1 Foundations of Exposure–Response Relationships

1.1 Definitions: exposure, dose, and response

In an exposure–response curve, the x-axis represents an exposure level or dose and the y-axis represents a measurable response. “Exposure” is a general term for what the system experiences, such as the concentration of a substance in the environment or the intensity of a stressor. “Dose” often emphasizes the amount delivered over time or integrated impact, such as accumulated drug exposure or delivered dose to a target. “Response” is the outcome being quantified—ranging from biomarker change and symptom severity to the probability of an event.

The curve summarizes how the response varies as the exposure level changes, typically across a range of conditions rather than at a single point in time.

1.2 Common shapes and interpretations

Exposure–response curves can take many forms depending on the underlying mechanism and measurement process. Common patterns include:

  • Monotone increasing: response rises with exposure up to a plateau.
  • Monotone decreasing: response falls as exposure increases.
  • Threshold-like behavior: little change occurs until exposure reaches a characteristic level, followed by a rapid shift.
  • Saturating behavior: response approaches an upper limit at high exposures.
  • Biphasic or hormesis-like patterns: response may increase at low exposures and decrease at higher levels, or the opposite, sometimes reflecting competing effects or measurement artifacts.

Interpretation usually focuses on relative changes, curvature, and asymptotic behavior rather than single-point comparisons.

1.3 Units, scaling, and transformations

Because exposure and response may span several orders of magnitude, analysts often use transformations to stabilize variance and improve interpretability. Examples include:

  • Log-scaling exposure when effects appear multiplicative.
  • Semi-log or log-log axes for sigmoidal and power-like relationships.
  • Centering and scaling response measures to aid numerical stability and interpret coefficients.

When transforming variables, it is important to preserve the meaning of derived quantities (such as effective concentrations) on the original scale used for decision-making.

1.4 Assumptions underlying curve usage

Exposure–response modeling relies on assumptions that vary by method. Typical assumptions include:

  • The chosen form captures the relationship reasonably across the observed range.
  • Observations at a given exposure share a common mean response up to random error.
  • The exposure level is measured with acceptable accuracy relative to the effect size.
  • Independence or a defensible dependence structure holds (e.g., within-subject correlation in longitudinal studies).
  • Covariates included in the model account for systematic differences rather than acting as proxies for unmeasured factors.

Violations may bias parameter estimates and distort derived metrics, motivating diagnostics and sensitivity analyses.

2 Data Types and Experimental Designs

2.1 Study designs that generate exposure–response data

Exposure–response data arise from many settings: dose-ranging experiments, concentration–effect analyses, repeated measurements under different conditions, and observational studies with quantified exposure levels. The key requirement is coverage across a range of exposures so the curve can be identified. Designs vary in how exposures are assigned (randomized vs. observational), how many exposure levels are sampled, and how response is recorded.

Quality also depends on whether the exposure level corresponds to the same conceptual quantity across participants, such as comparable timing, bioavailability, or measurement definitions.

2.2 Continuous vs. categorical response variables

Responses may be continuous (e.g., biomarker concentration), binary (event occurred vs. not), ordinal (graded severity), or count data. Each response type motivates different likelihoods and link functions. Continuous outcomes often use least-squares or normal-error likelihoods, while binary outcomes often use logistic or probit formulations. Ordinal and count responses require methods tailored to their distributional structure.

The modeling strategy should match the response’s measurement properties rather than forcing a continuous approximation.

2.3 Cross-sectional vs. longitudinal measurement structures

Cross-sectional data observe each unit once at a given exposure level. Longitudinal data track response over time and may involve repeated exposure changes. Longitudinal structures introduce correlation across time points and may require random effects or explicit time components. In some settings, an exposure–response curve can be constructed at multiple time horizons, effectively yielding a family of curves.

Timing can be crucial when exposure effects manifest after a lag.

2.4 Dose-ranging and concentration–time considerations

If the exposure is time-varying, a curve may depend on how exposure is summarized (e.g., peak concentration, area under the concentration–time curve). Concentration–time relationships can also influence the apparent shape of the exposure–response curve, especially when response reflects delayed or cumulative effects.

