Statistics Graduate Level Quiz: Poisson Regression

10 questions · exam conditions

0:00

Poisson RegressionQuestion 1 of 10

Independent count observations are analyzed using a Poisson log-linear model. The fitted means appear well calibrated across covariate patterns, but the Pearson statistic divided by its residual degrees of freedom is $2.4$ . The analyst originally reported model-based Poisson standard errors.

If the conditional mean model is correctly specified, which response is most appropriate?

Retain the coefficient estimates but inflate their covariance estimate, for example using a quasi-Poisson or sandwich procedure

Divide every coefficient estimate by

2.4

because the fitted log rates are inflated by the dispersion

Retain the model-based standard errors because dispersion affects prediction but not coefficient inference

Replace the log link with an identity link because overdispersion indicates that the original link is inconsistent

← Back to quizzes

Statistics Graduate Level Quiz

Statistics Graduate Level Quiz: Poisson Regression

Practice Poisson Regression in Statistics Graduate Level with focused quiz questions that help you check what you know, review explanations, and build confidence with test-style prompts.

What this quiz covers

This quiz focuses on Poisson Regression, giving you a quick way to practice the rules, question types, and explanations that matter most for Statistics Graduate Level.

How to use this quiz

Try each quiz question before looking at the correct answer. Use the explanations to review missed ideas, then come back to similar questions until the pattern feels familiar.

All questions

Question 1

If the conditional mean model is correctly specified, which response is most appropriate?

Retain the coefficient estimates but inflate their covariance estimate, for example using a quasi-Poisson or sandwich procedure (correct answer)
Divide every coefficient estimate by $2.4$ because the fitted log rates are inflated by the dispersion
Retain the model-based standard errors because dispersion affects prediction but not coefficient inference
Replace the log link with an identity link because overdispersion indicates that the original link is inconsistent

Explanation: Whenever you encounter overdispersion in a Poisson model, the critical distinction to make is: what is affected and what isn't? Overdispersion — indicated here by a Pearson statistic divided by residual degrees of freedom (the dispersion factor

\hat{\phi} = 2.4

) exceeding 1 — means the observed variance is larger than the Poisson model assumes (

\text{Var}(Y) = \mu

). However, if the conditional mean structure

E[Y|\mathbf{x}] = \exp(\mathbf{x}^\top\boldsymbol{\beta})

is correctly specified, the coefficient estimates

\hat{\boldsymbol{\beta}}

remain consistent and unbiased. The problem is purely one of underestimated uncertainty: Poisson standard errors assume no extra variance, so they are too small, making inference anti-conservative. The correct response is A: retain the coefficient estimates but correct the covariance matrix. A quasi-Poisson approach scales the variance-covariance matrix by

\hat{\phi} = 2.4

, inflating standard errors by

\sqrt{2.4} \approx 1.55

. A sandwich (robust) estimator achieves a similar correction empirically. Both fix the inferential problem without discarding the useful point estimates. B is wrong because the dispersion factor does not distort the fitted log rates or the coefficients — dividing

\hat{\boldsymbol{\beta}}

2.4

would destroy valid estimates with no statistical justification. C is wrong because overdispersion does affect inference: understated standard errors inflate Type I error rates, making this anything but ignorable. D is wrong because the link function choice is about the mean structure, not variance; switching to an identity link solves nothing about overdispersion. A handy rule: overdispersion corrupts standard errors, not coefficients — always adjust the covariance estimate, not the point estimates themselves.

Question 2

Three independent units have event counts $2$ , $7$ , and $3$ and corresponding exposures $1$ , $2$ , and $1$ person-years. An intercept-only Poisson rate model is fitted as $\log(\mu_i)=\beta_0+\log(t_i)$ .

What is the maximum likelihood estimate of $\beta_0$ ?

