Selecting Starting Values for Emax Estimation

Why Starting Values Matter

Estimating an Emax curve is an iterative numerical task: an algorithm begins with parameter values and updates them to improve agreement between model predictions and observed responses. Starting values are not results, and they need not be exact. They do need to put the model in a plausible region of parameter space. A start that gives the wrong response direction, an implausible plateau, or a transition at concentrations far from those observed can make estimation inefficient or lead to a poor or unsuccessful fit.

This tutorial uses the baseline-relative stimulatory Emax model with Hill coefficient fixed at 1. Concentration is denoted by $C$, and the response predicted at $C$ by $E(C)$. The parameter $E_0$ is the response at zero concentration, $E_{\max}$ is the maximal change from that baseline, and $EC_{50}$ is the concentration producing half of that maximal change. These parameter meanings follow the baseline-relative parameterization; in an absolute-plateau parameterization, the plateau itself is a parameter and the response increment is its difference from $E_0$.

Definition: A starting value is an initial numerical value supplied for a model parameter before estimation begins. For the baseline-relative stimulatory Emax model, $$ E(C)=E_0+\frac{E_{\max}C}{EC_{50}+C}, $$ a natural starting-value set is an approximate baseline, an approximate response increment to the plateau, and a positive concentration near the midpoint of that increment. Starting values are provisional computational inputs, not estimates with established uncertainty.

The model assumes a direct, saturating concentration-effect relationship over the observations, with a common response scale and a meaningful concentration scale. The basic stimulatory model also implies $E_{\max}>0$ and $EC_{50}>0$. It does not describe delayed effects or hysteresis, and a visually curved response pattern alone does not establish that the direct Emax assumptions hold. Inspect the data and study context before using the curve to initialize estimation.

Reading the Plot Before Calculating Starts

Start by plotting response against concentration on a scale that makes the observed range and curve shape visible. If concentrations span orders of magnitude, a logarithmic horizontal axis often helps display the transition; zero concentration must be shown separately because its logarithm is undefined. The plot helps distinguish three features: the response near zero concentration, the direction and size of the response change, and whether the upper end appears to approach a plateau.

For the baseline-relative model, a first approximation to $E_0$ comes from control observations or responses at the lowest concentrations, especially when zero concentration is observed. A provisional $E_{\max}$ is the difference between a plausible upper plateau and that baseline—not simply the largest response in the data if saturation is not apparent. When observations approach the plateau, the concentration at which the response is about halfway between baseline and plateau gives a useful $EC_{50}$ start.

Derivation: At $C=EC_{50}$, the concentration-dependent fraction of the response increment is $$ \frac{C}{EC_{50}+C} = \frac{EC_{50}}{EC_{50}+EC_{50}} = \frac{1}{2}. $$ Therefore $E(EC_{50})=E_0+E_{\max}/2$. This midpoint relationship explains how a plot can suggest an $EC_{50}$ starting value. It is a direct consequence of the model, not a claim that a sparse or noisy plot identifies the parameter precisely.

Worked Example: Starts From an Observed Plateau

Suppose a hypothetical direct-effect experiment measures concentration in mg/L and response in units. The plotted observations are:

Concentration (mg/L)	Response (units)
0	10
1	20
2	27
4	36
8	44
16	50
32	54

The zero-concentration response is 10 units, so use $E_0=10$ units as a reasonable start. Responses rise and the last two observations increase by only 4 units despite the concentration doubling. This is compatible with an approach to a plateau, although the data do not prove that the plateau is exactly 55 units. Use 55 units as a provisional plateau; then the response increment start is $$ E_{\max}=55-10=45\text{ units}. $$

Halfway from baseline to this provisional plateau is $$ 10+\frac{45}{2}=32.5\text{ units}. $$ The response of 32.5 units lies between 27 units at 2 mg/L and 36 units at 4 mg/L. Interpolating on the log-concentration scale, its fractional position is $(32.5-27)/(36-27)=5.5/9=0.6111$. Thus a provisional concentration is $$ EC_{50}=2\left(\frac{4}{2}\right)^{0.6111}\approx3.05\text{ mg/L}. $$ A practical starting set is $E_0=10$ units, $E_{\max}=45$ units, and $EC_{50}=3.05$ mg/L. These values are internally consistent with the model's baseline, increment, and midpoint definitions. They remain approximations, not reported estimates.