Dose-ranging studies typically sample multiple dose levels to ensure that the curve includes regions where the response transitions and saturates.

2.5 Handling repeated measures and clustering

When multiple observations come from the same subject or experimental unit, observations are clustered and not independent. Standard curve fitting can understate uncertainty if clustering is ignored. Approaches include mixed-effects models, robust (sandwich) standard errors, or explicit correlation structures. The chosen method should reflect how clustering arises—such as repeated assessments on individuals or batch effects across experimental runs.

Properly addressing clustering affects both parameter uncertainty and the reliability of derived metrics.

3 Exploratory Analysis and Visualization

3.1 Descriptive statistics across exposure bins

Before fitting a model, analysts often compute summaries by grouping exposure values into bins. Typical summaries include mean or median response, along with variability measures. Binning can reveal whether the relationship is approximately linear, curved, or threshold-like. It can also highlight sparse regions where estimates are unstable.

Care should be taken: binning choices can influence perceived shape, especially if exposure values are unevenly distributed.

3.2 Choice of plot scales (linear, log, semi-log)

Visualization benefits from selecting scales consistent with the hypothesized mechanism and data range. Log scales can make multiplicative effects and power-law relationships more apparent. Semi-log plots are common when response approaches a plateau but the transition occurs over orders of magnitude in exposure. The scale choice should support qualitative interpretation and guide subsequent model selection.

Plots should always be labeled clearly so that viewers interpret values on the correct scale.

3.3 Smoothing methods for initial curve inspection

Smoothing techniques help explore the underlying pattern without committing to a specific functional form. Common options include kernel smoothing or local regression. Smoothing is best treated as an exploratory aid rather than a final inferential tool, because it can introduce bias near boundaries and can be sensitive to bandwidth choices.

In reporting, it is often useful to show the smoothed trend alongside raw summaries.

3.4 Detecting outliers and non-monotonic patterns

Outliers may arise from measurement error, data entry issues, or genuine deviations caused by unmodeled factors. Non-monotonic patterns can be informative—indicating possible biphasic behavior—or can reflect heterogeneity, covariate differences, or insufficient measurement precision. Examining residuals from a basic fit, overlaying smoothed curves, and checking leverage points can help distinguish these possibilities.

When non-monotonicity appears, analysts should verify whether it persists after adjusting for obvious covariates or data quality issues.

3.5 Assessing baseline and control/reference groups

When a baseline exists—such as untreated controls or reference conditions—it anchors interpretation of effect size relative to no exposure or minimal exposure. Visual comparisons between baseline and exposed groups help determine whether the curve should be modeled with a lower asymptote fixed or allowed to vary. Baseline behavior can also indicate whether observed effects are offset by systematic shifts in measurement or subject characteristics.

In many analyses, baseline informs both model constraints and the interpretation of derived thresholds.

4 Parametric Modeling Approaches

4.1 Logistic and sigmoid models

Logistic and related sigmoid models represent smooth transitions from lower to higher response levels. These forms are useful when response changes gradually around a central exposure level and approaches asymptotes at low and high exposure. Parameters typically include baseline level, upper asymptote, a location parameter (e.g., the inflection point), and a slope controlling transition steepness.

Such models provide a compact description and yield interpretable quantities for effective doses.

4.2 Emax and Hill-type (sigmoid Emax) models

Emax-type models describe response that increases toward a maximum effect. The Hill extension adds flexibility to the steepness via a shape parameter, allowing curves to range from shallow transitions to near step-like behavior. These models are widely used because they encode saturation: beyond a sufficiently large exposure, incremental increases produce diminishing returns.

In some settings the “maximum response” parameter can be interpreted mechanistically as a capacity limit of the system.

