The Gold Standard of Zero-Inflated Models | Why Matching the Mean and Zero Fraction Does Not Match a Count Mixture

Let Y be a nonnegative integer count over a fixed positive exposure. Imagine that each observational unit belongs to one of two rate groups, with conditional Poisson means a and b. Both a and b are strictly positive.

The group label is latent in the first calculation: the record contains Y but not its rate-group membership. We stipulate equal group probabilities initially. This is a mathematical mixture, not evidence that an actual population has exactly two biological classes.

A zero is possible in either group because a Poisson distribution assigns positive probability to zero at every finite mean. No component in this construction is forced to produce zero. That distinction makes it a useful comparison with a zero-inflated representation, which explicitly includes a point-mass component at zero. We will keep the stipulated generative descriptions separate from the observable probabilities they imply.

Contents · Next chapter

A Poisson distribution with mean λ has mass exp(−λ)λ^k/k! at count k≥0. With equal weights, our rate-mixture mass is P(Y=k)=[exp(−a)a^k+exp(−b)b^k]/(2k!).

Every probability comes from this expression, including P(Y=0). The mean is μ=(a+b)/2. The total variance is μ+(a−b)²/4, obtained by adding the average conditional Poisson variance to the variance of the conditional means.

This last formula shows overdispersion relative to a homogeneous Poisson with the same mean when a≠b. It does not prove the two-rate mechanism from a dataset’s observed variance. Many observation processes can generate similar summaries. Our task is narrower: use an explicitly defined distribution as a counterexample to the idea that its mean and zero fraction uniquely determine its positive-count shape.

Contents · Previous chapter · Next chapter

Set a=ln 2 and b=ln 8=3 ln 2. Denote ln 2 by L when this shortens an equation. Then exp(−a)=1/2 and exp(−b)=1/8.

The unconditional mean is μ=2L, approximately 1.386294. The zero probability is q=(1/2)(1/2+1/8)=5/16=0.3125. Both rates are positive; even the lower-rate component can produce any nonnegative integer count.

These logarithmic values were chosen to make the zero calculation exact. The high-rate component is not a high-count cutoff: Poisson distributions overlap. A low count does not reveal the rate group with certainty, and a high count is not impossible in the lower group. The model assigns probabilities rather than deterministic labels to observed counts. We will exploit that changing group weighting to understand the positive-count shape.

Contents · Previous chapter · Next chapter

A homogeneous Poisson distribution with mean μ=2L has zero probability exp(−2L)=1/4. Our rate mixture has the same mean but zero probability 5/16. The difference is 1/16.

Thus the mixture has more zeros than a Poisson model formed by inserting the pooled mean. This discrepancy does not require a structural-zero state: it was generated entirely by two positive conditional rates.

The calculation is a specific version of the root guide’s warning about omitted rate heterogeneity. It is not yet our main result, because the existing ecosystem already explains that general warning. The next step is to let a ZIP model match both the mean and the zero probability and ask whether that stronger agreement matches the entire count law. A model can repair two diagnostic summaries and still miss information in the positive observations.

Contents · Previous chapter · Next chapter

For a random nonnegative Poisson mean Λ, the mixed zero probability is E[exp(−Λ)]. A homogeneous Poisson using the pooled mean instead gives exp[−E(Λ)].

The function exp(−x) is strictly convex, so Jensen’s inequality gives E[exp(−Λ)]≥exp[−E(Λ)]. The inequality is strict for a nonconstant rate distribution under the usual finite-mean conditions. Our two-rate example supplies an exact numerical instance.

This is a comparison between two specified probability models. It does not establish that any particular dataset’s excess zeros arose from heterogeneity, or that a mixture is the only alternative to ZIP. Detection, exposure and event definition still need examination. Convexity identifies a possible source of a discrepancy; it does not identify the source of an observed discrepancy without further evidence.

Contents · Previous chapter · Next chapter

A zero-inflated Poisson model has probability π of an additional zero-generating state and probability 1−π of an ordinary Poisson component with mean λ. Its overall mean is (1−π)λ.

Its zero probability is π+(1−π)exp(−λ), while its positive mass at k≥1 is (1−π)exp(−λ)λ^k/k!. The count component can still generate zero, so π is not the total zero fraction.

We seek π and λ that reproduce the original mixture’s μ and q. This is a summary-matching construction, not a maximum-likelihood fit to simulated or observed counts. It will be useful precisely because we can compare the resulting full distributions without estimation noise. If their positive masses differ, that difference reflects their mathematical laws rather than an optimizer, finite sample or plotting choice.

