral, but it goes ahead and fits the model anyway. If you want to suppress the warning (and you're sure it's not a problem), use family=quasibinomial instead.