4.3 Linear and piecewise-linear models

Linear models may be appropriate when the relationship appears approximately constant over the observed range and saturation is not evident. Piecewise-linear models introduce breakpoints to capture changes in slope, such as an approximate threshold where the relationship becomes steeper. These approaches are often simpler but require careful choice of breakpoint locations or selection procedures.

Piecewise models can be robust when the true shape is irregular, but uncertainty in breakpoint estimation can be substantial.

When the response is a probability, link functions connect the mean response to a linear predictor. Logistic regression uses a logit link and probit uses a normal-based link. If exposure affects the probability in a monotone fashion, these models can be translated into exposure–response curves by mapping predicted probabilities across exposure levels.

The choice between logit and probit can be guided by convention, interpretability, or empirical fit, though differences are often modest.

4.5 Saturating, biphasic, and hormesis-like behaviors

Some systems show decreases at high exposures after initial increases, producing biphasic shapes. To model such behavior, analysts may use functions that allow a peak at intermediate exposure or incorporate two competing components. “Hormesis-like” patterns—where low exposure yields effects opposite to high exposure—are sometimes observed in practice, though they may also reflect confounding or measurement structure. Flexible parametric models can accommodate such patterns while still yielding smooth curves.

When the curve shape is complex, constraints and diagnostic checks become especially important.

4.6 Flexible models (splines, GAMs)

Flexible parametric families such as regression splines and generalized additive models (GAMs) allow the curve to bend without specifying a single global form. Spline bases or smooth terms can capture curvature and local deviations. These methods can reduce misspecification bias at the cost of more tuning choices (e.g., number of knots or smoothing penalties) and sometimes reduced interpretability.

Regularization and validation help prevent overfitting.

5 Nonparametric and Semi-parametric Methods

5.1 Kernel smoothing and local regression

Kernel smoothing and local regression estimate the curve directly from the data by averaging nearby observations with weights determined by distance in exposure. These techniques adapt to local structure and can reveal patterns that parametric models might miss. However, performance depends strongly on bandwidth or neighborhood size, and edges near the minimum or maximum exposure can be less reliable.

Nonparametric estimates often serve as benchmarks for parametric fits.

5.2 Regression splines and monotonic constraints

Regression splines generalize local smoothing while offering structured control of smoothness. Adding monotonic constraints can enforce plausible directionality when the response should not decrease with exposure. Such constraints can improve stability and interpretability, especially when data are sparse in some exposure regions.

Constrained splines can also align statistical estimation with mechanistic expectations.

5.3 Gaussian process regression for curve estimation

Gaussian process regression (GPR) models the response as a random function, producing not only point estimates but also uncertainty across the exposure range. A covariance kernel governs smoothness and correlation length, shaping how rapidly the estimated curve can change. GPR can naturally represent complex patterns and provides predictive intervals useful for uncertainty reporting.

Computation can be demanding for very large datasets, and kernel selection matters for realism.

5.4 Semi-parametric mixture and hazard-based views

Semi-parametric approaches may combine parametric and nonparametric elements, such as mixtures of subpopulations that share different response dynamics. In survival-type settings, hazard-based formulations connect exposure to event risk over time, producing an implied exposure–response relationship. These frameworks can be useful when responses appear to come from heterogeneous mechanisms rather than a single smooth process.

Interpretation focuses on the implied structure—subgroup mixture weights, hazard ratios, or latent components.

6 Parameter Estimation and Inference

6.1 Likelihood-based estimation basics

Likelihood-based methods choose parameters that maximize the probability of observing the data under a specified model. The likelihood depends on the assumed error distribution and the mapping from exposure to the mean response. For parametric exposure–response curves, this procedure yields estimates of curve parameters such as asymptotes, slopes, and location metrics.

Using maximum likelihood also sets the stage for principled uncertainty quantification.

6.2 Bayesian estimation and credible intervals

Bayesian methods treat parameters as random variables with prior distributions. The observed data update the priors to produce posterior distributions, from which credible intervals are derived. Bayesian modeling can be advantageous when incorporating prior scientific knowledge, handling complex dependence structures, or obtaining uncertainty for derived quantities.