Contents · Previous chapter · Next chapter

Matching the mean requires π=1−μ/λ. A valid nonnegative π therefore requires λ≥μ, with π<1 for positive finite μ and λ.

Substitute this relation into the ZIP zero equation. The zero probability along the mean-matching curve is q_Z(λ)=1−[μ/λ][1−exp(−λ)]. We want q_Z(λ)=q.

At λ=μ, π=0 and q_Z=exp(−μ), the homogeneous Poisson value. As λ increases, a smaller fraction of observations enters a more intense Poisson component while preserving the same unconditional mean. This tradeoff can increase the total zero probability. Keeping the mean constraint visible prevents a common mistake: changing π while holding λ fixed and assuming the average count stayed unchanged. The two parameters must move together in this construction.

Contents · Previous chapter · Next chapter

Differentiate q_Z with respect to λ. Its derivative is μ[1−(λ+1)exp(−λ)]/λ², positive for λ>0 because exp(λ)>1+λ.

Thus q_Z increases strictly from exp(−μ) at λ=μ toward one as λ tends to infinity. Any target q strictly between exp(−μ) and one has exactly one finite mean-matching ZIP pair with π>0.

Our mixture has q=5/16 between 1/4 and one, so the match exists uniquely. This does not mean its latent interpretation is scientifically unique. It means two supplied summaries determine one parameter pair within this particular ZIP family. The rate-mixture construction already supplies a different model with those summaries and no point mass at rate zero. Parameter identification inside an assumed family and identification of the family itself are different questions.

Contents · Previous chapter · Next chapter

Solving the one-dimensional equation gives λ approximately 1.6156333274 and π approximately 0.1419498858. These decimal values are numerical solutions, not exact logarithmic identities.

Check the mean: (1−π)λ returns 2 ln 2. Check the zero mass: π+(1−π)exp(−λ) returns 5/16. A simple bisection routine can bracket the solution above μ and reduce the interval until a declared tolerance is reached.

Store the matching equations and numerical tolerance rather than merely copying the displayed decimals. Repeated rounding can make the two summary checks differ slightly even when the exact construction is valid. A small numerical mismatch at the last decimal should not be confused with the substantial positive-count differences we will calculate next. The model comparison rests on complete formulas, with numerical values used as independent checks.

Contents · Previous chapter · Next chapter

For the original mixture, P(Y=1)=(1/4)L+(1/16)(3L)=7L/16, approximately 0.3032518915.

For the matched ZIP model, P(Y=1)=(1−π)exp(−λ)λ, approximately 0.2755464485. The probabilities differ by about 0.0277054430 despite agreeing on the mean and zero mass.

If one imagined many independent draws under either stipulated law, the expected fraction of ones would approach its corresponding probability. That is a prediction of the two models, not an observed study result. It already disproves full distributional equality. Matching the two summaries did not constrain enough of the positive shape. Checking a single positive mass can reveal the failure in this example, but the following ratio argument will explain the general structural distinction.

Contents · Previous chapter · Next chapter

For k≥1, define R_k=(k+1)P(Y=k+1)/P(Y=k), provided the denominator is positive. Multiplication by k+1 removes the factorial step in a Poisson mass.

For any ZIP with λ>0 and π<1, the common positive multiplier cancels, leaving R_k=λ for every positive k. The ratio does not depend on the inflation probability.

This is an identity of the supplied distribution, not a recommendation to estimate λ from a sparse pair of empirical bins. A small observed denominator can make a sample ratio highly unstable or undefined. As a mathematical check, however, the ratio cleanly separates the positive Poisson shape from its overall positive probability. The matched ZIP model predicts one constant value across all positive-count transitions.

Contents · Previous chapter · Next chapter

The mixture mass can be written P(Y=k)=L^k[4+3^k]/(16k!). Therefore R_k=L[4+3^(k+1)]/[4+3^k].

At k=1 this is 13L/7, approximately 1.287273. At k=2 it is 31L/13, approximately 1.652889. At k=3 it is 85L/31, approximately 1.900565. These are not equal to one another, so no ZIP positive component can reproduce the entire positive-count law.

The ratio approaches 3L as k grows, because the larger-rate component increasingly dominates the probabilities of very large counts. It does not jump to that rate at a fixed count threshold. This gradual change explains why a pooled positive Poisson shape can miss the relative frequencies of small and large positive counts even when its mean and total probability of zero have been matched.