When the Plateau Is Not Observed

If the largest concentrations still show a substantial rise, the observed maximum is not a reliable estimate of the asymptote. Setting the plateau equal to the largest observed response can underestimate $E_{\max}$, which in turn can make the inferred midpoint too low. Conversely, an arbitrarily large plateau can push the implied midpoint beyond the observed concentration range. This is the practical consequence of weak information about saturation: several combinations of $E_{\max}$ and $EC_{50}$ may describe much of the observed range similarly.

In that situation, use the response range to propose one or more plausible plateau values, and label the resulting values as exploratory starts. Check whether the implied midpoint is compatible with the plotted transition. If it is not, revise the proposed plateau or use multiple reasonable starting sets. Do not present one guessed plateau or midpoint as if the data had established it. The earlier material on identifiability of Emax and EC50 explains why sparse coverage of the plateau limits what the data can determine.

Worked Example: A Rising Curve With No Clear Plateau

Consider hypothetical responses of 12, 18, 30, 46, and 65 units at concentrations 0.1, 0.3, 1, 3, and 10 mg/L, respectively. The response is still rising at 10 mg/L, so the observed maximum of 65 units does not establish the plateau. The response at zero concentration was not measured either; for illustration, suppose the study's control observations support an approximate $E_0$ of 12 units. This is an assumption for constructing starts, not something established by these five observations alone.

One exploratory choice is a plateau of 82 units. It gives $$ E_{\max}=82-12=70\text{ units}, \qquad E_0+\frac{E_{\max}}{2}=12+35=47\text{ units}. $$ The midpoint response, 47 units, is just above the observed 46 units at 3 mg/L. On the plotted log-concentration scale, this suggests an $EC_{50}$ start slightly above 3 mg/L; for example, 3.1 mg/L. This is a plausible starting set, not evidence that the plateau is 82 units.

To see the sensitivity to the plateau assumption, try a second provisional plateau of 132 units. Then $E_{\max}=132-12=120$ units, and the midpoint response is $12+120/2=72$ units, which lies above every observed response. This choice implies a midpoint beyond the observed range and is less directly supported by the plot. It may still be worth testing as a deliberately broader alternative, but it should not be mistaken for a data-based midpoint. Comparing such starts can reveal whether the estimation is sensitive to assumptions about the unobserved plateau.

The correct lesson is not to select the 82-unit plateau because it gives a more convenient answer. It is to recognize that the high-concentration data have not shown a plateau, document plausible alternatives, and interpret any resulting estimate in light of that limitation.

Checking Direction, Scale, and Plausibility

Before estimation, check each proposed value against the observed response scale and concentration units. For a stimulatory curve, a positive $E_{\max}$ gives an increasing response toward $E_0+E_{\max}$. For an inhibitory response expressed with the same baseline-relative equation, the response change is negative; the sign of $E_{\max}$ must match the direction of change. A wrong sign can make the initial curve move away from the observations rather than toward them.

Also check that $EC_{50}$ uses the same concentration units as $C$. If the data are in mg/L, a value entered in micrograms/L without conversion is 1,000-fold too large. That mismatch moves the model's transition far beyond the plotted range. Response units must also match: subtracting a baseline in one scale from a plateau in another produces an invalid $E_{\max}$ start.

Worked Example: Starting Values for a Decreasing Response

Suppose a hypothetical inhibitory direct-effect dataset has a baseline of 100 units and responses of 82, 66, and 54 units at concentrations 1, 4, and 16 mg/L. The response decreases with concentration. The high-concentration observation suggests a plateau near 50 units, so a provisional signed response change is $$ E_{\max}=50-100=-50\text{ units}. $$ The halfway response is $100+(-50/2)=75$ units. It lies between 82 units at 1 mg/L and 66 units at 4 mg/L. Interpolating on the log-concentration scale gives a fraction $(82-75)/(82-66)=7/16=0.4375$, so $$ EC_{50}=1\left(\frac{4}{1}\right)^{0.4375}\approx1.83\text{ mg/L}. $$ A plausible start is therefore $E_0=100$ units, $E_{\max}=-50$ units, and $EC_{50}=1.83$ mg/L. If instead $E_{\max}$ were incorrectly started at $+50$ units, the model would predict an increase toward 150 units, opposite to the observed direction. This example uses the baseline-relative signed-effect convention; the interpretation of a parameter labelled $E_{\max}$ must always match the parameterization being used.

