R语言glm()函数中weight参数的作用机制探究

Great question—this is such a common source of confusion, especially if you’ve worked with weighted models that use variable scaling elsewhere. Let’s break down exactly how glm() handles weights, and why it’s not the same as dividing your response by the weight vector.

First: It’s Not About Dividing the Response #

Unlike some weighted modeling approaches you might have used, the weight parameter in glm() does not transform your response variable by dividing each element by its corresponding weight. Instead, it integrates weights directly into the model’s fitting criterion (either the log-likelihood function for non-Gaussian families, or weighted least squares for Gaussian).

How It Works by Distribution Family #

The exact interpretation of weight depends on the family you specify in glm()—here are the most common cases:

1. Gaussian (Normal) Family (family = gaussian())

When fitting a linear regression via glm() (which is equivalent to lm() with weights), the weight argument represents precision weights—the inverse of the variance for each observation.

In plain terms, the model minimizes the weighted sum of squared residuals:

Σ [ weight_i * (y_i - ŷ_i)² ]

This gives more influence to observations with higher weights (since their residuals contribute more to the total sum we’re minimizing).

For example, if observation 2 has a weight of 2, its residual is squared and doubled before being added to the sum—so the model will prioritize fitting that point more closely than an observation with weight 1. This is not the same as fitting y/weight ~ x, which would scale the response directly and lead to a different model entirely.

2. Binomial Family (family = binomial())

If your response is a proportion (e.g., fraction of successes), the weight argument should be set to the number of trials per observation.

For instance, if you have a data point where 3 out of 10 trials succeeded (y = 0.3), setting weight = 10 tells glm() to treat this as 10 individual Bernoulli trials (3 successes, 7 failures) rather than a single observation. The log-likelihood is calculated based on the full set of trials, making the model more reliable than fitting to proportions alone.

3. Poisson Family (family = poisson())

Here, weight typically represents the exposure or base count for each observation. For example, if your response is the number of accidents per month, and some observations cover 2 months instead of 1, setting weight = 2 tells the model to account for the longer time period. The model effectively fits the rate (y / weight) while using the weight to adjust the variance of the Poisson distribution.

Key Takeaway #

If you want to model y_i / weight_i, don’t use the weight parameter—instead, create a new response variable directly:

y_scaled <- y / weight
model <- glm(y_scaled ~ x, family = gaussian())

But if you want to assign relative importance to observations (for weighted least squares) or account for group sizes/exposures (for binomial/Poisson), the weight parameter in glm() is designed to handle that correctly without transforming the response.

内容的提问来源于stack exchange，提问作者Elizabeth Eason