Contents · Previous chapter · Next chapter

Write c_i=w_i exp(−λ_i) for a finite positive-rate mixture. Apart from k!, its count mass is S_k=Σ_i c_i λ_i^k.

Define weights α_i(k)=c_i λ_i^k/S_k. They are nonnegative and sum to one. Then R_k=S_(k+1)/S_k=Σ_i α_i(k)λ_i. The adjacent-count ratio is the mean rate under weights tilted toward components that can more readily produce count k.

In the original example, the relative high-rate to low-rate weight is 3^k/4. It grows with k even though the original group proportions remain one-half each. The change comes from conditioning on a count, not from changing the population composition. Confusing those two kinds of weights can lead a report to describe a selected high-count subset as if it represented the original mixture proportions.

Contents · Previous chapter · Next chapter

Under the α_i(k) weights, let the mean of λ_i be R_k and its variance be V_k. The next ratio satisfies R_(k+1)−R_k=V_k/R_k.

To derive it, write R_(k+1)=S_(k+2)/S_(k+1). Subtract S_(k+1)/S_k and combine denominators. The numerator is S_(k+2)S_k−S_(k+1)², which equals S_k² V_k. Dividing gives V_k/R_k.

For at least two distinct strictly positive component rates with positive weights, V_k>0 and R_k>0, so the sequence increases strictly. A ZIP’s positive ratios are constant. This proof shows why the distinction is structural rather than a numerical accident in our chosen logarithms. It also states the relevant limits: coincident rates or zero mixture weights collapse the positive-rate mixture to a simpler case.

Contents · Previous chapter · Next chapter

If a mixture component has rate exactly zero, it produces zero with probability one. A two-component mixture of a zero-rate Poisson and a positive-rate Poisson is exactly a ZIP representation.

For positive counts k≥1, the zero-rate term contributes nothing. The positive adjacent ratios are therefore constant even though the overall rate distribution is not constant. That boundary is why the increasing-ratio proof explicitly required distinct strictly positive rates with positive weights.

Our original a=ln 2 and b=ln 8 satisfy those conditions. Moving a very small positive rate closer to zero can make a positive-rate mixture look increasingly like ZIP over a limited observed range, while remaining mathematically different before the boundary. A finite dataset may have little power to distinguish close distributions. Exact inequality of model laws does not guarantee easy practical discrimination.

Contents · Previous chapter · Next chapter

Both models match μ and q, so they necessarily match E(Y|Y>0)=μ/(1−q). In the original example this is (2L)/(11/16)=32L/11, approximately 2.016428.

Thus checking the average among positive observations adds no independent constraint once the overall mean and zero probability are already fixed. It cannot rescue a claim of full positive-shape agreement.

The positive distribution still has many degrees of structure: the relative mass at one, two and higher values can change while preserving its conditional mean. This is a useful diagnostic lesson. Before adding another summary to a comparison, ask whether it is algebraically determined by summaries already matched. A derived restatement may be useful for interpretation, but it is not a new test of distributional equality.

Contents · Previous chapter · Next chapter

For a mixed Poisson count, E[Y(Y−1)]=E(Λ²). Our original two-rate mixture gives (a²+b²)/2=5L².

For ZIP, the second factorial moment is (1−π)λ²=μλ along the mean-matching curve. Using the matched numerical λ gives a different value from 5L². Accordingly, their unconditional variances also differ, since Var(Y)=E[Y(Y−1)]+μ−μ².

This provides another observable distinction, but it is not a guarantee that a small-sample moment estimate will be reliable. Large counts can strongly affect empirical second moments. In actual model assessment, examine predictive distributions and uncertainty rather than treating a single moment discrepancy as a mechanistic verdict. The exact calculations are valuable because they specify what the competing models predict before any data-driven fitting choices are made.

Contents · Previous chapter · Next chapter

Adding more summary constraints can narrow a model comparison, but a finite list of moments does not generally specify an arbitrary count distribution. Different distributions can share several summaries while differing elsewhere.

Our result does not claim that every pair of mixture families can match any desired collection of moments. It establishes one explicit, feasible counterexample with two matched summaries and different positive probabilities. Its general ratio proof supplies a stronger shape check for the specified finite positive-rate mixtures.

A report should therefore state the level of agreement it established: matched mean, matched zero frequency, matched selected moments, or equality of the entire probability mass function. Those are different claims. Mechanism identification can require still more information even when observable distributions agree, as the existing ZIP–hurdle equivalence guide demonstrates for its shared parameter region.

