{
  "id": "poisson-regression",
  "title": "Poisson Regression",
  "lang": "en",
  "source_lang": "en",
  "summary": "Poisson regression models nonnegative counts or conditional means through a logarithmic link, relating expected outcomes to explanatory variables and exposure.",
  "category": "mathematics",
  "subcategory": "probability-statistics",
  "type": "concept",
  "aliases": [
    "Poisson log-linear regression",
    "Poisson log-linear model"
  ],
  "generated_at": "2026-11-04",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/poisson-regression/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/poisson-regression/",
      "json": "https://aiwiki.page/zh/poisson-regression.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "Poisson regression is a statistical method for relating a nonnegative count to explanatory variables. In its standard form, the response follows a Poisson distribution conditional on the predictors, and the logarithm of its expected value is a linear combination of those predictors. It is a principal example of a generalized linear model, extending regression analysis to outcomes such as the number of accidents at a road junction, defects produced during a manufacturing shift, or events observed over a specified period.\n\nThe method is especially useful when observations represent different amounts of time, population, distance, or other exposure. Incorporating exposure directly into the model allows comparisons of event rates rather than raw counts. Its coefficients describe multiplicative changes in expected counts or rates, unlike the additive changes ordinarily represented by linear regression.\n\nPoisson regression also has a broader interpretation as an estimating method for a log-linear conditional mean. Under suitable conditions, its coefficient estimates remain consistent even when the response does not have a Poisson distribution. This distinction between a full probability model and a conditional-mean model underlies its use in epidemiology, econometrics, demography, insurance, and other fields, as well as the development of robust inference and alternative count models."
    },
    {
      "heading": "Probability model and mathematical structure",
      "text": "Let \\(Y_i\\) denote the response for observation \\(i\\), and let \\(\\mathbf{x}_i\\) contain its explanatory variables, usually including a constant for an intercept. Standard Poisson regression specifies\n\n\\[Y_i\\mid\\mathbf{x}_i\\sim\\operatorname{Poisson}(\\mu_i), \\qquad \\log\\mu_i=\\mathbf{x}_i^\\top\\boldsymbol{\\beta}.\\]\n\nHere \\(\\boldsymbol{\\beta}\\) is a vector of unknown coefficients, and\n\n\\[\\mu_i=E(Y_i\\mid\\mathbf{x}_i) =\\exp(\\mathbf{x}_i^\\top\\boldsymbol{\\beta}).\\]\n\nThe exponential function ensures that fitted means are positive for every finite linear predictor. Observed responses, however, may equal zero. The model concerns the logarithm of the conditional mean, not the logarithm of each observed response.\n\nThe conditional probability mass function is\n\n\\[P(Y_i=y_i\\mid\\mathbf{x}_i) = \\frac{\\exp(-\\mu_i)\\mu_i^{y_i}}{y_i!}, \\qquad y_i=0,1,2,\\ldots.\\]\n\nA defining property of this distribution is conditional equidispersion:\n\n\\[\\operatorname{Var}(Y_i\\mid\\mathbf{x}_i)=\\mu_i.\\]\n\nThis equality is stronger than the mean specification alone. It determines the model’s conditional uncertainty and the probabilities assigned to individual counts. For example, an observation with mean 2 has variance 2 and probability \\(e^{-2}\\), approximately 0.1353, of producing zero events.\n\nThe logarithm is the canonical link for the Poisson generalized linear model. Writing \\(\\eta_i=\\log\\mu_i\\), the log probability becomes\n\n\\[y_i\\eta_i-\\exp(\\eta_i)-\\log(y_i!).\\]\n\nThe parameter-dependent part has a particularly simple form, which leads to convenient likelihood derivatives and stable optimization procedures. Other links can be paired with a Poisson response distribution, but identity and square-root links require additional care to prevent invalid negative fitted means. Unqualified references to Poisson regression usually mean the logarithmic-link model.\n\nConditional equidispersion should not be confused with equality of the overall sample mean and variance. By the law of total variance,\n\n\\[\\operatorname{Var}(Y) = E\\{\\operatorname{Var}(Y\\mid X)\\} + \\operatorname{Var}\\{E(Y\\mid X)\\}.\\]\n\nEven under a correct conditional Poisson model, the second term makes the marginal variance exceed the marginal mean when expected counts differ across predictor values. A heterogeneous population can therefore exhibit substantial overall dispersion without violating the conditional model.\n\nStandard likelihood calculations also generally assume that observations are conditionally independent. A Poisson marginal distribution for each observation does not establish independence between observations. Repeated measurements, shared environments, or spatial proximity can create dependence that must be represented separately."
