Choosing a Count Model for Imputation
A count variable may be nonnegative and integer-valued, but that alone does not determine its imputation model. A count of visits in a fixed period, a number of successes among a known number of trials, and a count with many zero observations have different possible supports and may have different mean–variance patterns. Using a model that ignores those features can generate implausible imputations or misrepresent relationships with other variables.
In MICE, the count model is a conditional model: it describes the count given the other variables in the imputation model. The comparison here focuses on choosing among common models for that conditional distribution. The model should respect known limits, fit the observed distribution reasonably well across relevant predictor patterns, and reflect a plausible data-generating process. A marginal count histogram is useful, but it is not enough by itself.
What Each Candidate Model Assumes
For a Poisson regression, the conditional count \(Y\) has mean \(\mu\), with \(\log(\mu)\) modeled as a linear predictor (and, when appropriate, an offset for exposure). The conditional variance is also \(\mu\). This equality is a defining feature of the Poisson distribution, not a universal property of count data.
A negative-binomial model allows the conditional variance to exceed the mean. One common parameterization has mean \(\mu\), size parameter \(r>0\), and variance \(\mu+\mu^2/r\). It still has support \(0,1,2,\ldots\), but the additional dispersion can represent variability beyond the Poisson model. A negative-binomial fit does not, by itself, assert that zeros arise from a separate zero-generating process.
A zero-inflated model combines a point mass at zero with a count distribution that can itself produce zeros. If \(\pi\) is the probability of belonging to the extra-zero component and the count component is Poisson with mean \(\lambda\), then \(P(Y=0)=\pi+(1-\pi)e^{-\lambda}\) and \(E(Y)=(1-\pi)\lambda\). A hurdle model also models zero probability separately, but all positive values come from a count distribution conditioned to be positive. Thus, a hurdle model has one mechanism for zero versus positive and another for the size of positive counts.
A binomial model is appropriate when the count is the number of successes among a known number \(m\) of trials. Given success probability \(p\), its mean is \(mp\), its variance is \(mp(1-p)\), and its support ends at \(m\). The bound must be real—for example, a known number of opportunities—not a convenient maximum observed in the sample.
| Model | Support | Useful clue | Key constraint |
|---|---|---|---|
| Poisson | 0, 1, 2, ... | Conditional variance is near the conditional mean | Mean equals variance under the model |
| Negative binomial | 0, 1, 2, ... | Conditional variance exceeds the mean | Does not separately identify a zero process |
| Zero-inflated | 0, 1, 2, ... | A plausible extra-zero group as well as ordinary count variation | Requires a defensible interpretation and estimable mixture |
| Hurdle | 0, 1, 2, ... | Zero versus positive and positive count size may have distinct predictors | Positive counts follow a zero-truncated distribution |
| Binomial | 0, 1, ..., m | Known number of trials or opportunities | Imputed values cannot exceed the trial count |
Worked Examples: Match the Model to the Evidence
Worked Example: A Count Bounded by Five Trials
Suppose a hypothetical data set records the number of successful outcomes out of five trials for each of 12 individuals. The counts are \(0,1,1,2,2,2,2,3,3,3,3,3\). There is one zero, the total is 25, and the sum of squared counts is 63. The known five-trial structure makes the binomial model a natural candidate; the observed maximum of three is not the reason for the bound.
The fitted binomial success-probability estimate is \(\hat p=25/(12\cdot5)=5/12\approx0.4167\). The fitted mean count is \(5\hat p=2.083\), and the fitted variance is \(5(5/12)(7/12)=175/144\approx1.215\).
Thus the model predicts \(12(0.06754)\approx0.8105\) zeros, compared with one observed zero. For comparison, a Poisson model with fitted mean \(25/12\) gives \(P(Y=0)=e^{-25/12}\approx0.1245\), or about \(12(0.1245)=1.494\) expected zeros. Neither small-sample comparison establishes that one model is better: the known trial structure is the stronger reason to prefer binomial here.
The sample variance is \((63-25^2/12)/11=131/132\approx0.9924\), below the fitted binomial variance of about 1.215. This does not contradict the binomial choice; sample variance fluctuates, particularly with only 12 records. It does give no clear indication that the observed counts require an overdispersed model. In imputation, the five-trial bound must be retained.
Worked Example: Many Zeros and Extra-Poisson Variation
Consider hypothetical event counts from 10 equal observation periods: \(0,0,0,0,0,0,2,2,3,3\). The total is 10, the sum of squares is 26, the mean is \(10/10=1\), the sample variance is \((26-10^2/10)/9=16/9\approx1.778\), and six of the 10 counts are zero. There is no known fixed trial maximum.
A Poisson model fitted at mean 1 predicts a zero probability of \(e^{-1}\approx0.3679\), or \(10e^{-1}\approx3.679\) zeros. The observed six zeros and sample variance above the mean are reasons to examine alternatives, not proof that a particular alternative is correct.
For a rough negative-binomial comparison, equate the observed mean and sample variance to \(\mu\) and \(\mu+\mu^2/r\). Solving gives \(r=\mu^2/(s^2-\mu)=1/(16/9-1)=9/7\approx1.286\). Its implied zero probability is \((r/(r+\mu))^r=(9/16)^{9/7}\approx0.4773\), corresponding to about 4.773 zeros. The negative-binomial model allows more zeros than the Poisson model through overdispersion, but this fitted probability remains below the observed proportion of 0.600.