$\log(3)$ , obtained by dividing the total count by the total exposure (correct answer)
$\log(4)$ , obtained by averaging the three observed counts before taking logs
$\log(10/3)$ , obtained by averaging the three individual observed rates
$\log(12)$ , obtained by treating the total count as the common fitted mean

Explanation: When fitting a Poisson rate model, the key insight is that maximum likelihood estimation works by pooling all the data through the likelihood function — not by averaging pre-computed quantities across units. In this model,

\log(\mu_i) = \beta_0 + \log(t_i)

, which means

\mu_i = e^{\beta_0} \cdot t_i

. So

e^{\beta_0}

represents the common underlying rate

\lambda

. The Poisson log-likelihood is proportional to

\sum_i y_i \log(\mu_i) - \mu_i

. Substituting and differentiating with respect to

\beta_0

, setting the score equation to zero yields:

\sum_i y_i = \sum_i \mu_i = e^{\beta_0} \sum_i t_i

Solving:

\hat{\lambda} = e^{\hat{\beta}_0} = \frac{\sum_i y_i}{\sum_i t_i} = \frac{2+7+3}{1+2+1} = \frac{12}{4} = 3

Therefore

\hat{\beta}_0 = \log(3)

, confirming A is correct. The MLE pools total counts over total exposure — exactly what you'd intuitively call the overall observed rate. Choice B is wrong because averaging the raw counts

(2+7+3)/3 = 4

ignores the exposure weights entirely, conflating counts with rates. Choice C computes the unweighted average of unit-level rates

(2/1 + 7/2 + 3/1)/3 = (2+3.5+3)/3 = 8.5/3

, which gives equal weight to each unit regardless of its exposure — this is not what the MLE does. Choice D mistakes the total count (12) for the estimated mean rate, forgetting to divide by total exposure. The study tip: whenever you see an intercept-only Poisson rate model, the MLE of the rate parameter is always total events divided by total exposure — a consequence of the score equation, not a guess.

Question 3

Each subject contributes a count over one year to a Poisson regression with a log-exposure offset. An analyst divides every subject-year into two half-year records, allocates each subject's observed events to the half-year in which they occurred, keeps covariates constant within the year, and uses $\log(0.5)$ as the offset for each new record.

Assuming a constant event rate within each subject-year, how should this restructuring affect estimation?

The coefficient estimates remain the same, but their standard errors decrease by a factor of approximately $\sqrt{2}$
The coefficient estimates and likelihood-based information remain the same, apart from likelihood terms unrelated to the coefficients (correct answer)
The intercept decreases by $\log(2)$ , while all slope estimates and their standard errors remain unchanged
All coefficient estimates double because each original subject now contributes two independent Poisson observations

Explanation: Whenever you see a question about splitting time-period records in Poisson regression, think carefully about what information is actually being added versus merely rearranged. In Poisson regression with a log-exposure offset, the model for a full subject-year with exposure

t=1

\log(\mu) = \mathbf{x}^\top\boldsymbol{\beta} + \log(1)

. When you split that year into two half-years, each half gets offset

\log(0.5)

, so the expected count per half becomes

\mu_i = 0.5\exp(\mathbf{x}^\top\boldsymbol{\beta})

. The two halves must sum to the original year's count, and since covariates are held constant within the year, the joint likelihood of the two half-year counts factors into a product that is mathematically equivalent to the original full-year likelihood — up to combinatorial terms (how events split between halves) that do not depend on

\boldsymbol{\beta}

. Therefore, B is correct: coefficient estimates and all likelihood-based information (including standard errors and the Fisher information matrix) are preserved, apart from those likelihood terms unrelated to the coefficients. A is wrong because splitting existing data into smaller windows does not inject new information; you cannot reduce standard errors simply by subdividing records. C confuses the offset adjustment with an intercept shift — the offset absorbs the exposure change so the intercept need not move; all

\boldsymbol{\beta}

estimates are unchanged. D is the most dangerous distractor: treating the two half-year records as independent observations would be wrong because their counts are constrained to sum to the original count, so no new independent information exists. The key study tip: restructuring existing exposure time never creates new data. Watch for traps suggesting that subdividing records inflates precision or shifts coefficients — it doesn't, as long as the offset correctly accounts for each sub-period's exposure.

