Worked Ordinal and Count PK/PD Model Interpretation

From Fitted Coefficients to Endpoint Predictions

A fitted coefficient is useful only when its meaning is clear on the response scale. For an ordinal endpoint, interpretation means translating thresholds and the concentration effect into probabilities for the ordered categories. For a count endpoint, it means translating the log-rate effect into an expected event rate or count, while keeping exposure time explicit. These are different response scales: an ordinal odds ratio is not a probability difference, and a count-rate ratio is not a ratio of event probabilities.

The examples below use hypothetical fitted models and data. They assume that the model form has been selected and estimated; the focus is on interpreting the fit, its uncertainty, and diagnostics relevant to each endpoint.

Interpreting a Proportional-Odds Fit

Definition: For an ordered response $Y$ with categories $1,\ldots,K$, a proportional-odds model can be written as $$ \operatorname{logit}\{P(Y\leq k\mid C)\}=\alpha_k-\beta C,\qquad k=1,\ldots,K-1, $$ where $C$ is concentration, $\alpha_k$ is the threshold for cumulative split $k$, and $\beta$ is the common concentration effect. The thresholds must be ordered, $\alpha_1<\cdots<\alpha_{K-1}$. With this sign convention, a positive $\beta$ shifts probability toward higher categories. The proportional-odds assumption is that the same $\beta$ applies at every cumulative split.

The model first gives cumulative probabilities $F_k(C)=P(Y\leq k\mid C)$. Individual category probabilities follow by subtraction: $P(Y=1\mid C)=F_1(C)$, $P(Y=j\mid C)=F_j(C)-F_{j-1}(C)$ for an interior category, and $P(Y=K\mid C)=1-F_{K-1}(C)$. This subtraction is valid because the thresholds are ordered and the cumulative probabilities consequently increase with $k$.

Worked Example: Predicting Three Ordinal Category Probabilities

Suppose a hypothetical three-category tolerability endpoint is coded 1 = mild, 2 = moderate, and 3 = severe. A fitted model has $\alpha_1=-1.20$, $\alpha_2=0.50$, and $\beta=0.70$ per mg/L. Calculate the category probabilities at concentrations of 0 and 2 mg/L.

At $C=0$, the cumulative probabilities are $F_1=\operatorname{logit}^{-1}(-1.20)=0.2315$ and $F_2=\operatorname{logit}^{-1}(0.50)=0.6225$. Thus the probabilities for mild, moderate, and severe are $0.2315$, $0.6225-0.2315=0.3910$, and $1-0.6225=0.3775$, respectively. They sum to 1.000.

At $C=2$ mg/L, the cumulative logits are $-1.20-(0.70)(2)=-2.60$ and $0.50-(0.70)(2)=-0.90$. Therefore $F_1=\operatorname{logit}^{-1}(-2.60)=0.0691$ and $F_2=\operatorname{logit}^{-1}(-0.90)=0.2891$. The category probabilities are $0.0691$, $0.2891-0.0691=0.2199$, and $1-0.2891=0.7109$, subject to rounding.

The fitted probabilities imply a shift toward more severe categories at 2 mg/L in this hypothetical example. They are model-based estimates, not observed proportions or proof that concentration causes severity.

The coefficient also has an odds interpretation. Rearranging the model gives $\operatorname{logit}\{P(Y>k\mid C)\}=-\alpha_k+\beta C$. Therefore, for a concentration increase $\Delta C$, the odds of being above category $k$, rather than at or below it, are multiplied by $\exp(\beta\Delta C)$, for every split $k$. The common multiplier is what “proportional odds” means.

Worked Example: Interpreting an Ordinal Effect and Its Uncertainty

For the hypothetical fit above, suppose the estimated standard error of $\beta$ is 0.25 per mg/L. Interpret a 2 mg/L increase and calculate a Wald 95% confidence interval.

The estimated cumulative odds ratio is $\exp(0.70\times2)=\exp(1.40)=4.055$. Thus, under the fitted proportional-odds model, a 2 mg/L increase multiplies the odds of being above any given category threshold by about 4.06. This does not mean that the probability of a severe response is multiplied by 4.06; probabilities depend on the thresholds and starting concentration.

The coefficient-scale interval is $0.70\pm1.96(0.25)=[0.210,1.190]$ per mg/L. For a 2 mg/L contrast, multiply the endpoints by 2 and exponentiate: $$ \left[\exp\{2(0.210)\},\ \exp\{2(1.190)\}\right]=[1.522,\ 10.80]. $$ The interval is wide, so the odds multiplier is imprecisely estimated even though the fitted direction is positive. This Wald interval describes uncertainty in the odds multiplier under the model and its large-sample approximation; it is not a probability statement about where the fixed parameter lies.