However, results can depend on prior choice and must be checked for sensitivity.

6.3 Maximum likelihood, least squares, and robust estimation

Least squares is common for continuous outcomes under approximate normal errors, while maximum likelihood generalizes estimation across many distributions. Robust estimation strategies can reduce sensitivity to outliers or departures from model assumptions. Examples include heavy-tailed error models or M-estimators, which alter the contribution of points with large residuals.

Robust approaches improve practical reliability when the data deviate from ideal conditions.

6.4 Confidence intervals and uncertainty quantification

Confidence intervals quantify uncertainty around parameter estimates or derived metrics. Techniques include:

Uncertainty should reflect both measurement variability and model uncertainty where possible, especially when comparing alternative curve forms.

6.5 Model comparison using information criteria and tests

To decide between competing models, analysts may use information criteria such as AIC or BIC, which balance fit and model complexity. Likelihood ratio tests can be used under certain regularity conditions, while cross-validation and predictive checks provide more assumption-light evaluation. Model comparison is particularly important when flexible models can overfit.

The preferred model is the one that offers adequate predictive performance with plausible complexity.

7 Key Quantities Derived from Curves

7.1 Thresholds and benchmark effects

Thresholds denote exposure levels associated with a minimal meaningful response, while benchmark effects specify a target effect size used to define an effective exposure. These quantities translate the curve into actionable summary numbers. Estimating them requires a model that allows inversion: finding the exposure that corresponds to a chosen response level.

Because thresholds depend on definitions, careful selection and justification are required.

7.2 ED/EC/IC-style metrics (effective/concentration/inhibitory values)

ED/EC/IC metrics represent exposure values corresponding to a specific response level. For example:

  • ED: effective dose producing a specified effect,
  • EC: effective concentration producing a specified effect,
  • IC: inhibitory concentration producing a specified degree of inhibition.

These are typically derived by inverting a fitted curve and interpolating between observed exposure levels, with extrapolation used cautiously.

Reporting usually includes uncertainty ranges to reflect estimation variability.

7.3 BMD/BMC concepts and implementation

Benchmark dose (BMD) and benchmark concentration (BMC) generalize ED/EC ideas by tying the benchmark to a predetermined change relative to background, often expressed as an absolute or relative deviation. Implementation involves selecting benchmark criteria and fitting a model to estimate the exposure producing that criterion, along with confidence intervals.

This approach standardizes how “effective” is defined across studies.

7.4 Margin of exposure concepts (general framework)

Margin of exposure (MOE) is a ratio-like concept that compares a reference exposure level (such as a baseline or exposure estimate) to an effective benchmark exposure. In a general framework, a larger MOE indicates a greater separation between background or typical exposure and an effective threshold derived from the curve. MOE depends on both the benchmark metric and the exposure estimate definition and uncertainty.

Interpretation requires attention to units and the uncertainty of both components.

7.5 Sensitivity analysis for derived metrics

Derived metrics can be sensitive to modeling choices such as the curve family, transformation settings, and covariate adjustments. Sensitivity analyses evaluate how benchmark or effective values change when these choices vary within reasonable alternatives. Analysts may also vary assumptions about error structure or constrain parameters differently.

This helps assess robustness and prevents overreliance on a single fitted specification.

8 Model Diagnostics and Goodness-of-Fit

8.1 Residual analysis and predictive checks

Diagnostics often start with residual plots and distributional checks. Residual analysis can show systematic departures such as nonlinearity, incorrect variance structure, or missing covariate effects. Predictive checks compare observed responses to simulated or predicted responses from the fitted model, highlighting regions where the model fails to capture variability.

These diagnostics guide whether to refit with different assumptions or model forms.

8.2 Assessing heteroscedasticity and variance structure

Variance may change with exposure, producing heteroscedasticity. If the error variance increases or decreases across exposure levels, standard intervals can be miscalibrated. Methods to handle this include variance-stabilizing transformations, heteroscedastic regression, or distributional choices that embed changing variance.

