Two Different Questions About Drug Effects
A drug concentration and a clinical response may vary together, but a curve through those observations does not by itself explain why. Is the concentration changing a biological process that changes the endpoint, or are other factors influencing both the measured exposure and the outcome? This distinction matters when a model is used beyond the conditions in which its data were collected.
The earlier material on mechanistic states, observables, and clinical endpoints distinguishes internal biological quantities from measurements and clinical outcomes. Here the central distinction is between a causal account of how drug action produces an effect and an empirical exposure–response fit that summarizes a pattern in observed data. Neither label guarantees that a model is reliable: the assumptions and evidence supporting its use still matter.
An empirical curve can be useful for summarizing data, estimating response over the observed exposure range, and comparing candidate descriptions. It does not automatically show that changing exposure would change response by the amount indicated by the curve. A mechanistic model makes a stronger kind of claim: its equations represent causal relationships. That claim is credible only to the extent that the mechanisms, parameters, and relevant alternative pathways are supported and the model is adequate for the intended use.
What an Exposure–Response Fit Says
A common empirical model for a response that rises toward a plateau is the Emax curve. Let \(C\) denote measured drug concentration, \(E(C)\) the mean response at that concentration, \(E_0\) the response at zero concentration, \(E_{\max}\) the maximum increase above baseline, and \(EC_{50}\) the concentration producing half of that maximum increase.
The equation gives a compact description, not a biological explanation. Its parameters characterize the curve under the conditions represented by the data. A good fit cannot establish that exposure caused the response, that the same curve applies to a different population, or that the fitted parameters correspond to a particular biological process.
Worked Example: Calculating an Empirical Emax Fit
Suppose a hypothetical study summarizes a response as a fraction from 0 to 1 and produces the fitted values \(E_0=0.10\), \(E_{\max}=0.60\), and \(EC_{50}=2.0\) concentration units. Calculate the fitted response at \(C=1.0\), \(2.0\), and \(4.0\) concentration units.
At \(C=1.0\), substitution gives \(E(1)=0.10+0.60(1.0)/(2.0+1.0)=0.10+0.20=0.30\). At \(C=2.0\), \(E(2)=0.10+0.60(2.0)/(2.0+2.0)=0.10+0.30=0.40\). At \(C=4.0\), \(E(4)=0.10+0.60(4.0)/(2.0+4.0)=0.10+0.40=0.50\).
Thus the fitted mean response increases from 0.30 to 0.50 across these concentrations, approaching the model's plateau of \(E_0+E_{\max}=0.70\). These are predictions from the specified curve, not proof that increasing an individual's concentration would cause that change. The hypothetical values also say nothing about precision; uncertainty would require information about the data and the fitting procedure.
Interpreting the curve as a causal dose or exposure effect requires more than a close fit. The exposure must precede the response, and relevant common causes of exposure and response must be controlled or otherwise addressed. There must be adequate data across the exposure values for which predictions are made. The response definition and measurement should be comparable across observations, and the functional form should describe the relationship adequately. If exposure is affected by disease severity, adherence, dose adjustment, or another determinant of response, an observed curve may combine drug effect with those influences.
A curve based on randomized assignment to dose can support causal comparisons of assigned regimens under the trial's design assumptions. But measured concentration is often not itself randomized: participants assigned the same dose can have different concentrations because of clearance, adherence, or other factors. A concentration–response association therefore needs its own justification before it is interpreted as the causal effect of changing concentration.
Why an Observed Pattern Can Mislead
The following hypothetical calculation illustrates how a difference in patient mix can create an exposure–response pattern even when exposure has no effect within either severity group. Let response mean a favorable outcome. Suppose the response probability is 0.20 for patients with severe disease and 0.80 for patients with mild disease, regardless of exposure. The high-exposure group contains 80% severe and 20% mild patients, while the low-exposure group contains 20% severe and 80% mild patients.
Worked Example: A Severity-Mix Difference Produces an Association
Calculate the response proportions in the two exposure groups and their difference. In this hypothetical example, severity influences both the exposure-group mix and the chance of response. Within each severity group, the response probability is identical at high and low exposure.
For the high-exposure group, the response proportion is the weighted average across severity groups:
For the low-exposure group, it is:
The high-minus-low difference in response proportions is \(0.32-0.68=-0.36\). On the proportion scale, the difference is \(-0.360\); equivalently, it is \(-36.0\) percentage points. The observed groups differ substantially, but by construction exposure has no effect on response within either severity group. The aggregate contrast reflects the groups' different severity composition, so it cannot be interpreted as the causal effect of exposure.
The calculation is a simple form of confounding: a third factor is associated with exposure and also predicts the outcome. One way to see the issue is to write an observed group mean as a weighted average of severity-specific means. The weights can differ between exposure groups, so a difference in group means can occur even if the severity-specific means are equal across exposure. A curve fitted to the aggregate observations may represent this pattern accurately while giving a misleading account of what changing exposure would do.
What a Mechanistic Causal Model Adds
A mechanistic model states how an intervention is represented in the system and how its effects propagate through specified biological processes to an outcome. Its equations may represent drug binding, pathway activity, feedback, or other processes, depending on the scientific question and available evidence. The important distinction is not that such a model has more equations, but that its structure is intended to encode causal relationships.