Contents · Previous chapter · Next chapter

The original mixture and its matched ZIP pair were stipulated by equations. We did not generate a finite sample, optimize a likelihood or select predictors. This isolates the information loss caused by matching only two summaries.

A maximum-likelihood ZIP fitted to data from the mixture need not preserve those two summaries exactly in every finite sample. Its selected parameters depend on all observed counts and the constraints of the family. A resulting fit difference should not be attributed to the summary-matching construction unless the fitting procedure actually imposed it.

Likewise, a package’s parameter named zero inflation might use a different link or convention. Reconstruct the mass function from the documented parameterization before comparing predictions. The value π≈0.14195 here belongs to one deliberately constructed ZIP law; it is not an estimate of the proportion of real units unable to experience an event.

Contents · Previous chapter · Next chapter

For observed data, compare predicted frequencies of several positive counts or ranges, not only the total zeros. Use a held-out sample or an appropriate predictive assessment when model fitting and checking could otherwise reuse the same information optimistically.

Sparse high-count bins can be combined into interpretable ranges, but preserve the predicted probability for each range and the sampling assumptions. A visual histogram without uncertainty can make random variation look like a model failure or hide a meaningful tail discrepancy.

The adjacent-ratio identity is a conceptual diagnostic, not a mandate to divide noisy empirical bins. A likelihood comparison or simulation-based predictive check can use the whole sample more coherently. Whatever method is chosen, state which features it can assess and which latent interpretations remain dependent on assumptions. Better count prediction does not automatically verify the rate-group story that produced a convenient mathematical example.

Contents · Previous chapter · Next chapter

The two conditional means a and b refer to the same stated positive exposure in the original example. If exposure e changes under a constant-rate assumption, the conditional means become ea and eb.

The mixture zero probability becomes [exp(−ea)+exp(−eb)]/2, while its mean becomes eμ. A ZIP pair matched at one exposure will not generally match this entire exposure-response curve if its structural-state probability is held fixed and its active mean simply scales with e.

This creates a possible additional comparison when exposure is measured and the scaling assumptions are justified. It does not automatically identify structural ineligibility, because changing detection, unobserved rate variation or dependence may also affect the curve. Record whether exposure represents time, area, opportunities or another denominator. A mathematical match at one observation frame should not silently become a universal prediction across different frames.

Contents · Previous chapter · Next chapter

For strictly positive finite rates a and b, expand exp(−eλ)=1−eλ+e²λ²/2+… near e=0. The mixture zero probability is approximately 1−eμ+e²E(Λ²)/2.

For a fixed ZIP mechanism with active rate r and inflation probability π, the corresponding expansion uses (1−π)r for its mean rate and (1−π)r² for its second factorial-rate contribution. Agreement at one exposure does not force all expansion coefficients to agree.

These are local model calculations, not instructions to reduce every study’s observation duration. Extremely short exposures can leave nearly all counts zero and provide little practical information. The useful point is that a family of exposure conditions can constrain more than one isolated mean-zero pair when the observation process and scaling law are independently defensible.

Contents · Previous chapter · Next chapter

Use a new mixture with weight 2/3 on rate a=ln 3 and weight 1/3 on rate b=ln 27=3 ln 3. Before reading the answer, calculate its mean, zero probability and count-one probability.

Then derive R_k directly from the probability mass function. Decide whether a ZIP can have the same mean and zero probability, and whether it can have the same entire positive-count law. Explain which part of your answer follows from convexity and which follows from the changing ratio.

The new weights matter. Reusing the earlier factor four in the ratio would be incorrect because the coefficients combine both the original group probabilities and their Poisson zero factors. This exercise tests the construction of the weights rather than recognition of a formula with different logarithms substituted into it.

Contents · Previous chapter · Next chapter

Write A=ln 3. The mean is (2/3)A+(1/3)(3A)=5A/3. The zero probability is (2/3)(1/3)+(1/3)(1/27)=19/81.

For positive k, P(Y=k)=A^k[18+3^k]/(81k!). Count-one probability is 21A/81=7A/27. The adjacent ratio is R_k=A[18+3^(k+1)]/[18+3^k]. At k=1 it is 9A/7; at k=2 it is 5A/3.

These distinct ratios again rule out full equality with ZIP. Convexity places the zero probability above exp(−5A/3), so the unique mean-zero ZIP match exists. You can locate it numerically using the same one-dimensional equation with the changed μ and q. None of these calculations requires interpreting an observed individual zero as belonging to a verified structural class.

