Why Profile Likelihood Helps With Weak Identification
A point estimate can look precise even when the data permit a broad range of parameter combinations. This can happen in PK/PD analysis when the observed concentrations cover only a narrow range, when PK sampling is sparse, or when uncertainty in exposure can be offset by a change in PD parameters. A standard error summarizes local curvature near an estimate; it may not show a long, curved trade-off that extends far from that estimate.
A likelihood profile makes that trade-off visible. For each fixed value of a parameter, it asks how well the model can fit after the other parameters are re-optimized. Comparing profiles from a sequential analysis and a joint analysis is especially useful: a sequential PD fit typically conditions on estimated PK parameters, whereas a joint fit can adjust PK and PD parameters together.
In a joint PK/PD model, \(\psi\) might be EC50, Emax, or clearance (CL), while \(\lambda\) contains the other PK and PD parameters and any estimated variance or residual-error parameters. If the model has separate PK and PD data contributions, the joint log-likelihood may be written as their sum, but the profile still comes from optimizing the full joint likelihood over all nuisance parameters at each fixed \(\psi\).
Likelihood-Ratio Intervals and Their Conditions
For a single parameter, a commonly used approximate 95% likelihood-ratio confidence set contains the values satisfying \(q(\psi)\leq 3.84\). The cutoff is the 95th percentile of a chi-square distribution with one degree of freedom. More generally, a nominal \(100(1-\alpha)\%\) set uses the \(1-\alpha\) quantile of that distribution. Under regular likelihood theory, the likelihood-ratio statistic is approximately chi-square distributed when the null value is true.
That approximation relies on conditions such as a correctly specified likelihood, an interior parameter value, sufficient information, and local identifiability. Weak identification is precisely a situation in which the usual approximation may be poor: the profile can be unusually flat, highly asymmetric, or poorly approximated by a parabola. Parameters at boundaries, such as a variance constrained to be nonnegative, also require particular care. The value 3.84 is therefore a useful reference, not a guarantee of accurate coverage in every weakly identified PK/PD problem. Simulation or bootstrap calibration can be considered when the standard approximation is doubtful.
Worked Example: Profiling PD Potency
Suppose a hypothetical joint PK/PD analysis estimates EC50 in mg/L. The unrestricted maximum occurs at EC50 \(=3.0\) mg/L with log-likelihood \(\ell=-120.0\). At each listed fixed EC50, all other parameters are re-optimized. The resulting values are shown below.
| Fixed EC50 (mg/L) | Profile log-likelihood | Likelihood-ratio statistic \(q\) |
|---|---|---|
| 1.0 | -122.6 | 5.2 |
| 2.0 | -121.0 | 2.0 |
| 3.0 | -120.0 | 0.0 |
| 4.0 | -120.5 | 1.0 |
| 5.0 | -121.5 | 3.0 |
| 6.0 | -122.6 | 5.2 |
For EC50 \(=2.0\) mg/L, for example, $$ q(2.0)=2\{ -120.0-(-121.0)\}=2.0. $$ The same calculation at 5.0 mg/L gives \(q(5.0)=2\{-120.0-(-121.5)\}=3.0\). Both are below 3.84, while the grid values at 1.0 and 6.0 mg/L are above it.
Linear interpolation in \(q\) between adjacent grid points estimates where the profile crosses 3.84. On the lower side, this is $$ 1.0+\frac{5.2-3.84}{5.2-2.0}(2.0-1.0)=1.425\ {\rm mg/L}. $$ On the upper side, it is $$ 5.0+\frac{3.84-3.0}{5.2-3.0}(6.0-5.0)=5.382\ {\rm mg/L}. $$ The approximate profile interval is therefore 1.43 to 5.38 mg/L. This interpolation is a numerical approximation based on the stated grid, not an exact boundary calculation. Under regularity conditions the interval is a nominal 95% likelihood-ratio interval; it does not mean there is a 95% probability that the fixed, unknown EC50 lies inside this particular interval.
Sequential and Joint Profiles Answer Different Questions
In a sequential analysis, PK parameters are estimated first and then treated as fixed inputs in the PD fit. Profiling EC50 in that second-stage PD likelihood allows the other PD parameters to adjust, but it does not generally allow PK parameters to vary. The resulting profile is conditional on the chosen PK estimates. It is not a profile of the full joint likelihood, and its interval does not automatically include uncertainty in PK estimates.
A joint profile instead fixes the parameter of interest and re-optimizes the remaining PK and PD parameters together. If a change in CL can be offset by a change in EC50, that adjustment can keep the joint likelihood relatively high and produce a broad profile. This is one way profile likelihood displays the uncertainty-transfer issue addressed by earlier material on plug-in PK estimates and the lost-uncertainty problem.
Worked Example: Comparing Conditional and Joint EC50 Profiles
Consider hypothetical concentration and response data fitted with an Emax PD model. A sequential analysis fixes PK estimates and gives a conditional EC50 profile with a maximum at 2.5 mg/L. Suppose its likelihood-ratio curve crosses 3.84 at approximately 2.1 and 3.0 mg/L. The resulting conditional interval is approximately 2.1 to 3.0 mg/L.