Interpreting a Negative-Binomial Concentration-Response Fit

Definition: A negative-binomial count model with exposure time $T$ can be parameterized by $$ \log(\mu)=\log(T)+\beta_0+\beta_1 C,\qquad \operatorname{Var}(Y\mid C,T)=\mu+\phi\mu^2, $$ where $\mu=E(Y\mid C,T)$, $T$ is the observation time, and $\phi>0$ is the overdispersion parameter. The term $\log(T)$ is an offset with coefficient fixed at 1. For a concentration contrast $\Delta C$ at the same exposure time, the fitted count-rate ratio is $\exp(\beta_1\Delta C)$.

The interpretation assumes that the log mean is correctly specified, the exposure-time offset is appropriate, and the conditional count distribution has the stated mean-variance relationship. Conditional independence is needed for the usual likelihood-based uncertainty calculations; repeated observations require an appropriate treatment of subject dependence. The negative-binomial variance allows more variability than a Poisson model, but it does not automatically account for excess structural zeros, an incorrect concentration-response shape, or omitted heterogeneity.

Worked Example: Converting a Count Coefficient to Rates and Counts

In a hypothetical study of symptom episodes, suppose the fitted model is $\log(\mu)=\log(T)+\log(2.400)-0.800C$, with $T$ measured in days and $C$ in mg/L. Compare the rate at $C=1$ mg/L with the rate at $C=0$, then calculate the expected count over two days at $C=1$.

At $C=0$, the fitted rate is $2.400$ episodes per day. The rate ratio for an increase of 1 mg/L is $\exp(-0.800)=0.4493$. Thus the fitted rate at 1 mg/L is $2.400(0.4493)=1.078$ episodes per day, about 55.1% lower than at 0 mg/L. The model describes a rate association conditional on the modeled covariates; it does not by itself establish a causal effect.

For two days at $C=1$, use the unrounded rate in the multiplication: $\mu=2\{2.400\exp(-0.800)\}=2.156779$, or 2.157 expected episodes. This is an expected count, not a guarantee that an individual will experience a fractional number of episodes.

For a confidence interval on the concentration effect, first form an interval for $\beta_1$, then transform it to the scale of the desired contrast. If an interval for $\beta_1$ is $[L,U]$, the rate-ratio interval for $\Delta C>0$ is $[\exp(L\Delta C),\exp(U\Delta C)]$. The coefficient standard error must be on the same concentration scale as the coefficient. A prediction interval for a future count is a different quantity: it must also reflect count variability and is generally wider than uncertainty in the mean rate alone.

Worked Example: Rate-Ratio Uncertainty and a Count Diagnostic

For the hypothetical model above, suppose the standard error of $\beta_1=-0.800$ is 0.250 per mg/L and the estimated overdispersion parameter is $\phi=0.450$. Calculate a 95% Wald interval for the 1 mg/L rate ratio. Then assess a hypothetical one-day observation of 4 episodes at $C=1$ mg/L using a standardized residual based on the fitted negative-binomial variance.

The coefficient interval is $-0.800\pm1.96(0.250)=[-1.290,-0.310]$ per mg/L. Exponentiating gives a rate-ratio interval of $[\exp(-1.290),\exp(-0.310)]=[0.2753,0.7334]$. In this hypothetical fit, the interval is below 1; its interpretation is an estimated decrease in rate for a 1 mg/L increase, with uncertainty quantified by these endpoints.

At $C=1$ and $T=1$ day, the fitted mean is $1.0783895$ episodes. The fitted variance is $\mu+\phi\mu^2=1.0783895+0.450(1.0783895^2)=1.602$ episodes squared. The standardized residual is $(4-1.0783895)/\sqrt{1.602}=2.309$. This is a potentially unusual observation under the fitted model and merits investigation, but one residual alone does not prove model failure. Check its context, other residuals, and predictive fit before drawing a conclusion.

Diagnostics, Alternatives, and Limitations

Diagnostics should match the endpoint and the model’s assumptions. For an ordinal fit, inspect observed versus predicted category frequencies across concentration ranges, calibration of cumulative probabilities at each threshold, and whether residual patterns or lack of fit are concentrated in particular categories. Also check that estimated thresholds are ordered and that predicted category probabilities are nonnegative and sum to one. A systematic concentration-related discrepancy at only one threshold may indicate that the common-slope assumption is inadequate; the earlier material on proportional-odds assessment and partial proportional-odds models provides the relevant alternatives. If observations are repeated within subjects, predictions and uncertainty must respect that dependence rather than treating records as independent.

