Find your next learning step
Read the power-variance law as a fitted mean–variance relationship, then test leverage, prediction and tail claims before trusting the model.
Open the chapter contents · Probability root · Gold Standard of Education · How X Works
Chapter contents
Choose the route closest to the question, or read from Chapter 1 for the complete argument. The teaching chapters remain open; these grouped routes form the index.
Route 1 · Build the family
Route 2 · Work the variance law
Route 3 · Fit and transfer
Route 4 · Respect the limits
Route 5 · Verify and continue
CHAPTER 1 OF 12 · BUILD THE FAMILY
1. Why one power-variance family can clarify several count models
Count data often refuse to behave like a simple Poisson variable. The mean may be four while the variance is eight, twenty or more. The upper tail may also grow faster than a single overdispersion number can explain.
Poisson–Tweedie regression gives one organised way to compare several mechanisms. Its working variance is V(Y)=μ+φμ^p, where μ is the mean, φ is a dispersion parameter and p is the power parameter. Changing p changes how rapidly extra variation grows with the mean.
The family is useful because it puts familiar count models on a shared diagnostic map. It is not a licence to choose a power after looking only at one summary. The mean pattern, dispersion pattern and held-out tail must agree.
CHAPTER 2 OF 12 · BUILD THE FAMILY
2. The power-variance equation is a diagnostic contract
The equation V(Y)=μ+φμ^p separates ordinary Poisson variation μ from extra variation φμ^p. Dividing by μ gives V(Y)/μ=1+φμ^(p-1), so the variance-to-mean ratio may change with the mean.
When p=1, the extra term grows linearly. At p=2 it grows quadratically, as in the negative-binomial variance form. At p=3 it grows cubically, matching the Poisson–inverse-Gaussian variance form under a common parameterisation.
These relationships concern moments. Two distributions can share a variance function and still differ in finite-sample probabilities, parameterisation or fitting method. State which Poisson–Tweedie construction and software convention are being used.
CHAPTER 3 OF 12 · WORK THE VARIANCE LAW
3. Worked ledger: power matters most at larger means
Take φ=0.25 and compare p=1, 2 and 3 at means two, four and eight. The same dispersion coefficient produces modest differences at a low mean and very large differences at a high mean.
| Mean μ | p=1 variance | p=2 variance | p=3 variance |
|---|---|---|---|
| 2 | 2.50 | 3.00 | 4.00 |
| 4 | 5.00 | 8.00 | 20.00 |
| 8 | 10.00 | 24.00 | 136.00 |
At μ=8, the variance-to-mean ratios are 1.25, 3 and 17. A model checked only near μ=2 could therefore look adequate and still fail badly for high-exposure observations.
CHAPTER 4 OF 12 · WORK THE VARIANCE LAW
4. The power parameter is not literally a tail exponent
Authors sometimes describe p as a tail or shape index because it orders familiar count families and changes skewness and upper-tail behaviour. That shorthand can mislead. The parameter p is defined through a power-variance relationship; it is not automatically the exponent in a regular-variation law for the survival probability.
A practical tail check must use probabilities or predictive simulation. Compare observed and replicated zero frequencies, upper quantiles, exceedance counts and maxima within relevant mean or exposure bands.
If a fitted p is large but the held-out upper tail is too light, the fitted variance relationship has not solved the prediction problem. If it is too heavy, a few influential counts may be driving the power estimate.
CHAPTER 5 OF 12 · WORK THE VARIANCE LAW
5. Regression keeps the mean model and exposure visible
Use a log link such as log μ_i=log e_i+x_i^Tβ, where e_i is exposure. The offset changes the expected count while β describes rate ratios. Omitting exposure can make the dispersion parameters absorb a mean-model error.
Check functional form before interpreting p. A missing nonlinear covariate, unmodelled season or omitted group effect can create residual mean–variance growth that resembles a Poisson–Tweedie pattern.
Report predictions on the count scale and, where useful, on the rate scale. A rate-ratio coefficient does not by itself describe the absolute probability of an extreme count.
CHAPTER 6 OF 12 · FIT AND TRANSFER
6. Estimating functions trade a full likelihood for robust moment equations
One established implementation combines a quasi-score for β with Pearson estimating functions for φ and p. This can fit a flexible mean–variance system without evaluating every count probability.
That convenience changes the evidence available. Likelihood-ratio comparisons may not be justified in the ordinary way, while sandwich uncertainty, estimating-function diagnostics and simulation become important.