Consider a deliberately simple hypothetical model. Let \(C\) be drug concentration, \(K_d\) a binding constant, and \(\theta\) the fraction of target occupied by drug. Suppose downstream pathway activity \(A\) and a clinical score \(Y\) are described by:
Here \(A_0\) is baseline pathway activity and \(Y_0\) is the baseline clinical score. The model proposes that exposure changes target occupancy, occupancy reduces pathway activity, and the reduction in pathway activity improves the score. The coefficients and equations are assumptions about biology, not facts established by writing them down. Their use requires evidence that the target is engaged, that the downstream relationship is adequate, and that important alternative routes to the endpoint are not omitted or misrepresented.
Worked Example: Following a Hypothetical Mechanistic Model
Suppose \(C=3.0\) nM, \(K_d=3.0\) nM, \(A_0=100\) pathway units, and \(Y_0=50\) score units. Use the model above to calculate occupancy, pathway activity, and the predicted score.
First, \(\theta=3.0/(3.0+3.0)=0.50\), a dimensionless occupancy fraction. Next, \(A=100[1-0.40(0.50)]=100(0.80)=80\) pathway units. The pathway reduction is \(A_0-A=100-80=20\) pathway units. Finally, \(Y=50-0.25(20)=45\) score units.
Under these equations and parameter values, the model predicts a score of 45 rather than the baseline score of 50. That prediction follows from the represented causal chain; it is not confirmed merely because the arithmetic is coherent. If the drug does not engage this target, if another process dominates the endpoint, or if the pathway-to-score relationship differs in the population, the prediction may fail.
A mechanistic representation can make a changed setting more interpretable when the model explicitly represents what differs and when the relevant biology and parameters remain valid. For example, if a second setting has a different baseline pathway activity, an equation containing \(A_0\) gives a place to represent that difference. But the model cannot safely predict the setting merely because it has a biological label. Unrepresented changes, such as a different pathway response or a co-treatment effect, can invalidate the assumed relationships. More mechanistic detail does not remove the need to check whether the represented system matches the intended use.
Assumptions, Diagnostics, and Model Choice
Both approaches require scrutiny, but their failure modes differ. For an empirical fit, inspect whether observations cover the exposure range used for predictions, whether the chosen curve captures the data pattern, and whether important covariates or group differences remain. Residual patterns, influential observations, and large uncertainty near the edges of the observed range can warn that the fit is inadequate. A good residual plot does not diagnose unmeasured confounding, however; that is a design and causal-identification problem, not simply a curve-fitting problem.
For a mechanistic model, check whether the proposed processes and directions of effect are biologically plausible and whether predicted intermediate quantities can be compared with measurements. Check that parameter values have empirical support and that the model behaves sensibly across the conditions for which it will be used. A mismatch between predicted and observed intermediate or clinical quantities can expose an incorrect mechanism, a missing process, or an unsuitable parameter value. Agreement with one data set is not sufficient proof that the causal structure is correct; multiple explanations may produce similar observed outcomes.
| Question | Empirical exposure–response fit | Mechanistic causal model |
|---|---|---|
| Primary aim | Summarize a pattern in observed exposure and response | Represent biological processes linking intervention and outcome |
| What supports interpretation? | Data quality, adequate fit, relevant design, and control of confounding for causal claims | Supported causal structure, credible parameters, and adequate representation of the relevant setting |
| Typical weakness | Association may reflect patient mix or apply only to the sampled range | Incorrect or unsupported mechanisms can give precise but unreliable predictions |
| Use in a changed setting | Requires evidence that the observed relationship transports to that setting | May represent relevant biological differences, but only if they are included and supported |
The approaches are not mutually exclusive. An empirical curve can summarize a data set or provide a comparison for a more mechanistic representation. A mechanistic model can be assessed against observed exposure–response patterns. But their claims should not be conflated: an empirical association does not become causal because it is fitted well, and a mechanistic model does not become validated because it is biologically detailed.
A practical choice starts with the question. If the goal is to describe outcomes in a similar population and exposure range, an empirical fit may be adequate. If a decision depends on explaining how an intervention acts or on representing a particular biological change, a causal mechanistic account may be needed. In either case, state the population, intervention, outcome, and conditions represented, and limit conclusions to what the assumptions and evidence support.
Common Pitfall
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- An empirical curve fits observed concentration and response data closely. Name two reasons this alone does not establish a causal exposure effect.
- In a hypothetical setting, severe patients respond with probability 0.30 and mild patients with probability 0.70, independent of exposure. The high-exposure group is 60% severe and the low-exposure group is 20% severe. Calculate each group's response proportion and the high-minus-low difference.
- What does \(EC_{50}\) mean in the Emax model defined here? Does it necessarily equal a biological binding constant?
- A mechanistic model predicts an endpoint through target occupancy and pathway activity. Name one assumption about the biological system and one diagnostic that could help assess the model.
Solutions. (1) The observed association may be confounded, and the fitted curve may only be supported over the observed exposure range or in the sampled setting. A good fit does not resolve either issue. (2) High exposure: \((0.60)(0.30)+(0.40)(0.70)=0.18+0.28=0.46\). Low exposure: \((0.20)(0.30)+(0.80)(0.70)=0.06+0.56=0.62\). The high-minus-low difference is \(0.46-0.62=-0.16\), or \(-16\) percentage points. Because response probabilities are exposure-independent within each severity group, the aggregate difference arises from the different severity mixes. (3) \(EC_{50}\) is the concentration at which the modeled increase above \(E_0\) reaches half of \(E_{\max}\). It is a curve parameter and need not equal a binding constant. (4) An assumption could be that drug occupancy changes pathway activity in the specified direction and by the specified relationship. A diagnostic could compare predicted pathway activity with measured pathway activity across relevant conditions; disagreement would challenge the model's adequacy.