Worked Comparisons of Offsets and Weight Types

Why These Comparisons Matter

In a generalized linear model, several quantities can affect a fit without playing the same role. An exposure offset fixes a known contribution to the linear predictor. A binomial trial total describes how many Bernoulli outcomes contribute to an observation. A frequency weight represents repeated observations, while a variance weight expresses relative precision. Confusing these meanings can leave a coefficient looking plausible while changing its interpretation, likelihood, or standard error.

The comparisons below focus on what each quantity means in a model and what follows from that meaning. For the worked uncertainty calculations, standard errors are model-based; the Gaussian examples state explicitly how dispersion is estimated.

Offsets: Fixing Exposure at Coefficient One

For a Poisson count \(Y_i\) observed over known exposure \(t_i>0\), a log-rate model can be written \( \log E[Y_i]=\log(t_i)+\beta_0+\beta_1 x_i \). The offset is \(\log(t_i)\): its coefficient is fixed at one rather than estimated. Exponentiating gives \( E[Y_i]=t_i\exp(\beta_0+\beta_1 x_i) \), so the linear predictor after removing exposure describes the log rate per exposure unit.

For a group indicator \(x_i\in\{0,1\}\), the coefficient \(\beta_1\) is the log ratio of the two expected rates, and \(\exp(\beta_1)\) is their rate ratio. Omitting the offset instead models expected counts directly. If exposure differs between groups, that is a different model and generally answers a different question.

Binomial Trial Totals: Information From Repeated Trials

For grouped binomial data, an observation records \(S_i\) successes out of \(n_i\) trials. Under a logistic model, \(S_i\sim\operatorname{Binomial}(n_i,p_i)\) and \( \operatorname{logit}(p_i)=\eta_i \). The total \(n_i\) is part of the sampling distribution: it determines how much information that group contributes about \(p_i\). A proportion \(S_i/n_i\) alone does not say whether it came from a few trials or many.

Frequency Weights and Variance Weights

An integer frequency weight \(m_i\) means that the observation represents \(m_i\) records with the same response and covariates. Its log-likelihood contribution is multiplied by \(m_i\), just as it would be if those records were written out individually. A variance weight has a different interpretation. In a common Gaussian model, a known relative precision weight \(w_i>0\) means that observation \(i\) has variance \(\sigma^2/w_i\). It changes how precisely that observation informs the fit; it does not assert that the observation occurred \(w_i\) times.

Proof. Let \(f(y_i\mid x_i;\theta)\) be the likelihood contribution from one record. The \(m_i\) identical records contribute \( \prod_{j=1}^{m_i} f(y_i\mid x_i;\theta)=f(y_i\mid x_i;\theta)^{m_i} \). Taking logarithms gives \(m_i\log f(y_i\mid x_i;\theta)=m_i\ell_i(\theta)\). Summing over records produces the frequency-weighted log-likelihood. Its derivatives with respect to \(\theta\), including the score and information, are therefore the same as those from the expanded data. For a Gaussian model, dispersion estimates also depend on the correct residual degrees of freedom for the expanded sample. \(\square\)

A Rate-Model Result for Exposure

The effect of an offset can also be seen directly from the likelihood. Suppose independent counts satisfy \(Y_i\sim\operatorname{Poisson}(t_i\lambda)\), with known \(t_i>0\) and a common unknown rate \(\lambda=\exp(\beta)\). This is an intercept-only log-rate model with offset \(\log(t_i)\).

Derivation. Ignoring terms that do not depend on \(\beta\), the log-likelihood is \( \ell(\beta)=\sum_i[Y_i(\log t_i+\beta)-t_i\exp(\beta)-\log(Y_i!)] \). Its score and second derivative are \( \ell'(\beta)=\sum_iY_i-\exp(\beta)\sum_i t_i \) and \( \ell''(\beta)=-\exp(\beta)\sum_i t_i<0 \). The score is zero exactly when \( \exp(\hat\beta)=\sum_iY_i/\sum_i t_i \), and the negative second derivative shows this is the unique maximum. Therefore \(\hat\lambda=\exp(\hat\beta)=\sum_iY_i/\sum_i t_i\). The offset is what makes total exposure, rather than the number of records, the denominator of the fitted rate. \(\square\)

How to Choose the Right Interpretation

Before interpreting coefficients or standard errors, ask what the quantity attached to each record represents. Exposure belongs in a log-rate model when expected counts scale proportionally with known time, area, or another exposure measure. A binomial total belongs in the response distribution when successes are counted out of known trials. A frequency weight represents replicated records. A variance weight represents relative precision under a stated variance model.

These meanings can yield similar coefficient estimates while implying different uncertainty. For example, the weighted mean in the Gaussian example is the same under both weight interpretations, but the replication structure and estimated dispersion differ. The binomial example shows a parallel warning: the observed proportions determine the fitted group logits, but the trial totals determine the information supporting them.

Common Mistake

• Wrong reasoning: “The model accepts a weight column, so I can put exposure, trial totals, replication counts, or precision values into that column interchangeably.”
• Correction: A model's weight argument does not define the scientific meaning of its input. Specify whether the quantity is an offset, a binomial trial total, a frequency count, or a variance weight, then use a likelihood and uncertainty calculation that match that interpretation.
• Another warning: Do not infer equivalence from matching point estimates. Check the likelihood contributions and the standard-error calculation as well.