    },
    {
      "heading": "Coefficients, exposure, and comparisons of rates",
      "text": "The logarithmic link makes coefficient interpretations multiplicative. For a predictor \\(x_j\\) entering without interactions or nonlinear transformations,\n\n\\[\\frac{\\mu(x_j+1,\\mathbf{x}_{-j})} {\\mu(x_j,\\mathbf{x}_{-j})} = \\exp(\\beta_j).\\]\n\nThus \\(\\exp(\\beta_j)\\) is the ratio of expected counts associated with a one-unit increase in \\(x_j\\), holding the remaining predictors fixed. A coefficient of 0.2 corresponds to a ratio of approximately 1.221, or a 22.1% increase. The expression \\(100\\beta_j\\%\\) is only a small-coefficient approximation to the exact percentage change \\(100\\{\\exp(\\beta_j)-1\\}\\%\\).\n\nFor a binary predictor, exponentiating its coefficient compares the expected count in the category coded 1 with that in the reference category coded 0. For a continuous predictor changed by \\(\\Delta\\), the corresponding ratio is \\(\\exp(\\Delta\\beta_j)\\). The change on the original count scale depends on the baseline mean:\n\n\\[\\frac{\\partial\\mu_i}{\\partial x_{ij}} = \\beta_j\\mu_i.\\]\n\nConsequently, a constant multiplicative effect produces larger absolute count differences where expected counts are already high.\n\nWhen observations have different exposure \\(t_i>0\\), the usual model is\n\n\\[\\mu_i=t_i\\lambda_i, \\qquad \\log\\mu_i=\\log t_i+\\mathbf{x}_i^\\top\\boldsymbol{\\beta}.\\]\n\nHere \\(\\lambda_i\\) is the expected rate per unit exposure. The term \\(\\log t_i\\) is an offset: its coefficient is fixed at 1 rather than estimated. Examples include person-years in population studies, vehicle-kilometres in transport analysis, and operating hours in equipment reliability. If a model implies a rate of 0.004 events per unit exposure, exposures of 500 and 2,000 units imply expected counts of 2 and 8, respectively.\n\nAn offset embodies the substantive assumption that expected counts are proportional to exposure, conditional on other variables. It does not merely adjust for a nuisance measurement. If doubling exposure does not plausibly double the expected count, an estimated exposure effect or a more detailed model may be needed. Zero exposure ordinarily implies a structurally zero count and cannot be inserted into a logarithmic offset without separate handling.\n\nA statistical interaction changes these interpretations. In a model containing \\(\\beta_1x+\\beta_2z+\\beta_3xz\\), a one-unit increase in \\(x\\) has expected-count ratio \\(\\exp(\\beta_1+\\beta_3z)\\). There is then no single ratio that describes the effect of \\(x\\) at every value of \\(z\\).\n\nRate ratios are not automatically causal effects. Differences in exposure measurement, confounding, selection, or omitted predictors can explain observed associations. A causal interpretation requires an appropriate design and identification assumptions beyond the Poisson regression equation itself."