Contents · Previous chapter · Next chapter

Suppose a record retains only whether Y is zero, discarding the positive count value. Under independent common-parameter observations, that reduced record has a Bernoulli law determined by q alone.

Our original mixture and matched ZIP therefore give exactly the same law for every such zero-versus-positive record. If an independently supplied mean is also retained, both match it by construction. The discarded positive shape cannot be recovered merely by increasing the number of zero indicators under the same recording scheme.

This is information loss caused by data reduction, not evidence that the original full count laws were identical. Before choosing a reporting summary, consider what scientific question the retained record must answer. Zero frequency is useful for some operational decisions, but it cannot carry every distinction between two count-generating models. Preserve full counts when their positive distribution matters to later analysis.

Contents · Previous chapter · Next chapter

The ratio proof establishes that two specified count laws differ. It does not prove that an observed dataset came from the two-rate mixture whenever ZIP fits poorly. Another family, a detection model or a changing observation frame might explain the discrepancy.

Likewise, a ZIP matching the mean and zero fraction does not establish an ineligible subgroup. Its π is the mixing parameter of that representation. Independent eligibility measurements, documented exposure or other design information can make a substantive interpretation more credible, but they must actually address the proposed mechanism.

This distinction protects both modelling and interpretation. A predictive distribution can be useful without its latent states being literally observed. A scientific mechanism claim is stronger and needs its own evidence. State which conclusion is supported by the calculations and which would require information outside the unlabelled count histogram.

Contents · Previous chapter · Next chapter

The Stan User's Guide on finite mixtures documents mixture probability laws and zero-inflated constructions. Those definitions anchor the notation used here. The logarithmic rate mixtures, summary-matching calculation and increasing-ratio proof are independently worked teaching examples.

For the existing discussion of heterogeneous arrival rates, read the Cox-process guide in Queueing Theory. It addresses latent intensity, overdispersion and operational time structure. This article adds the narrower test of a ZIP matched on two particular summaries and a positive-count shape certificate.

Do not infer temporal persistence from the independent mixture in our first example. A unit-specific persistent rate, an occasion-specific redraw and a time-varying intensity process can share some marginal count features while predicting different histories. The joint process must be specified separately.

Contents · Previous chapter · Next chapter

More zeros than a pooled Poisson proves a structural-zero class: our two-positive-rate construction disproves that implication. Matching the mean and zero fraction proves the full law matches: the different probabilities of one disprove it.

The mean among positives is an independent third check after μ and q are matched: it equals μ/(1−q), so it is already determined. A changing empirical adjacent ratio proves two rate groups: finite-sample noise and other count families prevent that mechanistic conclusion.

A useful repair sequence is to write the complete mass function, calculate the matched summaries, then identify one observable not fixed by them. Next change a parameter and repeat the argument. This keeps the distinction between prediction, summary and interpretation visible throughout the calculation instead of adding a generic caveat after an overconfident model claim.

Contents · Previous chapter · Next chapter

Does this contradict ZIP–hurdle equivalence?

No. The ZIP–hurdle equivalence guide derives equality of every marginal count probability for a particular mapping. Here only two summaries match, and the remaining probabilities differ.

Are the fitted-looking ZIP numbers estimates from data?

No. They solve two stipulated matching equations. Their decimals reflect numerical root finding, not sampling uncertainty or evidence for an actual structural group.

Can a finite sample distinguish the two laws reliably?

That depends on sample size, design, parameter separation and the checking procedure. Exact mathematical inequality is not a guarantee of high statistical power.

Must I use ratios of histogram bins in practice?

No. The ratios explain the structural distinction. Sparse empirical bins can make them unstable; use an appropriate likelihood or predictive assessment with uncertainty.

Contents · Previous chapter · Next chapter

Calculate the original mixture masses directly for k=0 through at least five. Solve q_Z(λ)=5/16 along the mean curve μ=2 ln 2, then reconstruct π. Verify the mean and zero probability through both formulas.

Compare count-one probabilities, the positive conditional mean and the adjacent ratios. The first differs, the conditional mean agrees by algebra, and the ratio sequence increases for the mixture while remaining constant for ZIP.

Repeat with the changed weights and rates. Preserve the exact logarithms until numerical display, and state the root-finding tolerance. A successful reproduction should explain why each check agrees or differs; agreement between two software outputs without a stated law is weaker evidence. The result is a transparent mathematical comparison, not a fitted scientific classification of real observational units.