Question 4

Two individuals have covariate values $x=-1$ and $x=1$ . Their counts follow a Poisson regression with fitted means $\mu(x)=\exp(\alpha+\beta x)$ . An analyst proposes predicting their combined expected count by evaluating the fitted mean at the average covariate value and multiplying by two.

Which expression gives the correct combined expected count, and how does it compare with the analyst's proposal when $\beta\ne0$ ?

$2\exp(\alpha)$ , which equals the analyst's proposal because the covariate values average to zero
$2\exp(\alpha+\beta)$ , which exceeds the proposal because the positive covariate determines the total
$\exp(2\alpha)$ , which may exceed or fall below the proposal depending on the intercept
$2\exp(\alpha)\cosh(\beta)$ , which exceeds the proposal because the mean function is convex (correct answer)

Explanation: Whenever you see a question involving a nonlinear model like Poisson regression, pause before aggregating predictions — the order in which you average versus apply the nonlinear function matters enormously. The correct combined expected count is simply the sum of each individual's fitted mean:

\mu(-1) + \mu(1) = \exp(\alpha - \beta) + \exp(\alpha + \beta)

. Factor out

\exp(\alpha)

and recognize the hyperbolic cosine identity:

\exp(\alpha)[\exp(-\beta)+\exp(\beta)] = 2\exp(\alpha)\cosh(\beta)

. The analyst's proposal instead evaluates the mean at the average covariate

\bar{x}=0

, giving

2\exp(\alpha+\beta\cdot 0)\cdot 1 = 2\exp(\alpha)

. Since

\cosh(\beta) \geq 1

for all real

\beta

, with equality only at

\beta=0

, the correct answer strictly exceeds the analyst's proposal whenever

\beta\neq 0

. This is a direct consequence of Jensen's inequality: because

\exp(\cdot)

is convex,

\frac{1}{2}[\exp(\alpha-\beta)+\exp(\alpha+\beta)] \geq \exp(\alpha)

. So D is correct. Choice A claims both expressions are equal, which would only hold if convexity had no effect — it ignores Jensen's inequality entirely. Choice B incorrectly sums only the positive-covariate individual's mean, as if

x=-1

contributed nothing; there's no justification for discarding half the data. Choice C introduces

\exp(2\alpha)

, which has no grounding in the model — squaring the intercept is a purely algebraic error with no probabilistic meaning. Study tip: On any exam question involving predictions from nonlinear models, always sum predictions at the actual covariate values — never plug in the average covariate. Jensen's inequality guarantees these differ whenever the function is strictly convex.

Question 5

A Poisson regression for accident counts includes the logarithm of observation time as an offset: $\log(\mu_i)=\beta_0-0.30G_i+0.02A_i+\log(t_i),$ where $G_i$ is a group indicator, $A_i$ is age in years, and $t_i$ is observation time in years.

Compare a group- $1$ person aged $60$ observed for $2$ years with a group- $0$ person aged $50$ observed for $1$ year. What is the ratio of their expected accident counts?

$\exp(-0.10)\approx 0.90$ , after combining the group and age differences but ignoring exposure
$2\exp(-0.30)\approx 1.48$ , after accounting for group and exposure but not age
$2\exp(-0.10)\approx 1.81$ , after accounting for group, age, and exposure (correct answer)
$\exp(0.10)\approx 1.11$ , after treating the offset as a rate adjustment rather than count exposure

Explanation: When comparing expected counts in Poisson regression with an offset, you need to carefully track every term in the linear predictor for both individuals, then take their ratio. Set up the two linear predictors explicitly. For the group-1 person aged 60 observed 2 years:

\log(\mu_1) = \beta_0 - 0.30(1) + 0.02(60) + \log(2)

. For the group-0 person aged 50 observed 1 year:

\log(\mu_0) = \beta_0 - 0.30(0) + 0.02(50) + \log(1)

. The ratio

\mu_1/\mu_0

is found by subtracting the second equation from the first:

\log(\mu_1/\mu_0) = -0.30 + 0.02(60-50) + \log(2) - \log(1) = -0.30 + 0.20 + \log(2) = -0.10 + \log(2)

. Therefore

\mu_1/\mu_0 = 2\exp(-0.10) \approx 1.81

, confirming answer C. Answer A drops the offset entirely, treating the ratio as a pure covariate comparison. This ignores the fact that longer observation time mechanically increases expected counts — a critical omission when exposure differs between subjects. Answer B correctly incorporates the group effect and the 2-year exposure but omits the age coefficient. The age difference of 10 years contributes

0.02 \times 10 = 0.20

to the log-ratio, which cannot be ignored. Answer D misinterprets the offset conceptually, treating

\log(t_i)

as merely a rate scaling factor rather than a structural component of the expected count, leading to an incorrect exponent sign. Your study tip: when computing ratios from a fitted Poisson model, always write out the full linear predictor for each unit, subtract, and exponentiate. The offset is never optional — it directly governs how counts scale with exposure.

Question 6

A Poisson regression for the number of infections during a fixed follow-up period uses the model $\log(\mu)=\beta_0+0.40T+0.30x-0.20Tx,$ where $T=1$ denotes treatment and $T=0$ denotes control. The covariate $x$ is centered, so $x=0$ represents the reference covariate value.

According to the fitted model, what is the treatment-to-control incidence rate ratio for a patient with $x=2$ ?

$\exp(-0.40)\approx 0.67$ , because the interaction reverses the main treatment effect
$\exp(0)=1.00$ , because the treatment effect is zero at this covariate value (correct answer)
$\exp(0.40)\approx 1.49$ , because the treatment coefficient is the adjusted log rate ratio
$\exp(0.80)\approx 2.23$ , because both treatment-related coefficients contribute positively

Explanation: When a regression model includes an interaction term, the effect of one variable depends on the level of another — and this is exactly what's being tested here. To find the treatment-to-control incidence rate ratio at a specific covariate value, you compute

\log(\mu_{T=1}) - \log(\mu_{T=0})

at that value of

x

, which gives you the log IRR. At

x = 2

\log(\mu_{T=1}) = \beta_0 + 0.40(1) + 0.30(2) - 0.20(1)(2) = \beta_0 + 0.40 + 0.60 - 0.40

, and

\log(\mu_{T=0}) = \beta_0 + 0.30(2) = \beta_0 + 0.60

. Subtracting gives a log IRR of

0.40 + 0.60 - 0.40 - 0.60 = 0

, so the IRR is

\exp(0) = 1.00

. Answer B is correct. Answer A claims the IRR is

\exp(-0.40)

, which would require the interaction term alone to flip the effect — but the full arithmetic doesn't support this; you must combine all treatment-related terms. Answer C uses only

\beta_T = 0.40

, ignoring the interaction entirely. This would only be valid at

x = 0

(the reference value), not at

x = 2

. Answer D adds

0.40 + 0.40 = 0.80

, treating both coefficients as if they contribute positively — but the interaction coefficient is

-0.20

, which at

x = 2

contributes

-0.40

, not

+0.40

. Your key strategy: in any model with an interaction, never read off a single coefficient as "the effect." Always plug in the specific covariate value and compute the full contrast between groups — that's the only way to get the correct conditional effect.

Question 7

Two nested Poisson regressions are fitted to the same observations. The reduced model has residual deviance $428$ , and the full model has residual deviance $419$ . The full model adds two slope parameters. Assume the usual large-sample likelihood-ratio approximation is appropriate.

Which conclusion follows from comparing the two models?