    },
    {
      "heading": "Likelihood estimation and computational methods",
      "text": "For conditionally independent observations, maximum likelihood estimation maximizes the log-likelihood\n\n\\[\\ell(\\boldsymbol{\\beta}) = \\sum_{i=1}^{n} \\left[ y_i\\mathbf{x}_i^\\top\\boldsymbol{\\beta} -\\exp(\\mathbf{x}_i^\\top\\boldsymbol{\\beta}) -\\log(y_i!) \\right].\\]\n\nWith an offset \\(o_i=\\log t_i\\), the linear predictor becomes \\(o_i+\\mathbf{x}_i^\\top\\boldsymbol{\\beta}\\). The score equations retain the same form:\n\n\\[U(\\boldsymbol{\\beta}) = X^\\top(\\mathbf{y}-\\boldsymbol{\\mu}) = \\mathbf{0},\\]\n\nwhere \\(X\\) is the design matrix. In general, these nonlinear equations have no closed-form solution. In an intercept-only model without exposure, however, the fitted mean equals the sample mean. With exposure and only an intercept, the estimated rate is \\(\\sum_i y_i/\\sum_i t_i\\).\n\nThe Hessian matrix of the log-likelihood is\n\n\\[H(\\boldsymbol{\\beta})=-X^\\top W X, \\qquad W=\\operatorname{diag}(\\mu_1,\\ldots,\\mu_n).\\]\n\nThe log-likelihood is concave, so an interior maximum, when it exists and the design has full column rank, is unique. Concavity does not guarantee that a finite maximum exists. If every response is zero, for example, the likelihood increases as the intercept tends to negative infinity. Related boundary behavior can occur when predictor patterns perfectly distinguish zero outcomes from positive outcomes.\n\nA Newton–Raphson update is\n\n\\[\\boldsymbol{\\beta}^{(k+1)} = \\boldsymbol{\\beta}^{(k)} + (X^\\top W^{(k)}X)^{-1} X^\\top(\\mathbf{y}-\\boldsymbol{\\mu}^{(k)}).\\]\n\nFor the canonical Poisson model, the observed and expected information matrices coincide, making Newton–Raphson and Fisher scoring identical. In actual computation, linear systems are solved rather than forming the inverse explicitly.\n\nThe same update can be expressed through iteratively reweighted least squares. At each iteration, the algorithm constructs the working response\n\n\\[z_i=\\eta_i+\\frac{y_i-\\mu_i}{\\mu_i}\\]\n\nand solves a weighted least-squares problem with weights \\(\\mu_i\\). When offsets are present, the fixed offset is subtracted from the working response in the coefficient update. Iteration continues until likelihood changes, coefficient changes, or score magnitudes satisfy convergence criteria.\n\nNumerical difficulties can result from nearly dependent predictors, very large linear predictors, or sparse categories. Exponentiation can overflow, while nearly singular weighted design matrices can produce unstable updates. Scaling continuous predictors, using stable matrix factorizations, and step control address distinct computational problems, but cannot resolve a fundamentally unidentified model.\n\nRegularization modifies the optimization criterion, for example by adding a quadratic coefficient penalty or an absolute-value penalty. These penalties can stabilize high-dimensional estimation and perform shrinkage or variable selection. They also change the estimator and its inferential interpretation: uncertainty formulas derived for an unpenalized interior maximum do not automatically apply."
