More Detail Is Not Automatically More Useful
A mechanistic model can represent a biological process with one effective response or divide it into several states, cell groups, or pathway steps. The more detailed version may describe more possible biology. But if the decision depends on a predicted clinical outcome, extra description is useful only when it improves that prediction or changes how confidently the decision can be made. Detail alone does not guarantee either result.
This distinction matters because every added part of a model brings assumptions. A split into two cell groups requires assumptions about their proportions and responses. A sequence of pathway steps requires assumptions about each step and how the steps combine. If those assumptions are not supported by relevant evidence, the detailed model may look more explanatory while making a decision-relevant prediction no better than a simpler one.
The earlier material on mechanistic states and observables distinguishes internal model quantities from measured outcomes. Here the focus is narrower: whether representing more internal biology changes the prediction that a decision actually uses.
A decision-invariant refinement is not automatically pointless. It may help answer a different question, support an explanation, or become useful when conditions change. But for the current decision, it has not demonstrated added value merely by containing more states or parameters. The comparison must keep the decision question, outcome, and rule fixed.
Compare Predictions at the Decision Level
Before comparing representations, state the action rule. For example, a development team might proceed if predicted improvement in a clinical outcome is at least 5 percentage points. The rule makes the comparison concrete: a more detailed model matters to this decision if it changes whether the prediction meets the threshold, or materially changes the uncertainty relevant to that choice.
Suppose a simplified model predicts an outcome \(Y_s\), while a detailed model predicts \(Y_d\). If the decision rule is “choose action A when \(Y\) is at least \(c\),” the actions are the same whenever both predictions lie on the same side of \(c\). Matching actions do not prove that the models are biologically equivalent; they show that their difference does not change this particular threshold-based decision at those stated predictions.
A useful first check is aggregation. If the detailed model separates groups or steps that contribute to an overall outcome, calculate the aggregate explicitly and compare it with the simpler representation. This can reveal that the additional mechanisms merely distribute the same total effect among more components.
Worked Example: Splitting a Response Into Two Cell Groups
Hypothetical scenario. A simplified model predicts 60% inhibition of a disease-related signal at a candidate dose. A more detailed model divides the relevant cells into two groups. It assumes that each group makes up half of the total and predicts inhibition of 50% in the first group and 70% in the second. The decision rule is to proceed if the population-average inhibition is at least 55%.
Assumptions. For this calculation, the two groups are exhaustive, their weights are each 0.5, and the measured signal contributions are proportional to those weights. The model also assumes that an overall inhibition can be calculated as a weighted average. These are assumptions of the hypothetical comparison, not facts established by adding the groups to the model.
The detailed model’s population-average prediction is:
The simpler prediction is \(Y_s = 60\%\). Both predictions exceed the decision threshold: \(60\% \geq 55\%\). Thus, both representations recommend proceeding under this rule. The detailed model adds a description of how the response is distributed across groups, but in this example that detail does not change the average prediction or action.
The refinement could matter if the decision concerned a specific group, if the groups had unequal contributions, or if group-specific effects led to different clinical outcomes. Those possibilities would require a decision question and evidence capable of supporting them; they do not follow from the split alone.
Why Added Mechanisms Add Assumptions
A simple representation often compresses several processes into one effective quantity. A detailed representation can make those processes explicit, but then the prediction depends on more component quantities and on rules for combining them. Unless the added quantities are constrained by evidence, a more elaborate explanation may not narrow what the model predicts.
A simple derivation shows this for a serial pathway. Suppose a normalized upstream signal is 1, and two sequential steps transmit fractions \(a\) and \(b\), each between 0 and 1. Under the assumption that the steps act sequentially and multiplicatively, the downstream signal is \(ab\). A compressed model can represent the same pathway with one effective transmission fraction \(q\). If \(q=ab\), both models give the same downstream prediction.
This equality is conditional on the stated structure and combination rule. It does not show that the two underlying steps are known, or that they will remain equivalent under every perturbation. Indeed, if only the product is observed, many pairs \((a,b)\) can produce the same \(q\). A detailed account of the steps therefore needs evidence that informs those steps, especially if the intended decision depends on them separately.
Worked Example: Two Pathway Steps With the Same Net Prediction
Hypothetical scenario. A simplified model predicts that 60% of an upstream signal reaches a downstream biomarker. A detailed model represents two sequential steps, with transmission fractions \(a=0.80\) and \(b=0.75\). The decision is to select the candidate if the predicted downstream signal is no more than 65% of the upstream signal.
Assumptions. The upstream signal is normalized to 1; the steps are sequential; each fraction is constant in the scenario; and the two fractions combine multiplicatively. The simplified model’s effective fraction is assumed to represent the same conditions.
In the detailed representation:
In the simplified representation, \(Y_s=0.60\). Both satisfy the rule because \(0.60 \leq 0.65\), so both select the candidate. The detailed model attributes the net transmission to two steps, but the decision calculation receives the same value.
The assumptions are consequential. For example, the same net value could arise from \(a=0.60\), \(b=1.00\), or \(a=0.75\), \(b=0.80\). If the steps were affected differently by a new intervention, these alternatives might no longer predict the same response. Without evidence to distinguish them or a decision that depends on the separate steps, the additional parameterization does not improve this decision’s prediction.