A good diagnostic also evaluates whether the model’s uncertainty bands widen appropriately where data are noisier.

8.3 Goodness-of-fit metrics and calibration plots

Quantitative fit measures include log-likelihood, deviance, or squared error metrics depending on the model and response type. Calibration plots are particularly useful for probability responses: they compare predicted probabilities against observed event frequencies. Poor calibration indicates that predicted probabilities are systematically too high or too low.

These checks support whether the model is fit for estimation, inference, or prediction tasks.

8.4 Validating monotonicity and asymptotic limits

For mechanistic or practical reasons, some curves should be monotone and may have meaningful lower and upper bounds. Validation assesses whether the fitted curve respects these expectations across the exposure range. If the chosen model yields implausible behavior—such as response decreasing when it should increase—constraints or different functional forms may be needed.

Asymptotic limits also matter when extrapolating beyond observed exposures.

8.5 External validity and transportability considerations

A curve fitted in one dataset may not transfer directly to another setting with different populations, measurement protocols, or exposure distributions. External validity concerns how well the relationship holds under new conditions. Transportability assessments evaluate whether additional covariates, scaling adjustments, or stratified modeling are required.

Even with good internal fit, external performance determines practical usefulness.

9 Dealing with Special Data Challenges

9.1 Censored, truncated, and bounded responses

Some measurements cannot be observed beyond certain limits, leading to censoring or truncation. Bounded responses may include floor or ceiling effects. These data structures require specialized likelihoods or measurement models so that the curve fitting does not treat limited-range values as ordinary uncensored observations.

Correct handling can materially change estimated thresholds and uncertainty.

9.2 Measurement error in exposure or response

When exposure values are measured with error, parameter estimates can become biased, especially if the error is substantial relative to the effect. Response measurement error can similarly distort variability and curve steepness. Methods to address these problems include calibration models, regression calibration, errors-in-variables approaches, or using replicate measures when available.

Assessing measurement quality and magnitude of error is essential for credible inference.

9.3 Missing data handling strategies

Missingness may occur in exposures, responses, or covariates. Strategies range from complete-case analysis to multiple imputation and model-based missingness handling. The choice depends on the missingness mechanism and how much data are incomplete. Sensitivity analyses help evaluate how different assumptions about missingness affect curve parameters and derived metrics.

Underestimating uncertainty is a common risk if missingness is ignored.

9.4 Between-subject vs. within-subject variability

Response variation can arise both across individuals and across repeated measures within individuals. Disentangling these sources helps avoid overstating effect precision. Mixed-effects models and hierarchical structures provide a framework for distinguishing population-level curve shape from subject-specific deviations.

This distinction also informs how to interpret uncertainty for predictions at new subjects.

9.5 Confounding and covariate adjustment

In observational studies or non-randomized designs, exposure may correlate with other factors that influence response. Confounding adjustment uses covariates to reduce bias, often through regression modeling, propensity-based methods, or stratification. Because exposure–response curves can be sensitive to covariates, analysts must avoid including inappropriate variables that might introduce bias (e.g., mediators in some contexts).

A careful causal interpretation is supported by diagnostics and validation.

9.6 Zero-inflation and excess probability mass at boundaries

For outcomes with many zeros or at specific bounds, standard distributions may fit poorly. Zero-inflated models or hurdle approaches can separate the probability of being at the boundary from the distribution of values away from it. Boundary mass also affects probability response modeling where extreme outcomes are overrepresented. Addressing these issues improves estimation of curve shape near the lower or upper limits.

Ignoring excess mass can lead to distorted thresholds and miscalibrated uncertainty.

10 Covariates, Stratification, and Effect Modification

10.1 Incorporating covariates into exposure–response models

Covariates may influence the baseline response, the exposure sensitivity, or both. Incorporation typically occurs by adding covariates to the linear predictor, allowing parameters like asymptote or slope to depend on covariate values. In practice, this helps explain variability and reduces confounding. It also supports interpretation: changes in the curve due to covariates can be separated from changes due to exposure.

