Modeling Agricultural Disease Rates

Why Model Disease Rates?

Suppose crop plots differ in how many plants were inspected. A plot with more inspected plants has more opportunities for disease events to be observed, so comparing counts alone can be misleading. An exposure-adjusted count model uses the inspected plant count to distinguish a higher event rate from simply having surveyed more plants.

Consider disease lesions recorded on plots within farms. Lesions are the count outcome, and the number of surveyed plants is the exposure. Multiple plots from the same farm may share soil, weather, or management conditions. A farm-level random intercept represents this clustering while the exposure offset accounts for differences in the number of plants inspected.

Definition: Let $Y_{ij}$ be the number of disease events on plot $i$ in farm $j$, $E_{ij}$ the positive number of plants surveyed on that plot, and $x_{ij}$ an indicator equal to 1 for treatment and 0 for control. An exposure-adjusted Poisson random-intercept GLMM is $$ Y_{ij}\mid b_j \sim \operatorname{Poisson}(\mu_{ij}), \qquad \log(\mu_{ij})=\log(E_{ij})+\beta_0+\beta_1x_{ij}+b_j, $$ where $b_j\sim N(0,\sigma_b^2)$. The term $\log(E_{ij})$ is an offset: its coefficient is fixed at 1. The model describes event counts conditional on the farm effect and the included predictors.

The model assumes that, conditional on $b_j$ and the predictors, the count has a Poisson distribution, so its conditional mean and variance are both $\mu_{ij}$. It also assumes conditional independence of plot counts given the farm effects, a normally distributed farm intercept, and positive, correctly recorded exposure. The random effects are conventionally assumed independent of the predictors. These assumptions are not guaranteed by fitting the model; they require consideration of the study design and diagnostics.

What the Offset Does

Because the logarithm of the expected count contains $\log(E_{ij})$, the model can also be written

\mu_{ij}=E_{ij}\exp(\beta_0+\beta_1x_{ij}+b_j).

Thus, holding treatment and farm effect fixed, doubling the number of surveyed plants doubles the expected count. The corresponding expected rate per surveyed plant is

\frac{\mu_{ij}}{E_{ij}}=\exp(\beta_0+\beta_1x_{ij}+b_j).

This is the reason to use an offset rather than treat exposure as an ordinary predictor with an estimated coefficient: the model imposes proportional scaling of expected counts with the number of plants surveyed. If that scaling is not scientifically appropriate, the offset model may not describe the data well.

Interpretation: With control as the reference category, $\exp(\beta_1)$ is the treatment-to-control ratio of expected disease-event rates per surveyed plant, comparing plots with the same farm effect. It is a rate ratio, not a difference in counts. Its interpretation as a causal treatment effect additionally depends on the study design and adequate control of confounding.

A short derivation makes the rate-ratio interpretation explicit. At the same exposure $E$ and farm effect $b_j$, the ratio of the treated to control conditional expected counts is

\frac{E\exp(\beta_0+\beta_1+b_j)} {E\exp(\beta_0+b_j)} =\exp(\beta_1).

The exposure and farm effect cancel, so the same ratio applies to the conditional rates. The introductory material on Poisson GLMMs and exposure offsets establishes the general model and software framework; here the focus is applying that specification to farm-level disease data and interpreting its treatment contrast.

Worked Example: Translating Exposure Into Expected Counts

Suppose a hypothetical model has a control rate of 0.020 disease lesions per surveyed plant for a farm with $b_j=0$, and a treatment rate ratio of 0.75. Compare expected counts for 500 surveyed plants under control and treatment.

The control expected count is $500\times0.020=10$ lesions. Under treatment, the rate is $0.020\times0.75=0.015$ lesions per plant, so the expected count is $500\times0.015=7.5$ lesions.

For a farm with $b_j=\log(1.2)$, both rates are multiplied by $\exp(b_j)=1.2$. The expected counts become $10\times1.2=12$ under control and $7.5\times1.2=9$ under treatment. The treatment-to-control rate ratio remains $9/12=0.75$. These are expected counts, which need not be whole numbers; an observed count is an integer.

Fitting and Interpreting a Treatment Contrast

For plots nested within farms, a typical software specification is a Poisson model with a log link, treatment as a fixed effect, $\log(E_{ij})$ as an offset, and a random intercept for farm. In R, the model formula for a fit using glmer is: cases ~ treatment + offset(log(plants)) + (1 | farm), with family set to Poisson and link set to log. The variable plants must be positive because its logarithm is used. The treatment reference level should be set deliberately so that the sign and exponentiation of the coefficient match the intended contrast.

Worked Example: Estimating a Farm-Adjusted Rate Ratio

In a hypothetical randomized study, each of 48 farms contributes plots assigned to treatment and control. Disease lesions are counted on surveyed plants. A fitted model reports a treatment coefficient of $\hat{\beta}_1=-0.4055$ with standard error $SE=0.1800$. Assume the stated Poisson, exposure, and random-intercept specifications are appropriate for this illustration.

The estimated treatment-to-control rate ratio is

\exp(\hat{\beta}_1)=\exp(-0.4055)\approx0.6666.

Thus, the fitted treatment rate is about 0.667 times the control rate, or about 33.3% lower, conditional on the farm effect. This is an estimated association; randomization within farms supports a treatment-effect interpretation, but the model result alone does not establish that the assumptions or implementation are sound.

For an approximate Wald test of $H_0:\beta_1=0$, the statistic is

z=\frac{-0.4055}{0.1800}\approx-2.253.

The approximate two-sided normal p-value is 0.0243. At $\alpha=0.05$, this is evidence against $H_0$ under the model and its large-sample approximation. It is not the probability that the null hypothesis is true.