    },
    {
      "heading": "Statistical inference and the conditional-mean interpretation",
      "text": "Under a correctly specified independent Poisson model and standard large-sample conditions, the estimated coefficient vector is approximately multivariate normal. Its model-based estimated covariance matrix is\n\n\\[\\widehat{\\operatorname{Var}}_{\\mathrm{model}} (\\widehat{\\boldsymbol{\\beta}}) = (X^\\top\\widehat W X)^{-1}.\\]\n\nThe diagonal entries provide squared standard errors. A large-sample 95% confidence interval for a coefficient is approximately \\(\\widehat\\beta_j\\pm1.96\\,\\operatorname{SE}(\\widehat\\beta_j)\\). Exponentiating its endpoints gives an interval for the corresponding count or rate ratio.\n\nWald, score, and likelihood-ratio procedures provide related approaches to hypothesis testing. A likelihood-ratio statistic compares maximized log-likelihoods for nested models:\n\n\\[2\\{\\ell(\\widehat{\\boldsymbol{\\beta}}_{\\mathrm{full}}) -\\ell(\\widehat{\\boldsymbol{\\beta}}_{\\mathrm{restricted}})\\}.\\]\n\nUnder suitable regularity conditions, it has an asymptotic chi-squared distribution with degrees of freedom equal to the number of restrictions. Small samples, weakly identified coefficients, boundary parameters, and sparse data can make these approximations inaccurate.\n\nA broader justification comes from the score equation itself. If the true conditional mean satisfies\n\n\\[E(Y_i\\mid\\mathbf{x}_i) = \\exp(\\mathbf{x}_i^\\top\\boldsymbol{\\beta}_0),\\]\n\nthen\n\n\\[E\\left[ \\mathbf{x}_i \\{Y_i-\\exp(\\mathbf{x}_i^\\top\\boldsymbol{\\beta}_0)\\} \\right]=\\mathbf{0}.\\]\n\nAccordingly, solving the Poisson score equations can give a consistent estimator of \\(\\boldsymbol{\\beta}_0\\) even when the conditional distribution is not Poisson. This use is commonly called Poisson pseudo-maximum likelihood or Poisson quasi-maximum likelihood.\n\nCorrect mean specification does not make the usual Poisson standard errors valid. For independent observations, a sandwich covariance estimator takes the form\n\n\\[\\widehat V = A^{-1}BA^{-1},\\]\n\nwhere\n\n\\[A=\\sum_i\\widehat\\mu_i\\mathbf{x}_i\\mathbf{x}_i^\\top, \\qquad B=\\sum_i (y_i-\\widehat\\mu_i)^2 \\mathbf{x}_i\\mathbf{x}_i^\\top.\\]\n\nThis estimator allows the conditional variance to differ from the conditional mean, subject to appropriate moment and asymptotic conditions. Dependence requires additional changes, such as forming the middle matrix from cluster-level score sums.\n\nThe distinction is important for prediction. A well-specified mean model can estimate expected outcomes while assigning incorrect probabilities to zero, large counts, or future observations. Robust standard errors repair neither distributional predictions nor an incorrect conditional mean. Similarly, likelihood-ratio tests calibrated for a true Poisson likelihood are not automatically justified when that likelihood is used only as a quasi-likelihood.\n\nRobust Poisson methods have also been used with binary outcomes to estimate risk ratios. In that setting the response variance is Bernoulli rather than Poisson, and fitted means are not inherently constrained below 1. The estimating approach therefore differs from claiming that a binary response actually follows a Poisson distribution."
