The Dispersion Parameter φ
You'll learn what the dispersion parameter does and why it matters for controlling variance in generalized linear models.
What Is the Dispersion Parameter?
In the exponential family of distributions, the dispersion parameter φ (phi) acts like a "variance knob" that controls how spread out your data is around its mean. Remember from the previous lesson that the variance formula in exponential family distributions includes this parameter:
Variance = φ × V(μ)
where V(μ) is the variance function that depends on the mean.
Think of φ as a volume control: when φ = 1, you get the "standard" variance for that distribution. When φ > 1, the data is more dispersed (overdispersed). When φ < 1, it's less dispersed (underdispersed, though this is rare in practice).
Why Some Distributions Have φ = 1
Here's the important part: some distributions have a fixed dispersion parameter of 1, meaning you can't adjust it. These include:
- Binomial distribution: variance is fully determined by n and p
- Poisson distribution: variance equals the mean (λ), no extra wiggle room
These distributions have their variance "locked in" by their mathematical structure. You can't add an extra variance parameter because the mean already determines everything.
In contrast, distributions like Normal and Gamma have a free φ parameter that must be estimated from your data, giving them flexibility to match real-world variance patterns.
Practical Impact
When φ = 1 is fixed, your model assumes the theoretical variance structure is correct. If real data shows extra variance (overdispersion), you'll need special techniques to handle it.
Key Takeaway: The dispersion parameter φ controls variance scaling in exponential family distributions; some distributions (Binomial, Poisson) fix φ = 1, while others (Normal, Gamma) estimate it freely from data.