Modeling choices should reflect whether covariate effects are plausible across the exposure range.

10.2 Interactions and subgroup-specific curves

Effect modification is represented by interactions between exposure and covariates, yielding subgroup-specific curves. For instance, slope differences imply differing sensitivity, while shifts in asymptotes imply different maximal attainable responses. Subgroup modeling can improve fit but increases model complexity and requires adequate data per subgroup.

If subgroup sizes are small, partial pooling via hierarchical models can stabilize estimates.

10.3 Time-varying covariates and lagged exposures

Some covariates change over time and may influence the response dynamics. Lagged exposure effects model delayed response by incorporating exposure history rather than instantaneous exposure. This requires careful alignment between exposure timing and response measurement. Without appropriate lag handling, the curve can appear weaker or differently shaped than the true relationship.

Time-aware models improve interpretability when physiology or processes respond with delay.

10.4 Random effects and hierarchical modeling

Hierarchical models treat parameters as varying across groups or individuals, drawing from population-level distributions. Random effects capture unobserved heterogeneity and help with repeated measures and clustered data. Such models can also support borrowing strength: estimates for groups with limited data are informed by the overall distribution. This reduces instability in subgroup curve fitting.

Hierarchical structure often clarifies whether differences are systematic or mostly random.

10.5 Scaling or standardization across groups

When exposure or response scales differ across batches, instruments, or groups, standardization may be needed. Approaches include calibration to a common reference, scaling by baseline values, or fitting models with group-specific scaling parameters. Proper scaling ensures that curve comparisons reflect true differences in effect rather than measurement inconsistencies.

Scaling decisions should be documented so derived metrics remain interpretable.

11 Mixture, Joint, and Multi-exposure Models

11.1 Multi-dose/multi-concentration experimental layouts

Experiments may include multiple doses across time points, multiple concentration measures, or multiple regimes. Multi-dose layouts can be analyzed by pooling information while accounting for correlations induced by shared subjects or shared conditions. Alternatively, separate curves may be fitted for each regime and then compared using model-based summaries.

A careful layout specification is required so exposure definitions align with response measurement.

11.2 Additive vs. interactive exposure effects

When two exposures vary jointly, the response may depend on each factor additively (sum of effects) or interactively (synergistic or antagonistic combined effects). Additive models describe separate contributions, often via main effects and scaled parameters. Interaction models introduce joint terms that allow the combined effect to deviate from additivity. The choice can be guided by mechanistic plausibility and empirical fit.

Because interactions require sufficient data, sparse joint sampling can limit identifiability.

11.3 Joint modeling of multiple responses

Sometimes multiple response variables are collected, such as different biomarkers or several outcome endpoints. Joint models can exploit correlation among responses and may improve inference about exposure effects. Multi-response modeling can also distinguish whether an exposure produces a consistent pattern across endpoints or selectively affects certain measures.

Joint modeling increases complexity but can yield more coherent conclusions.

11.4 Handling correlated exposures

Correlated exposures occur when factors co-vary naturally or due to experimental design. Multi-exposure models address this by including both exposures in a joint predictor and accounting for their correlation structure. Without such handling, attributing effects to the wrong exposure can occur. Analysts may also use dimension reduction or regularization methods when the exposure set is large.

Correlated predictors require careful diagnostics to ensure stable estimates.

12 Practical Implementation and Reporting

A typical workflow includes data cleaning, exposure and response definition checks, exploratory visualization, selection of candidate model families, fitting with appropriate likelihoods, and diagnostics. After fitting, analysts compute derived quantities such as effective exposures and thresholds, along with uncertainty intervals. Finally, they validate predictive behavior using residual plots, calibration checks, or cross-validation depending on response type.

This end-to-end approach helps avoid errors that can appear when steps are skipped.

Model family selection is guided by outcome type (continuous, binary, count), expected curve shape (sigmoid, saturating, linear), and constraints such as monotonicity or asymptotic limits. Link functions connect the mean response to a linear predictor and are especially relevant for probability or count outcomes. Analysts can compare families using information criteria, predictive scoring, and diagnostic plots.