    },
    {
      "heading": "Model assessment, residuals, and prediction",
      "text": "Assessing a Poisson regression requires examining its mean structure, conditional variance, dependence assumptions, and predictive distribution separately. A model can predict average counts well while underestimating extreme observations or uncertainty. Conversely, apparent excess variation may arise because the mean omits important predictors, nonlinearities, or interactions rather than because a different count distribution is intrinsically necessary.\n\nPearson residuals standardize deviations using the Poisson variance:\n\n\\[r_{Pi} = \\frac{y_i-\\widehat\\mu_i}{\\sqrt{\\widehat\\mu_i}}.\\]\n\nThe Pearson statistic \\(X^2=\\sum_i r_{Pi}^2\\), divided by the residual degrees of freedom \\(n-p\\), is often used as a rough dispersion diagnostic, where \\(p\\) is the number of estimated coefficients. Values substantially above 1 can indicate excess variation. Its interpretation is less straightforward with sparse counts, high leverage, many fitted parameters, or a misspecified mean, so it is not a universal test of the Poisson assumption.\n\nThe Poisson deviance compares the fitted model with a saturated model that assigns each observation its own fitted mean:\n\n\\[D = 2\\sum_i \\left[ y_i\\log\\left(\\frac{y_i}{\\widehat\\mu_i}\\right) -(y_i-\\widehat\\mu_i) \\right],\\]\n\nwhere \\(0\\log(0/\\widehat\\mu_i)\\) is defined as zero. Signed square roots of the individual contributions are deviance residuals. They are often more symmetric than raw residuals, although their distributions remain discrete and need not resemble a normal distribution when means are small.\n\nResidual plots against fitted values, exposure, predictors, time, and location can reveal different failures. Curvature may suggest an inadequate functional form; increasing residual spread may suggest a variance problem; sequential patterns may indicate dependence. Unusually influential observations combine discrepancy with leverage, so a large raw count alone does not establish that an observation dominates the fit.\n\nPredictive checks compare observed features with those generated under the fitted model. Relevant features include the number of zeros, the upper tail, group totals, and variation within predictor strata. Under a Poisson model,\n\n\\[P(Y_i=0\\mid\\mathbf{x}_i)=e^{-\\mu_i}.\\]\n\nThe expected number of zeros in a dataset is therefore \\(\\sum_i e^{-\\widehat\\mu_i}\\), not generally \\(n e^{-\\overline{\\widehat\\mu}}\\). Variation in fitted means matters when judging whether zeros are unusually frequent.\n\nA confidence interval for \\(\\mu_i\\) describes uncertainty about an expected count. A prediction interval for a future response additionally includes the response’s intrinsic count variability and is generally wider. Predictive intervals based only on a fitted Poisson distribution can be too narrow when overdispersion is present.\n\nCross-validation can compare predictive performance using criteria such as held-out log probability, Poisson deviance, or squared prediction error. The split structure must match the prediction task: randomly splitting observations from the same subject or neighboring time periods can overstate performance when the intended application concerns new subjects or future periods."
    },
    {
      "heading": "Overdispersion, heterogeneity, and dependent observations",
      "text": "Overdispersion means that conditional variability exceeds that allowed by the reference distribution. In a Poisson regression context, this usually means\n\n\\[\\operatorname{Var}(Y_i\\mid\\mathbf{x}_i)>\\mu_i.\\]\n\nPossible sources include unobserved differences in event propensity, clustering, contagion, omitted predictors, measurement error, and mixtures of distinct processes. Underdispersion, in which variance is smaller than the mean, also occurs, especially when events are constrained or inhibited. Neither phenomenon is adequately described by comparing the unadjusted sample variance with the unadjusted sample mean.\n\nA quasi-Poisson specification assumes\n\n\\[E(Y_i\\mid\\mathbf{x}_i)=\\mu_i, \\qquad \\operatorname{Var}(Y_i\\mid\\mathbf{x}_i)=\\phi\\mu_i.\\]\n\nFor a common dispersion parameter \\(\\phi\\), coefficient estimates ordinarily coincide with those from Poisson regression using the same mean structure, while model-based standard errors are multiplied by \\(\\sqrt{\\phi}\\). Quasi-Poisson analysis specifies a mean–variance relationship rather than a complete probability distribution. It therefore does not directly supply probabilities for individual counts or an ordinary full-likelihood information criterion.