An approximate 95% Wald interval on the log-rate-ratio scale is

-0.4055\pm1.96(0.1800) =[-0.7583,-0.0527].

Exponentiating both endpoints gives the rate-ratio interval

[\exp(-0.7583),\exp(-0.0527)] =[0.4685,0.9487].

The interval indicates that, under the model and its approximation, rate ratios from about 0.469 to 0.949 are compatible with the data at the 95% confidence level. Since the interval is below 1, it corresponds to treatment rates estimated to be lower than control rates. It does not mean there is a 95% probability that the fixed parameter lies in this particular interval.

Predicting Counts and Checking the Fit

Worked Example: Predicting Disease Events on a Plot

Continue with the hypothetical fitted model from the previous example. Suppose the estimated intercept is $\hat{\beta}_0=-3.600$, and a particular farm has estimated random intercept $b_j=-0.200$. For a treated plot with 500 surveyed plants, the conditional fitted mean is

\hat{\mu}_{ij} =500\exp(-3.600-0.4055-0.200) =500\exp(-4.2055) \approx7.46\text{ disease events}.

This is the model’s estimated conditional mean count for that exposure and farm, not a prediction that the next observed count must be 7.46. For comparison, the fitted control mean for the same farm and exposure is $500\exp(-3.600-0.200)\approx11.19$ events. Their ratio is approximately $7.46/11.19=0.667$, consistent with the estimated treatment rate ratio.

For a practical analysis, inspect both the model fit and the data construction. Useful checks include:

Exposure: confirm that every included plot has a positive plant count, that the counts correspond to the area or units at risk, and that they were measured comparably across treatment groups.
Convergence: check software convergence messages and whether the estimated farm-effect variance is on a boundary near zero. Such warnings can indicate numerical or model-structure problems.
Residual patterns: compare observed and fitted counts across farms, treatment groups, and exposure sizes. Systematic discrepancies or unusually frequent large counts can indicate inadequate conditional variation or a poor mean structure.
Clustering: inspect whether residual patterns remain farm-specific. A random intercept represents a shared farm-level shift, but does not automatically handle every form of dependence.

The conditional Poisson assumption equates conditional variance and mean. If substantial extra variation remains after accounting for farm effects and predictors, standard errors and fitted probabilities may be unreliable. Investigate data errors and omitted structure before selecting an alternative count distribution or model. If each surveyed plant can only be classified as diseased or not diseased, a grouped binomial model may better represent the sampling process than a Poisson count model.

How This Approach Compares With Alternatives

Approach	What it does	Important limitation
Crude count comparison	Compares totals without modeling exposure or farm clustering.	Can mistake differences in surveyed plant numbers or farm composition for rate differences.
Poisson model without a farm effect	Models exposure-adjusted counts but omits shared farm heterogeneity.	May fail to account for dependence among plots from the same farm.
Exposure-adjusted Poisson GLMM	Uses an offset for plant exposure and a farm random intercept.	Relies on the conditional Poisson mean-variance relationship and the specified random-effects structure.
Grouped binomial model	Models diseased and non-diseased counts when each plant has one binary status.	Does not represent multiple events per plant in the same way as an event-count model.

The offset GLMM is useful when the outcome is a count of events, exposure varies across plots, and farms create a plausible source of clustering. It is limited if exposure is measured poorly, the count-generating process is not approximately Poisson conditional on the modeled effects, or treatment is confounded with unmodeled farm differences. A crude rate ratio can be a useful descriptive summary, but it does not replace a model that accounts for the design and clustering.

Common Mistakes

Comparing counts without exposure. A plot with twice as many surveyed plants can have twice the expected count at the same rate. Compare rates or fit the offset model.
Estimating an exposure coefficient by accident. The offset uses $\log(E_{ij})$ with its coefficient fixed at 1. Including it as an ordinary predictor changes the model assumption.
Calling the exponentiated coefficient a count difference. $\exp(\hat{\beta}_1)$ is a rate ratio; expected counts also depend on exposure and the other model terms.
Ignoring the farm structure. Treating plots from one farm as unrelated can understate uncertainty when outcomes share farm-level conditions.
Overstating the conclusion. A fitted rate ratio below 1 is not, by itself, proof of a causal reduction. Consider randomization, exposure measurement, model fit, and the uncertainty interval.

Key takeaway: For a count of disease events, the log of the number of surveyed plants enters as an offset when expected counts are assumed proportional to exposure. With treatment coded against control, exponentiating its coefficient gives the treatment-to-control event-rate ratio, conditional on the modeled farm effect.

Check Your Understanding

Answer each question, then compare your work with the solutions below.

If exposure increases from 300 to 600 plants while the rate and farm effect stay fixed, what happens to the expected count under this offset model?
A fitted treatment coefficient is 0.200. Calculate and interpret the treatment-to-control rate ratio.
A fitted rate is 0.012 events per surveyed plant. What is the expected count for a plot with 800 surveyed plants?
A treatment coefficient is $-0.300$ with standard error 0.150. Calculate an approximate 95% rate-ratio interval and state whether it excludes 1.

Solutions. (1) The expected count doubles because it is proportional to exposure: $600/300=2$. (2) $\exp(0.200)\approx1.221$, so the estimated treatment rate is about 22.1% higher than the control rate, conditional on the farm effect. (3) $800\times0.012=9.6$ expected events. (4) On the log scale, $-0.300\pm1.96(0.150)=[-0.594,-0.006]$. Exponentiating gives approximately $[0.552,0.994]$. This approximate interval excludes 1, narrowly, indicating evidence of a lower treatment rate at the two-sided 5% level under the model’s Wald approximation.