For negative-binomial fits, compare observed counts with fitted means across concentration and exposure time, examine residuals for trends and influential observations, and use predictive checks that assess the count distribution, including its spread and frequency of zeros. A few large standardized residuals, persistent underprediction of high counts, or systematic variation with exposure time can signal problems. The estimated $\phi$ permits extra-Poisson variation, but a negative-binomial fit is not automatically adequate just because it converges. Excess zeros may motivate zero-inflated or hurdle models; repeated counts may require subject-level effects; and a misspecified concentration curve may require a different mean model. These approaches answer different modeling needs and should be compared using fit, diagnostics, and scientific plausibility.

Binary logistic regression is appropriate when the endpoint genuinely has two categories, but collapsing an ordered endpoint discards distinctions among its levels. Conversely, an ordinal model uses category ordering without assuming equal spacing between categories. A partial proportional-odds model relaxes the common slope for selected effects when supported by diagnostics. For counts, Poisson regression imposes $\operatorname{Var}(Y\mid C,T)=\mu$, whereas the negative-binomial model allows the variance to exceed the mean according to its parameterization. Zero-inflated and hurdle models distinguish processes governing zeros from positive counts; they are not interchangeable with ordinary negative-binomial regression without evidence for that structure.

All these interpretations are conditional on the chosen model. Plug-in category probabilities and expected counts omit parameter uncertainty unless an interval or predictive procedure includes it. Wald intervals can be unreliable with small samples, sparse ordinal categories, or poorly estimated coefficients. Finally, concentration-response associations may reflect design features, covariates, or subject differences; model coefficients alone do not establish causation.

A Practical Interpretation Sequence

Name the endpoint and scale.
For ordinal outcomes, state the category ordering and cumulative split. For counts, state the event, the time unit, and whether the quantity is a rate or an expected count.

Translate the coefficient.
For proportional odds, use $\exp(\beta\Delta C)$ as a cumulative odds ratio. For a log-rate count model, use $\exp(\beta_1\Delta C)$ as a count-rate ratio.

Calculate response-scale predictions.
Use thresholds to obtain ordinal category probabilities, or include exposure time in the count mean. Retain unrounded intermediate values until the final reported result.

Check uncertainty and fit.
Transform coefficient intervals to the target scale, then examine endpoint-specific calibration, residuals, and predictive behavior.

Common Mistakes

Calling a cumulative odds ratio a probability ratio. It compares odds at a cumulative split; the corresponding probabilities depend on the thresholds and baseline concentration.
Ignoring the exposure offset. A count mean for two days is twice the fitted daily rate when the offset specifies a constant rate over time; it is not the one-day mean.
Rounding too early. Compute $2\{2.400\exp(-0.800)\}=2.156779$, which rounds to 2.157 expected episodes, rather than multiplying a previously rounded rate of 1.078 and reporting 2.156.
Treating a significant coefficient as proof of adequate fit. A coefficient test does not verify proportional odds, the count variance structure, calibration, or independence.

Key takeaway: Interpret each fitted effect on its own scale, calculate predictions using the complete fitted model, and assess uncertainty and fit with diagnostics appropriate to the endpoint.

Check Your Understanding

Answer each question, then compare with the solutions below.

In a proportional-odds model with $\beta=0.40$ per mg/L, what is the cumulative odds ratio for a 1.5 mg/L increase, and how should it be described?
If ordinal cumulative probabilities at two thresholds are 0.20 and 0.65, what are the probabilities of categories 1, 2, and 3?
A count model has a rate ratio of 0.60 per 1 mg/L increase. What rate ratio corresponds to a 2 mg/L increase?
For a count model with rate 3 episodes per day, what is the expected count over 4 days if the rate is constant?
Name one diagnostic concern specific to ordinal predictions and one specific to negative-binomial counts.

Solutions. (1) $\exp(0.40\times1.5)=\exp(0.60)=1.822$. The odds of being above any cumulative threshold are estimated to be 1.822 times as large for the higher concentration, under proportional odds. (2) The probabilities are $0.20$, $0.65-0.20=0.45$, and $1-0.65=0.35$; they sum to 1. (3) The 2 mg/L rate ratio is $0.60^2=0.36$, assuming the same log-linear concentration effect. (4) The expected count is $4(3)=12$ episodes. (5) Ordinal checks include observed versus predicted frequencies at each category or cumulative split. Count checks include residual patterns, observed-versus-predicted count spread, and predictive assessment of zeros.