\n\nNegative binomial regression supplies a full alternative distribution. One common construction introduces an unobserved multiplicative factor \\(U_i\\):\n\n\\[Y_i\\mid U_i,\\mathbf{x}_i \\sim\\operatorname{Poisson}(\\mu_iU_i),\\]\n\nwith \\(U_i\\) gamma distributed, having mean 1 and variance \\(\\alpha\\). Integrating over this heterogeneity gives\n\n\\[E(Y_i\\mid\\mathbf{x}_i)=\\mu_i, \\qquad \\operatorname{Var}(Y_i\\mid\\mathbf{x}_i) = \\mu_i+\\alpha\\mu_i^2.\\]\n\nThis quadratic variance function differs from quasi-Poisson’s linear one. Negative binomial estimation generally changes coefficient estimates as well as standard errors because its score equations weight observations differently.\n\nDependence raises a separate issue. Repeated observations on a person, hospital, firm, or geographic area may remain correlated after observed predictors have been included. Cluster-robust standard errors can accommodate within-cluster dependence in large-sample inference when clusters are independent and sufficiently numerous. They do not explicitly model the correlation structure or improve the fitted distribution of individual outcomes.\n\nGeneralized estimating equations specify a marginal mean and a working within-cluster correlation structure. Poisson mixed models instead introduce random effects, for example\n\n\\[\\log\\mu_{ij} = \\mathbf{x}_{ij}^{\\top}\\boldsymbol{\\beta}+b_j.\\]\n\nWhen \\(b_j\\) is normally distributed, the resulting marginal count distribution is a Poisson–lognormal mixture. Conditional, subject-specific coefficients need not have the same interpretation as marginal, population-average coefficients.\n\nIn time-series and spatial data, residual dependence may persist over many observations rather than within a simple set of isolated clusters. Models may use lagged covariates, evolving latent effects, or structured spatial components. The relevant choice concerns both the expected count and how uncertainty propagates between observations; changing only the marginal distribution does not necessarily address dependence."
    },
    {
      "heading": "Excess zeros, truncation, and observation mechanisms",
      "text": "A large number of zeros does not by itself establish that Poisson regression is inappropriate. Small fitted means naturally imply high zero probabilities, and mixtures of observations with different means can produce many zeros even if each conditional distribution is Poisson. The relevant comparison is between observed zeros and the zeros predicted after accounting for covariates and exposure.\n\nA zero-inflated Poisson model introduces two latent states. With probability \\(\\pi_i\\), an observation belongs to a state that necessarily produces zero. With probability \\(1-\\pi_i\\), it follows a Poisson distribution with mean \\(\\lambda_i\\):\n\n\\[P(Y_i=0) = \\pi_i+(1-\\pi_i)e^{-\\lambda_i},\\]\n\n\\[P(Y_i=y) = (1-\\pi_i)e^{-\\lambda_i} \\frac{\\lambda_i^y}{y!}, \\qquad y\\geq1.\\]\n\nIts unconditional mean is \\((1-\\pi_i)\\lambda_i\\), not \\(\\lambda_i\\). Predictors may enter both the zero-state probability and the count component. A coefficient in the count component consequently describes the Poisson-state mean rather than necessarily the overall observed mean.\n\nA hurdle model separates the occurrence of any event from the number of events once a positive count occurs. If \\(q_i=P(Y_i>0)\\), a Poisson hurdle model has\n\n\\[P(Y_i=0)=1-q_i,\\]\n\n\\[P(Y_i=y) = q_i\\frac{e^{-\\lambda_i}\\lambda_i^y} {y!(1-e^{-\\lambda_i})}, \\qquad y\\geq1.\\]\n\nUnlike a zero-inflated model, its positive-count component cannot generate zero. Hurdle models can represent processes in which initiating an activity differs from determining its subsequent intensity. The distinction is a model of the data-generating process, not merely a technical adjustment for the observed zero count.\n\nTruncation concerns which observations enter the dataset. In a zero-truncated sample, observations with zero events are absent altogether. The probability for an observed positive count must therefore be divided by \\(1-e^{-\\mu_i}\\). Examples arise when sampling is based on having used a service at least once. Fitting an ordinary Poisson model to such data ignores the selection rule and generally distorts inference about the original population.\n\nCensoring is different: an observation is present, but its exact count is only partially known. If counts are reported as “10 or more,” the likelihood contribution for that category is \\(P(Y_i\\geq10)\\), not \\(P(Y_i=10)\\). Rounding, reporting thresholds, and detection failures likewise affect the observation model.\n\nThere is also a fundamental identification limit. Independent detection of each event with probability \\(p\\) transforms a Poisson count with mean \\(\\mu\\) into another Poisson count with mean \\(p\\mu\\). Without additional information about detection, recorded counts alone cannot generally distinguish a low underlying event rate from a high rate observed incompletely. Such problems cannot be resolved merely by adding a zero-inflation component."