Diagnosing Poor Starts and Comparing Approaches

A poor starting set can show itself before any final result is considered: the initial curve may be nearly flat over all observed concentrations, may turn at a concentration far outside the data, may predict responses on the wrong side of the baseline, or may imply an implausible plateau. During estimation, warnings, failure to converge, estimates stuck at extreme values, or materially different results from modestly different starts all call for investigation. None of these signs alone proves the model is wrong; they can indicate a poor start, insufficient information, an unsuitable model, or a combination of these.

Useful checks include plotting the starting curve over the data, confirming units and signs, and repeating estimation from several plausible starts when the plateau is uncertain. If different starts lead to substantially different fitted parameter values or predictions, do not select the most convenient result without investigation. Review whether the concentrations span the transition and plateau, whether the assumed direct-effect curve is suitable, and whether the parameters are identifiable in the observed range.

Three approaches are commonly contrasted. Plot-guided starts use the baseline, apparent plateau, and midpoint and are interpretable, but depend on visual judgment. Formula-based starts apply the model's midpoint relationship or solve for parameters from selected response points; these can be reproducible, but are sensitive to noise and are unreliable when the plateau is absent. Multiple plausible starts provide a practical check on sensitivity, but do not create information the experiment did not collect. Automated defaults may be convenient, but cannot know whether a response scale, sign convention, or concentration unit has been supplied incorrectly.

Starting values do not replace model checking or establish that a parameter is estimable. In particular, the selection task is not to force a desired answer by choosing convenient numbers. It is to begin with values consistent with the observed range and the stated model, then recognize when the data provide little guidance. Detailed fitting algorithms, weighting, parameter constraints, and formal uncertainty assessment are separate issues.

Common Pitfalls

Using the largest observation as the plateau automatically. A rising high-concentration trend gives no guarantee that the maximum observed response is close to the asymptote. Use it as a provisional clue only when the plot supports that interpretation.
Confusing the plateau with the response increment. In the baseline-relative model, the plateau is $E_0+E_{\max}$; $E_{\max}$ is the change from baseline.
Choosing a midpoint without first choosing a plateau. Half of an increment is defined relative to the assumed plateau. If the plateau changes, the midpoint response and the corresponding concentration can change.
Ignoring direction or units. Check that the sign of $E_{\max}$ agrees with the response trend and that $EC_{50}$ and measured concentrations use the same units.
Treating a successful numerical fit as proof of a good start or a well-supported model. A computation can finish even when the data weakly identify the parameters or the model assumptions are unsuitable.

Key takeaway: Use observed low-concentration responses to guide $E_0$, a defensible plateau assumption to guide $E_{\max}$, and the half-increment point to guide $EC_{50}$. When saturation is not visible, make the uncertainty in these starts explicit and check more than one plausible set.

Check Your Understanding

Answer each question, then compare with the brief solutions below.

In a baseline-relative Emax model, $E_0=8$ units and a plausible plateau is 38 units. What is the corresponding $E_{\max}$, and what response marks the midpoint?
Why is the largest observed response not necessarily a sound estimate of the plateau when the curve is still rising?
A stimulatory curve is recorded with concentration in mg/L. What is one consequence of entering an $EC_{50}$ start in micrograms/L without converting units?
A response falls from a baseline of 90 units toward an apparent plateau of 50 units. What sign should the baseline-relative $E_{\max}$ start have?
Two reasonable starting sets give substantially different parameter values. Does choosing the set that produces a completed fit resolve the issue? State one appropriate next check.

Solutions. (1) $E_{\max}=38-8=30$ units; the midpoint response is $8+30/2=23$ units. (2) Continued increase at the highest concentrations suggests the asymptote may be above the observed maximum, so the data have not established a plateau. (3) The entered value is 1,000 times its mg/L equivalent, moving the implied transition far from its intended concentration range. (4) The response change is $50-90=-40$ units, so the start should be negative. (5) No. Check the starting curves, units, concentration coverage, and whether the data sufficiently inform the plateau and transition.