A zero-inflated Poisson model can match these two sample summaries approximately. Its mean equation requires \((1-\pi)\lambda=1\), so \(\pi=1-1/\lambda\). Matching a zero probability of 0.600 then requires \(1-(1-e^{-\lambda})/\lambda=0.600\), or \((1-e^{-\lambda})/\lambda=0.400\). A numerical solution is approximately \(\lambda=2.23\), giving \(\pi\approx1-1/2.23=0.5516\). The resulting zero probability is about \(0.5516+0.4484e^{-2.23}\approx0.600\), and the mean is about \(0.4484(2.23)\approx1.000\).
A hurdle model could instead assign zero probability 0.600 and model the four positive counts separately. Their observed mean is \((2+2+3+3)/4=2.5\). For a zero-truncated Poisson positive-count component, the mean is \(\lambda/(1-e^{-\lambda})\); a value near \(\lambda=2.23\) gives a conditional positive mean near 2.5. These calculations illustrate how the models differ. With only 10 records, they do not justify a confident claim about a latent extra-zero group or a distinct hurdle process.
Worked Example: Distinguishing Fixed Opportunities from Event Rates
Suppose six hypothetical people each have exactly four opportunities for an outcome, and their success counts are \(0,1,1,2,2,3\). The total is 9, so the pooled binomial estimate is \(\hat p=9/(6\cdot4)=0.375\), with fitted mean count \(4(0.375)=1.5\). The binomial zero probability is \((1-0.375)^4=0.625^4\approx0.1526\), predicting \(6(0.1526)\approx0.9155\) zeros. One zero is observed.
A Poisson model with mean 1.5 instead predicts \(e^{-1.5}\approx0.2231\) zeros per person, or about \(6(0.2231)=1.339\) zeros. The numerical difference is not enough to override the design information: if the count is genuinely successes in four trials, binomial support is correct. The binomial model cannot impute five successes.
If these were instead event counts over periods of unequal length, a Poisson rate model with an exposure offset might be appropriate, because the number of possible events would not be a shared fixed four. Choosing between the models depends on what the count measures, not simply on which fitted zero probability is closer to the observed number in six records.
A Practical Selection and Checking Process
The following sequence helps organize model choice for a count variable in a MICE conditional model. It is a diagnostic strategy, not a rule that mechanically selects a model from a single statistic.
Check whether the count has a known upper bound, a known number of trials, or an exposure duration. Rule out models with incompatible support.
Assess whether Poisson-like variance is plausible after accounting for predictors and exposure. Overdispersion can motivate a negative-binomial candidate, but should not be diagnosed from a raw marginal variance alone.
Compare observed and model-predicted zero frequencies, then examine the positive-count distribution. Consider zero-inflated or hurdle models only when their distinct zero mechanisms are plausible and supported by enough information.
Evaluate whether fitted conditional distributions and eventual imputations are sensible across predictor patterns, including sparse or extreme regions. Check that imputations obey known bounds and preserve relevant associations.
Why the Distinctions Matter in MICE
Imputations are draws from conditional distributions, so an incompatible support can create impossible completed data. A Poisson or negative-binomial model can generate counts above a true trial limit; a binomial model can enforce that limit when the number of trials is known. Conversely, forcing an arbitrary upper bound on an unbounded event count can truncate plausible values.
Dispersion and zero patterns also affect the distribution of imputations, not just their average. A Poisson model that understates variation may draw too few large counts or too few zeros. A zero-inflated or hurdle model can represent a higher zero frequency, but it adds structure and parameters that may be weakly identified when the sample is small or the predictors have sparse patterns. A negative-binomial model may accommodate extra variation without asserting a separate zero process.
These are conditional-model decisions: the count distribution should be assessed given the other variables used to predict it, rather than selected only from an overall histogram. Model choice also does not guarantee that every conditional relationship in a MICE system is mutually compatible. The practical aim here is a defensible conditional distribution for this count, with diagnostics that reveal evident failures.
Useful warning signs include many imputed values above a known limit, systematic underprediction of zeros or large counts, implausible conditional distributions in particular predictor groups, and unstable fits caused by sparse cells. A model with a close overall zero count can still fail badly in subgroups. When a candidate model has convergence or estimation problems, that is evidence to investigate—not a reason to interpret its output as reliable. Each candidate also has limitations: Poisson can be too restrictive, negative binomial may not capture a distinct zero process, and zero-inflated and hurdle models can be difficult to estimate and interpret.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A count is the number of successes among exactly eight trials. Which candidate model respects its support, and what is the largest possible imputed value?
- A Poisson fit has conditional mean 2.0, but the data show substantial conditional overdispersion. Name one candidate model to examine and state what it changes.
- In a hypothetical sample of 20 counts, a Poisson model predicts 5.0 zeros and 6 are observed. Does that discrepancy establish zero inflation? Give one reason.
- State the difference between zero-inflated and hurdle models in how their positive counts are generated.
Solutions. (1) A binomial model with eight trials has support 0 through 8, so the maximum imputed count is 8. (2) Examine a negative-binomial model; it allows conditional variance greater than the mean through an additional dispersion parameter. (3) No. The observed difference is only one zero and may arise from sampling variation or model misspecification; examine the conditional distribution and positive counts as well. (4) A zero-inflated model mixes an extra-zero component with a count distribution that can also produce zeros. A hurdle model first distinguishes zero from positive, then generates positive values from a zero-truncated count distribution.