Decision Thresholds and Prediction Differences
A small change in a prediction can be unimportant when both models remain comfortably on the same side of a decision threshold. The same numerical difference can be important when predictions are close to the threshold. Consequently, comparing model outputs without considering the decision rule can misstate the value of added detail.
Worked Example: A Small Change Near a Decision Threshold
Hypothetical scenario. A team will advance a program if predicted net clinical benefit is at least 5 benefit units. A simplified model predicts 5.4 units. A detailed model, which separates two biological contributions and a toxicity penalty, predicts 4.8 units. The unit is a hypothetical decision scale defined for this example; it is not a clinical measurement unit.
Assumptions. Both predictions refer to the same population, time horizon, and definition of net benefit, and the stated threshold is the operative decision rule. The detailed model’s component estimates are assumed to combine on this scale as specified by its construction. These assumptions make the predictions comparable, but do not establish that either model is empirically correct.
The simplified representation gives \(5.4 \geq 5\), so the rule says advance. The detailed representation gives \(4.8 < 5\), so the rule says do not advance. Here the added detail changes the decision. That makes the distinction decision-relevant, but does not by itself establish that the detailed answer is better. The component assumptions and evidence supporting the toxicity penalty need scrutiny before treating the changed recommendation as an improvement.
If both models instead predicted 5.4 and 5.3 units, they would recommend the same action under this rule. The 0.1-unit difference might still matter under another decision rule or if uncertainty around the threshold were important, but the point predictions alone would not show that it changes the current action.
A Practical Comparison and Its Limits
A disciplined comparison asks whether the added representation earns its assumptions for the decision at hand. This does not require rejecting complexity. It requires checking that the detail is linked to a prediction the decision uses, rather than treating the number of mechanisms as evidence of greater reliability.
Name the available actions, the outcome used to compare them, and any threshold or ranking rule.
Check that the simple and detailed representations predict the same quantity for the same conditions. A biomarker prediction and a clinical outcome prediction are not interchangeable just because both are numerical.
Apply its stated weights, pathway rules, and outcome mapping. Compare the result with the simple model’s prediction rather than comparing model diagrams.
Identify which new quantities or relationships the added mechanisms require and whether available evidence informs them.
Check whether predictions change the selected action, or whether any difference matters to the decision’s stated needs.
This comparison has limits. Two representations can give the same prediction under one dose, schedule, or biological condition but diverge elsewhere. A decision-invariant refinement for one question is not proof that detail is never useful. Conversely, a different prediction is not proof that the more detailed model is more accurate. The decision-maker must consider the evidence behind each representation and whether the difference is robust enough to support a different action.
Diagnostics for an unhelpful refinement include a detailed prediction that cannot be traced back to measured or otherwise supported quantities, an aggregate that does not match the simpler model despite claims that they represent the same effect, or a changed decision driven by a newly introduced assumption with little support. Check the calculation, weights, units, and endpoint mapping first. Then ask whether the added detail answers a distinct decision-relevant question. If it does not, the simpler representation may be more useful: it is easier to inspect, communicate, and use without implying unsupported precision.
Common Pitfall
Wrong reasoning: “The detailed model includes more biology, so its prediction must be more accurate and should replace the simpler one.” This confuses descriptive richness with predictive support. The additional mechanisms may rely on uncertain proportions, rates, or interactions, and their errors can affect the prediction.
Correction: Compare both representations on the same decision-relevant quantity and under the same conditions. Show how the detailed components produce the aggregate, identify the assumptions they add, and check whether the resulting difference changes the decision. Prefer the more detailed representation when its extra structure is supported and useful for the decision—not simply because it is more elaborate.
Check Your Understanding
Use the stated assumptions and decision rules in each question. Brief solutions follow the questions.
- A simple model predicts 70% inhibition. A detailed model predicts 60% inhibition in one group and 80% in another, with equal weights. The rule is to act if average inhibition is at least 65%. Do the representations lead to different actions?
- Two sequential pathway steps transmit fractions 0.5 and 0.8. Under a multiplicative assumption, what effective fraction gives the same downstream prediction? If the decision threshold is at most 45% signal remaining, does it pass?
- A simple model predicts benefit of 6 units and a detailed model predicts 5.7 units. The decision threshold is 5 units. What can be concluded about the actions, and what cannot be concluded from these point predictions alone?
- Give one reason a detailed model could be useful even if it does not change the current decision, and one reason its added detail could be misleading.
Solutions. (1) The detailed average is \((0.5)(60\%)+(0.5)(80\%)=30\%+40\%=70\%\). Both models exceed 65%, so both recommend acting. (2) The effective fraction is \(0.5 \times 0.8=0.40\), or 40% remaining. Since \(0.40 \leq 0.45\), it passes the rule. (3) Both predictions exceed 5 units, so both recommend the same action. The point predictions alone do not establish which model is more accurate or whether uncertainty could affect the decision. (4) Detail could support a future question about one pathway step; it could mislead if its new component assumptions are unsupported and create unwarranted confidence in the prediction.