Selection should balance interpretability, goodness-of-fit, and robustness to misspecification.

12.3 Reproducibility: code, seeds, and versioning

Reproducibility requires consistent data processing steps and stable computational settings. Analysts typically record software versions, document preprocessing rules, and set random seeds when using stochastic algorithms such as bootstrapping or Bayesian sampling. Code organization—scripts for cleaning, modeling, plotting, and reporting—supports auditing and reuse.

These practices reduce accidental discrepancies across analysts or time.

12.4 Reporting standards for fitted curves and uncertainty

Good reporting includes describing the fitted model family, parameter estimates, link functions, and assumptions about error structure. Curves are typically accompanied by confidence or credible bands, and derived metrics like benchmark exposures are reported with uncertainty ranges. Reporting should also mention exposure scaling choices and how baseline or reference values were handled.

Transparent reporting supports interpretation and comparison across studies.

12.5 Common pitfalls in interpretation

Common issues include extrapolating beyond observed exposure ranges, interpreting statistical significance as practical importance, and ignoring uncertainty bands. Another pitfall is overinterpreting fitted curve shape when data are sparse at extremes. In probability models, poor calibration can lead to misleading probability estimates even if discrimination appears reasonable. Finally, correlational patterns should not be treated as causal without appropriate study design and assumptions.

Awareness of these pitfalls helps keep conclusions grounded.

13 Interpretation, Communication, and Use in Decision-Making

13.1 Translating curve features into actionable conclusions

Decision-making often uses curve features such as effective exposures, thresholds, saturation levels, and the steepness of transitions. Translating these features requires mapping them to pre-specified criteria relevant to the application, such as a target effect size or acceptable outcome probability. The curve’s shape also informs risk of rapid changes: steep regions indicate that small exposure changes may yield large response differences.

Actionable conclusions should cite the quantitative metrics and their uncertainty.

13.2 Communicating uncertainty and confidence levels

Uncertainty should be communicated through interval estimates for both parameters and derived quantities. Visual tools like confidence bands can convey the range of plausible curves, while tabulated intervals allow direct reading of effective exposures. For probability outcomes, uncertainty may be reflected in calibrated predicted probabilities across exposure levels. Clear communication helps prevent treating one fitted line as definitive.

Uncertainty is part of the result, not a footnote.

13.3 Sensitivity to model assumptions

Because exposure–response conclusions depend on modeling choices, analysts can present sensitivity results across reasonable alternatives. Examples include comparing parametric models to flexible fits, varying monotonic constraints, or adjusting for additional covariates. If derived metrics shift substantially across alternatives, conclusions should emphasize conditionality rather than asserting a single number as definitive.

Sensitivity analysis strengthens credibility by showing what is stable versus fragile.

13.4 Avoiding overinterpretation of correlation as causation

A fitted exposure–response curve summarizes association between exposure and response under the modeling assumptions. In observational or non-randomized settings, unmeasured confounding may produce an apparent relationship. Even with good curve fit, causal interpretation requires appropriate design, assumptions, or validation. Analysts should clearly separate what the curve estimates from what it can justify.

Where causality is not guaranteed, conclusions should focus on prediction or descriptive characterization.

13.5 Case studies: walkthrough of curve construction (generic)

A generic walkthrough typically starts with selecting the response type and defining exposure units and transformations. Next, exploratory plots determine whether a log scale, smoothing, or baseline adjustment is appropriate. Candidate models—such as a sigmoid parametric form and a spline-based alternative—are fitted using appropriate estimation methods. Diagnostics evaluate residual patterns, calibration for probability outcomes, and whether constraints like monotonicity hold. Finally, derived benchmark exposures are computed by inverting the fitted curve at a pre-specified effect criterion, and uncertainty intervals are reported.

The case study emphasizes transparency: documenting choices, checking assumptions, and validating that results hold under alternative plausible specifications.