Contents · Previous chapter · Next chapter

In the original mixture, the high-rate group has prior probability one-half. Among observed zeros, its conditional probability is [(1/2)(1/8)]/(5/16)=1/5. Among observed ones it is 3/7. More generally, conditional on count k it is 3^k/(4+3^k).

Among all positive observations, the high-rate group has weight [(1/2)(7/8)]/(11/16)=7/11. Thus the positive-only population has different group weights from the full population. Those selected weights help explain the positive shape but do not replace the original mixture probabilities.

These are exact conditional probabilities under the stipulated mixture, not observed labels. A zero can still come from the high-rate component, and a large positive count can still come from the low-rate component. The model makes some explanations more probable without turning them into certain identities.

This provides another practical checking route: when conditioning or filtering a dataset, recompute the relevant weights from the selection rule. Retaining the original one-half weights after restricting to positives would describe a different distribution. Selection can change the composition of a record even when the underlying population proportions and component laws remain fixed.

Contents · Previous chapter · Next chapter

For a nonnegative integer count, the probability-generating function is G(s)=E(s^Y) where the expectation exists. For our rate mixture it is G_M(s)=[exp(a(s−1))+exp(b(s−1))]/2. For ZIP it is G_Z(s)=π+(1−π)exp(λ(s−1)).

The zero mass is G(0), normalization is G(1)=1, and the mean is G′(1). Our matching construction makes those values agree. It does not make the entire functions equal, just as agreement of two curve values and a slope does not determine every point on a curve.

Higher derivatives at one give factorial moments when finite, connecting this representation to the second-moment check. Expanding either function as a power series in s returns its complete list of count probabilities. The different count-one coefficient directly certifies that the two functions differ.

A generating-function calculation is optional for applying the example, but useful for seeing the information structure. The two summary constraints are particular evaluations of a much richer object. Adding a graph of G(s) can illustrate the disagreement; the explicit coefficients and derivative identities are the reproducible evidence behind that illustration.

Contents · Previous chapter · Next chapter

Suppose two observation windows are recorded. If the mixture rate is independently redrawn for each window, the generating function of their sum is G_M(s)^2. If one latent rate persists across both windows and counts are conditionally independent, the sum instead has generating function [exp(2a(s−1))+exp(2b(s−1))]/2.

The two constructions have identical one-window marginal distributions but different joint structures and different distributions for the two-window total. Their mean totals agree at 2μ. In the persistent case, the latent-rate contribution to the total variance is four times the one-window rate variance; in the independently redrawn case it is twice that variance.

This comparison is separate from the mean-and-zero ZIP match. It shows why an operational forecast over longer periods must state persistence, not merely a one-window histogram. A ZIP state can likewise persist or be redrawn, so neither family name supplies a joint model automatically.

If real event timestamps or repeated-unit identifiers are available, retain them when the target concerns aggregation or persistence. Reducing everything to unlabelled single-window counts can discard the evidence needed to compare those proposals. Additional data are informative only when the compared models make different predictions for them under explicitly stated assumptions.

Contents · Previous chapter · Next chapter

A database may retain categories such as zero, one, two and three-or-more rather than the complete count. Our original two models still differ at count one, so that particular coarsening preserves an observable distinction. A record retaining only zero and positive removes it because both laws share q.

Other category choices can retain different amounts of information. Compute each category probability by summing the complete mass function over its specified values. Do not treat a category labelled three-or-more as the numerical count three when calculating a mean or a ratio.

The adjacent-ratio proof requires individual positive-count probabilities. If the record combines them, the proof remains a statement about the underlying laws, but those ratios cannot be estimated directly from the coarsened record. Design the recording scheme around the intended comparison when possible. Preserving exact counts can enable later checks that a convenient summary cannot support, while privacy, burden or instrument constraints may require a different explicitly documented analysis.

Contents · Previous chapter · Next chapter

For definitions, exposure and model choice, return to the Zero-Inflated Models subject guide. For full observable equivalence rather than summary matching, follow the ZIP and hurdle probability-mapping guide.

For latent-rate time structure, use the Cox-process guide. Connect the worked and changed-context reasoning through The Gold Standard of Education: Learning Transfers into Life and the How X Works hub.

A precise conclusion names the summaries that agree, the positive-count predictions that differ and the independent evidence needed for a mechanism claim. That is more informative than treating a matched average and zero fraction as a certificate for an entire count-generating story.

Contents · Previous chapter · Return to Zero-Inflated Models