Inspect convergence from several starting values. Profile or perturb p and φ together because they can compensate for one another over a narrow range of μ. A precise-looking p estimate from a restricted mean range can be structurally weak.
CHAPTER 7 OF 12 · FIT AND TRANSFER
7. Changed context: two models match at four and diverge elsewhere
Suppose two models both have variance eight at μ=4. Model A uses p=2 and φ=0.25. Model B uses p=3 and φ=0.0625. The calibration point cannot distinguish them.
| Mean μ | Model A: p=2, φ=0.25 | Model B: p=3, φ=0.0625 |
|---|---|---|
| 2 | 3.00 | 2.50 |
| 4 | 8.00 | 8.00 |
| 8 | 24.00 | 40.00 |
At μ=8 the models differ by sixteen variance units. This is why a changed-context check should span the fitted mean range and the deployment range, not just the average observation.
CHAPTER 8 OF 12 · FIT AND TRANSFER
8. Predictive-tail checks should be stratified by opportunity
Pool all observations and a model can appear calibrated because low-exposure cases dominate the count. Instead, group predictions by fitted mean, exposure, time or operational regime.
Within each group, compare the number of zeros, the 90th and 99th percentiles, and exceedances above a decision-relevant threshold. Use held-out or cross-validated predictions so the same extreme count is not used both to tune p and to praise the tail fit.
Also inspect probability integral transform variants or randomized quantile residuals suitable for discrete data. No single plot proves adequacy; agreement across the centre and tail is the stronger test.
CHAPTER 9 OF 12 · RESPECT THE LIMITS
9. Power and dispersion uncertainty must travel into decisions
Confidence intervals for p and φ can be strongly correlated. A useful report includes their joint uncertainty and the resulting range of predictive tail probabilities.
When a decision depends on P(Y≥k), propagate coefficient and dispersion uncertainty rather than inserting point estimates only. Parametric bootstrap, estimating-function simulation or a justified Bayesian procedure can expose decision reversals.
Pre-specify the tail quantity before comparing models. Choosing whichever threshold makes the preferred model look best is another form of overfitting.
CHAPTER 10 OF 12 · RESPECT THE LIMITS
10. Limits and common misconceptions
Poisson–Tweedie flexibility does not prove the data-generating mechanism. Excess zeros can come from a mixture of opportunities, recording rules or structural absence. A good marginal fit does not identify which explanation is true.
The variance function may be estimated well while individual count probabilities remain uncertain. Underdispersion extensions can rely on moment-based constructions that do not behave like every classical Poisson mixture.
Do not equate a large p with a universal heavy-tail law, and do not compare p values from incompatible parameterisations. Small samples, a narrow mean range and influential extremes can all destabilise the estimate.
Evidence and reproducibility
Bonat and colleagues set out the regression and estimating-function framework in Extended Poisson–Tweedie: Properties and regression models for count data. El-Shaarawi and colleagues demonstrate the family in ecological abundance modelling in Modelling species abundance using the Poisson–Tweedie family.
Archive the offset, covariates, parameterisation, starting values, estimating equations, covariance estimator, convergence history and predictive simulation code. Save checks by fitted-mean band, not only a single global residual plot.
CHAPTER 11 OF 12 · VERIFY AND CONTINUE
11. Frequently asked questions
Is p=2 always negative binomial?
It matches a familiar negative-binomial variance form, but the exact probability model and parameterisation still need to be stated.
Can effective exposure be estimated instead of supplied?
Sometimes, but then exposure uncertainty must enter the model. A convenient fitted offset is not the same as a measured opportunity.
Should I choose the model with the best in-sample tail?
No. Use held-out predictive scoring and decision-relevant exceedance checks.
Can p be fixed?
Yes, when a mechanism or comparison justifies it. Show how conclusions change across credible fixed powers.
What belongs in a minimum report?
Mean equation, offset, φ and p with joint uncertainty, mean-range coverage, tail checks, convergence evidence and the exact software convention.
CHAPTER 12 OF 12 · VERIFY AND CONTINUE
12. Final transfer check and connected reading
Calibrate two models to variance twelve at μ=4: one with p=2 and one with p=3. Solve for each φ, then compare their variances at μ=2 and μ=10. Explain which deployment range provides the stronger test.
Continue through the Probability root, negative-binomial guide, Neyman Type A guide, Poisson–inverse-Gaussian guide and Poisson–lognormal guide. Together they separate mean structure, variance growth and predictive tails.
Top · Contents · Previous · How X Works