The test statistic is $4.5$ on $1$ degree of freedom, indicating significance at $0.05$
The test statistic is $9$ on $2$ degrees of freedom, indicating significance at $0.05$ but not at $0.01$ (correct answer)
The test statistic is $9$ on $428$ degrees of freedom, indicating the added terms are not significant
The test statistic is $847$ on $2$ degrees of freedom, indicating significance at both $0.05$ and $0.01$

Explanation: When comparing two nested generalized linear models, the likelihood-ratio test (also called the deviance difference test) is your go-to tool. The test statistic is simply the difference in residual deviances between the reduced and full models, and the degrees of freedom equal the number of additional parameters in the full model. Here, the deviance difference is

428 - 419 = 9

, and the full model adds two slope parameters, giving

df = 2

. Under

H_0

, this statistic follows a

\chi^2_2

distribution. The critical values are

5.99

\alpha = 0.05

and

9.21

\alpha = 0.01

. Since

9 > 5.99

but

9 < 9.21

, the result is significant at

0.05

but not at

0.01

, confirming B is correct. Choice A is wrong on two counts: it halves the test statistic to

4.5

(perhaps confusing a two-sided z-test adjustment) and uses

df = 1

, ignoring that two parameters were added. Choice C correctly identifies the deviance difference as

9

but assigns

df = 428

, which is the reduced model's residual deviance — not a degrees-of-freedom value. Degrees of freedom for this test always reflect the number of extra parameters, not the deviance magnitude. Choice D sums the two deviances (

428 + 419 = 847

) instead of taking their difference — a fundamental misunderstanding of how the test statistic is constructed. Your rule of thumb: deviance difference, not deviance sum or raw deviance, tested against

\chi^2

with degrees of freedom equal to the parameter count difference. Always match your critical value to the correct

\alpha

threshold.

Question 8

A dataset of annual hospital-visit counts contains more zeros and a heavier upper tail than predicted by a fitted Poisson regression. A colleague concludes that the population must consist of a group that can never visit a hospital and another group following a Poisson process.

Which assessment of the colleague's conclusion is most defensible?

The conclusion is established because any excess of observed zeros identifies a structural-zero subpopulation
The conclusion is established only if the sample variance exceeds the sample mean after covariate adjustment
The Poisson model may be inadequate, but these features alone do not identify a structural-zero mechanism (correct answer)
The Poisson model remains adequate because fitted covariates automatically account for both excess zeros and heavy tails

Explanation: When you encounter questions about count data with excess zeros and overdispersion, your first instinct should be to separate two distinct problems: model fit and mechanism identification. These are not the same thing, and conflating them is exactly the trap this question sets. Excess zeros and a heavy upper tail together signal that a standard Poisson model is misspecified — but they do not, by themselves, tell you why. Two very different data-generating processes can produce identical symptom profiles: a zero-inflated Poisson (ZIP) model (which assumes a structural-zero subpopulation) and a negative binomial model (which assumes unobserved heterogeneity in a single population) can both fit data with excess zeros and overdispersion. Without additional subject-matter reasoning or formal model comparison, you cannot distinguish between them from symptoms alone. This makes C the most defensible answer — the Poisson is likely inadequate, but the observed features don't uniquely identify a structural-zero mechanism. A is wrong because excess zeros are a necessary but not sufficient condition for a structural-zero subpopulation. Overdispersion or mixing can also inflate the zero count without any "never-visit" group existing. B is wrong because variance exceeding the mean (

\text{Var}(Y) > \bar{Y}

) indicates overdispersion generally, not a zero-inflation mechanism specifically. Many models accommodate overdispersion without invoking two subpopulations. D is wrong in the opposite direction — fitted covariates reduce unexplained variation but cannot fix fundamental distributional misspecification. Residual overdispersion and zero-inflation persist even after covariate adjustment if the likelihood family is wrong. Study tip: On count-data questions, always ask: does the symptom have a unique cause? Excess zeros and heavy tails are symptoms, not diagnoses — multiple mechanisms can produce them.

Question 9

In a Poisson regression with a log link, the estimated coefficient for pollution exposure $x$ is $0.07$ per exposure unit. No interactions involving $x$ are included.

Holding the other covariates fixed, how does the expected count change when exposure increases from $20$ to $30$ units?