    },
    {
      "heading": "Historical development and major applications",
      "text": "The mathematical foundation is the Poisson distribution, introduced by Siméon-Denis Poisson in 1837 in his work on probability. Regression models for counts subsequently connected this distribution with systematic variation in expected frequencies. John Nelder and Robert Wedderburn’s 1972 formulation of generalized linear models placed Poisson regression within a common framework alongside normal linear models and binomial regression, unifying distributional assumptions, link functions, and iterative estimation.\n\nPopulation studies made log-linear Poisson models especially useful because event counts are naturally paired with exposure. In demography, births or deaths may be modeled within age, sex, and calendar-period groups, with the corresponding population exposure as an offset. In epidemiological rate analysis, person-time serves the same role. Such models distinguish a larger number of events caused by greater exposure from a larger underlying event rate.\n\nPoisson regression is closely related to piecewise-exponential models for event times. If a hazard is constant within specified intervals, event-time data can be reorganized into interval-specific event counts and time at risk. The resulting likelihood has the same parameter-dependent form as a Poisson likelihood with a log-time offset. This connection permits time-varying baseline hazards and covariates to be represented through a count-model design, provided the interval construction and risk exposure are correct.\n\nIn contingency-table analysis, Poisson log-linear models express expected cell frequencies through combinations of categorical factors. Interactions encode departures from forms of independence among table dimensions. Under appropriate sampling arrangements and conditioning, these models connect with multinomial analysis and logistic regression. Their purpose is then to describe relationships among categorical variables through expected cell counts rather than to model counts observed over time.\n\nEconometric applications emphasize the conditional-mean interpretation. A 1984 paper by Christian Gouriéroux, Alain Monfort, and Alain Trognon developed the theory of pseudo-maximum likelihood estimation. In 2006, João Santos Silva and Silvana Tenreyro demonstrated the importance of Poisson pseudo-maximum likelihood for multiplicative trade models. With a specified exponential conditional mean, the method can be applied to nonnegative trade values even though those values are not integer counts.\n\nThis application avoids replacing zero outcomes with arbitrary positive constants before taking logarithms. It also addresses a separate issue: under heteroskedasticity, a regression for \\(E(\\log Y\\mid X)\\) generally does not recover the parameters of \\(\\log E(Y\\mid X)\\). Poisson pseudo-maximum likelihood estimates the latter mean relationship directly, although it still requires correct mean specification and suitable inference.\n\nOther uses include insurance claim frequency, manufacturing defects, ecological abundance, and transport incidents. In each, the statistical specification depends on what constitutes an observation, how exposure is measured, whether events cluster, and whether zero counts are observable. Insurance claim counts, for example, may use policy duration as exposure, whereas ecological counts may require an additional detection model because the number recorded can differ systematically from the number present."
    }
  ],
  "links": [
    "poisson-distribution",
    "generalized-linear-model",
    "linear-regression",
    "epidemiology",
    "econometrics",
    "exponential-function",
    "probability-mass-function",
    "offset-statistics",
    "statistical-interaction",
    "confounding",
    "maximum-likelihood-estimation",
    "matrix",
    "sample-mean",
    "hessian-matrix",
    "newtons-method",
    "iteratively-reweighted-least-squares",
    "regularization",
    "covariance-matrix",
    "standard-error",
    "confidence-interval",
    "statistical-hypothesis-testing",
    "chi-squared-distribution",
    "consistent-estimator",
    "prediction-interval",
    "cross-validation",
    "overdispersion",
    "time-series",
    "statistical-independence",
    "logistic-regression"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}