13
$\begingroup$

I have been discussing with several lab members about this one, and we have gone to several sources but still don't quite have the answer:

When we say a GLM has a family of poisson let's say are we talking about the distribution of the residuals or the response variable?

Points of contention

  1. Reading this article it states that the assumptions of the GLM are The statistical independence of observations, the correct specification of the link and variance function (which makes me think about the residuals, not the response variable), the correct scale of measurement for response variable and lack of undue influence of single points

  2. This question has two answers with two points each, the one that appears first talks about the residuals, and the second one about the response variable, which is it?

  3. In this blogpost, when talking about assumptions, they state "The distribution of the residuals can be other, eg, binomial"

  4. At the beginning of this chapter they say that the structure of the errors has to be Poisson, but the residuals will surely have positive and negative values, how can that be Poisson?

  5. This question, which often is cited in questions such as this one to make them duplicated does not have an accepted answer

  6. This question the answers talk about response and not residuals

  7. In this course description from the University of Pensilvania they talk about the response variable in the assumptions, not the residuals

$\endgroup$
18
$\begingroup$

The family argument for glm models determines the distribution family for the conditional distribution of the response, not of the residuals (except for the quasi-models).

Look at this way: For the usual linear regression, we can write the model as $$Y_i \sim \text{Normal}(\beta_0+x_i^T\beta, \sigma^2). $$ This means that the response $Y_i$ has a normal distribution (with constant variance), but the expectation is different for each $i$. Therefore the conditional distribution of the response is a normal distribution (but a different one for each $i$). Another way of writing this model is $$ Y_i = \beta_0+x_i^T\beta + \epsilon_i $$ where each $\epsilon_i$ is distributed $\text{Normal}(0, \sigma^2)$.

So for the normal distribution family both descriptions are correct (when interpreted correctly). This is because for the normal linear model we have a clean separation in the model of the systematic part (the $\beta_0+x_i^T\beta$) and the disturbance part (the $\epsilon_i$) which are simply added. But for other family functions, this separation is not possible! There is not even a clean definition of what residual means (and for that reason, many different definitions of "residual").

So for all those other families, we use a definition in the style of the first displayed equation above. That is, the conditional distribution of the response. So, no, the residuals (whatever defined) in Poisson regression do not have a Poisson distribution.

$\endgroup$
13
$\begingroup$

Further to Kjetil's excellent answer, I wanted to add some specific examples to help clarify the meaning of a conditional distribution, which can be a bit of an elusive concept.

Let's say you took a random sample of 100 fish from a lake and you are interested in seeing how the age of the fish affects several outcome variables:

  1. Fish weight (Weight);
  2. Whether or not the fish are longer than 30cm;
  3. Number of fish scales.

The first outcome variable is continuous, the second is binary (0 = fish is NOT longer than 30 cm; 1 = fish IS longer than 30 cm) and the third is a count variable.

Simple Linear Regression

How does Age affect Weight? You are going to formulate a simple linear regression model of the form:

$$ \text{Weight} = \beta_0+\beta_1*\text{Age} + \epsilon $$

where the $\epsilon$'s are independent, identically distributed, following a Normal distribution with mean 0 and standard deviation $\sigma$. In this model, the mean of the Weight variable for all the fish in the lake sharing the same age is assumed to vary linearly with age. The conditional mean is represented by $\beta_0 + \beta_1*\text{Age}$. It is called conditional because it is the mean weight for all the fish in the lake with the same Age. (The unconditional mean weight would be the mean weight of all the fish in the lake, regardless of their age.)

Simple Binary Logistic Regression

How does Age affect whether or not the fish are longer than 30cm? You are going to formulate a simple binary logistic regression model of the form:

$$ log(\frac{p}{1-p}) = \beta_0+\beta_1*\text{Age} $$

where $p$ denotes the conditional probability that a fish of a given age is longer than 30cm. In this model, the conditional mean of the variable "whether or not the fish are longer than 30cm" corresponding to all fishes in the lake sharing the same age is assumed to vary linearly with age after being fed to the logit transformation. The logit-transformed conditional mean is represented by $\beta_0 + \beta_1*\text{Age}$. This model works because we assume that the distribution of values of the variable "whether or not the fish are longer than 30cm" for a given age is a Bernoulli distribution. Recall that for this distribution, the variance is a function of the mean value, so if we can estimate its mean value, we can also estimate its variance. (The mean of a Bernoulli variable is $p$ and the variance is $p*(1-p)$.) See also https://www.theanalysisfactor.com/link-functions-and-errors-in-logistic-regression/.

Simple Poisson Regression

How does Age affect the number of fish scales? You are going to formulate a simple Poisson regression model of the form:

$$ log(\mu) = \beta_0+\beta_1*\text{Age} $$

where $\mu$ denotes the conditional mean value of the outcome variable "number of fish scales" for fish of a given age (that is, the expected number of fish scales for all fish in the lake of a given age). In this model, the conditional mean of the outcome variable is assumed to vary linearly with age after being fed to the log transformation. The log-transformed conditional mean is represented by $\beta_0+\beta_1*\text{Age}$. This model works because we assume that the distribution of values of the variable "number of fish scales" for all the fish in the lake of a given age is a Poisson distribution. Recall that for this distribution, the mean and variance are equal so it is sufficient to model its mean value.

To sum up, a conditional distribution represents the distribution of the outcome values for specific values of the predictor variable(s) included in the model. Each type of regression model illustrated above imposes certain distributional assumptions on the conditional distribution of the outcome variable given Age. Based on these distributional assumptions, the model proceeds to formulate how (1) the mean of the conditional distribution varies as a function of age (simple linear regression), (2) the logit-transformed mean of the conditional distribution varies as a function of age (simple binary logistic regression) or (3) the log-transformed mean of the conditional distribution varies as a function of age.

For each type of model, one can define corresponding residuals for the purpose of model checking. In particular, Pearson and deviance residuals could be defined for the logistic and Poisson regression models.

$\endgroup$
  • 2
    $\begingroup$ EXCELLENT answers. Thanks to both of you. I never realized that the "actual" residual is never really explicit in the general GLM framework like it is in the normal distribution case. $\endgroup$ – mlofton Oct 30 '18 at 20:46
  • 1
    $\begingroup$ @mlofton: Thank you for your kind words. An excellent question invited excellent answers. We all benefit from this exchange of knowledge. $\endgroup$ – Isabella Ghement Oct 30 '18 at 21:06
  • 4
    $\begingroup$ I used GLM's a long time ( for a year or two like 10 years ago ) and that was always my confusion but I never knew that was my confusion until it was asked so clearly and explained so clearly. So sometimes confusion means not even being able to ask the right question. Thanks again. $\endgroup$ – mlofton Oct 30 '18 at 22:41
  • 1
    $\begingroup$ You are absolutely right! Confusion is part of learning - when we struggle with something for a while, we are primed to understand it better when we suddenly stumble upon a clear explanation. $\endgroup$ – Isabella Ghement Oct 30 '18 at 22:48
  • 1
    $\begingroup$ My pleasure and thank you for your excellent answer @IsabellaGhement $\endgroup$ – Patrick Apr 28 at 13:47

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.