It increases by approximately $7\%$ because the fitted coefficient is interpreted per comparison
It increases by approximately $70\%$ because the exposure rises by ten units
It increases by approximately $101\%$ because the count ratio is $\exp(0.70)$ (correct answer)
It increases by approximately $201\%$ because the count ratio itself is converted directly to a percentage

Explanation: Whenever you encounter a Poisson regression question involving a change across multiple units of a covariate, your first instinct should be to work with the multiplicative structure of the log-linear model. The model specifies

\log(\mu) = \beta_0 + \beta_1 x + \cdots

, which means

\mu = \exp(\beta_0 + \beta_1 x + \cdots)

. When

x

increases by

\Delta x

, the expected count is multiplied by

\exp(\beta_1 \cdot \Delta x)

— not

\beta_1 \cdot \Delta x

itself. Here,

\beta_1 = 0.07

and

\Delta x = 30 - 20 = 10

, so the count ratio is

\exp(0.07 \times 10) = \exp(0.70) \approx 1.997

. This means the expected count roughly doubles, corresponding to an increase of about

97\% \approx 101\%

— confirming that C is correct. Choice A is wrong because it treats

\beta_1

as a simple per-unit percentage and ignores the ten-unit change entirely. A coefficient of

0.07

gives approximately

7\%

per one unit, not per the full ten-unit shift. Choice B correctly multiplies by 10 but then treats the result (

0.70

) as a direct percentage change, confusing the log-scale quantity with the multiplicative factor on the original count scale. Choice D takes the ratio

\exp(0.70) \approx 2.01

and converts it directly to

201\%

, forgetting to subtract 1 — the ratio represents the new count relative to the old, so the change is

(\text{ratio} - 1) \times 100\%

. Study tip: Always remember the two-step process — compute

\exp(\beta \cdot \Delta x)

for the ratio, then subtract 1 to get the percent change. Mixing up the ratio and the change is the most common trap in Poisson regression interpretation questions.

Question 10

Suppose unobserved multiplicative heterogeneity causes otherwise independent Poisson rates to vary among individuals. A negative binomial regression is considered with the same log-linear mean model as the original Poisson regression and variance $\operatorname{Var}(Y_i\mid x_i)=\mu_i+\kappa\mu_i^2$ , where $\kappa>0$ .

Which statement best compares this negative binomial model with the original Poisson model?

It preserves the modeled conditional mean and rate-ratio interpretation while allowing variance to grow faster than the mean (correct answer)
It preserves equidispersion but changes the link from logarithmic to logistic to accommodate additional zeros
It forces every slope coefficient toward zero because heterogeneity must be absorbed entirely by the regression mean
It changes the response from a count to a continuous rate, so exposure offsets are no longer appropriate

Explanation: When a Poisson model is misspecified due to unobserved multiplicative heterogeneity, the negative binomial (NB) model offers a natural fix — and questions like this test whether you understand what changes and what stays the same in that transition. The NB model retains the same log-linear mean structure:

\log(\mu_i) = x_i^\top \beta

. This means the conditional mean is still

E(Y_i \mid x_i) = \mu_i

, and exponentiated coefficients still represent multiplicative rate ratios — the same substantive interpretation as Poisson. What changes is the variance:

\operatorname{Var}(Y_i \mid x_i) = \mu_i + \kappa\mu_i^2

. Because

\kappa > 0

, the variance grows quadratically with the mean rather than linearly (as in Poisson's equidispersion condition

\operatorname{Var} = \mu

). This quadratic growth accommodates overdispersion, making A the correct answer. B is wrong on two counts: NB does not preserve equidispersion (that's the whole point of switching models), and the link function remains logarithmic, not logistic. Logistic links belong to binary/binomial models. C is backwards. Unobserved heterogeneity is absorbed into the variance through

\kappa

, not into the regression coefficients. The slopes are not systematically shrunk toward zero; in fact, NB coefficient estimates are often more reliable than Poisson when overdispersion is present. D confuses model structure with outcome type. The response remains a count, and exposure offsets (e.g.,

\log(t_i)

) are just as appropriate in NB regression as in Poisson regression. Study tip: Always separate the mean model from the variance model. The NB extension targets the variance, leaving the mean — and thus the linear predictor and link function — completely intact.