Now suppose the unrestricted joint maximum also occurs at EC50 \(=2.5\) mg/L, with full joint log-likelihood \(-84.0\). Re-optimizing all nuisance parameters gives profile log-likelihoods of \(-86.0\) at 1.0 mg/L, \(-84.0\) at 2.5 mg/L, \(-84.8\) at 5.0 mg/L, and \(-86.5\) at 10.0 mg/L. The corresponding joint statistics are $$ q(1.0)=2\{-84.0-(-86.0)\}=4.0, $$ $$ q(5.0)=2\{-84.0-(-84.8)\}=1.6,\qquad q(10.0)=2\{-84.0-(-86.5)\}=5.0. $$
The joint profile therefore retains 5.0 mg/L at the 3.84 reference cutoff, even though the conditional sequential interval ended near 3.0 mg/L. The difference is consistent with the joint fit finding parameter combinations that preserve fit while PK and PD parameters adjust together. It does not by itself prove that the joint interval is better calibrated: both profiles depend on their likelihood assumptions, and the chi-square approximation may need checking when identification is weak.
Reading Parameter Trade-Offs
Profile curves are often more informative when accompanied by the nuisance-parameter estimates found at each fixed value of the parameter of interest. A broad profile for EC50 alone tells us that several potency values fit reasonably well; recording the corresponding Emax, CL, or other fitted parameters can reveal how the model accommodates those values.
For an Emax model, $$ E(C)=E_0+\frac{E_{\max}C}{EC_{50}+C}, $$ where \(E(C)\) is the modeled response at concentration \(C\), \(E_0\) is baseline response, and Emax is the maximum increase above baseline. If all observed concentrations are small relative to EC50, then \(C/EC_{50}\) is small and $$ E(C)-E_0 =\frac{E_{\max}C}{EC_{50}+C} \approx \frac{E_{\max}}{EC_{50}}C. $$ In that range, observations mainly inform the ratio \(E_{\max}/EC_{50}\), rather than the two parameters separately. Increasing both parameters in a way that preserves approximately the same ratio can yield similar predicted responses. This gives a shallow, curved likelihood ridge rather than a sharply defined maximum.
Worked Example: A Shallow Emax–EC50 Trade-Off
Suppose a hypothetical study observes concentrations only between 0.1 and 0.5 mg/L. Compare two candidate parameter pairs: \(E_{\max}=20\) response units with EC50 \(=5\) mg/L, and \(E_{\max}=40\) response units with EC50 \(=10\) mg/L. Both have \(E_{\max}/EC_{50}=4\) response units per mg/L.
At \(C=0.2\) mg/L, the first pair predicts an increase above baseline of $$ \frac{20(0.2)}{5+0.2}=\frac{4}{5.2}=0.7692\ {\rm response\ units}, $$ while the second predicts $$ \frac{40(0.2)}{10+0.2}=\frac{8}{10.2}=0.7843\ {\rm response\ units}. $$ At \(C=0.5\) mg/L, the corresponding increases are \(10/5.5=1.818\) and \(20/10.5=1.905\) response units. These predictions are close, especially relative to typical residual variation. Thus low-concentration data alone may not distinguish the pairs strongly. At concentrations approaching or exceeding EC50, the saturation behavior becomes more informative about Emax and EC50 separately.
The calculation establishes why a shallow ridge is plausible; it does not establish an actual likelihood-ratio value. That value must be obtained by evaluating the specified likelihood and profiling the nuisance parameters against the unrestricted maximum.
PK parameters can participate in similar trade-offs. For example, in a simplified linear PK setting with fixed dosing conditions, concentrations may scale approximately inversely with CL. If EC50 changes along with CL, the predicted PD response can remain similar over the observed exposure range. The exact relationship depends on the PK model, design, and data; it should be identified from the fitted profile rather than assumed from a simple proportionality.
Worked Example: Profiling Clearance Along a Joint Ridge
Suppose a hypothetical joint model has an unrestricted maximum at CL \(=5.0\) L/h and EC50 \(=2.0\) mg/L, with log-likelihood \(-50.0\). At each fixed CL below, EC50 and all other nuisance parameters are re-optimized. The table reports the resulting joint profile. In particular, each \(q\) value is calculated relative to the stated unrestricted joint maximum.
| Fixed CL (L/h) | Profile EC50 (mg/L) | Profile log-likelihood | Joint \(q\) |
|---|---|---|---|
| 4.0 | 2.50 | -50.5 | 1.0 |
| 5.0 | 2.00 | -50.0 | 0.0 |
| 6.0 | 1.67 | -50.75 | 1.5 |
| 8.0 | 1.25 | -52.1 | 4.2 |
For CL \(=6.0\) L/h, the calculation is \(q(6.0)=2\{-50.0-(-50.75)\}=1.5\). For CL \(=8.0\) L/h, \(q(8.0)=2\{-50.0-(-52.1)\}=4.2\). The fitted EC50 decreases as CL increases, showing the estimated trade-off along the profile. The profile is within the 3.84 reference cutoff at 6.0 L/h and outside it at 8.0 L/h. This conclusion uses the explicitly stated unrestricted joint maximum as the reference; a likelihood difference between two arbitrary parameter combinations alone would not establish either combination’s joint likelihood-ratio statistic.
Diagnostics, Alternatives, and Limitations
A profile should be computed by re-optimizing nuisance parameters at every fixed value, preferably from multiple starting points when the likelihood may have several local maxima. Check that the optimizer converges, that constraints are respected, and that the reported profile is the best likelihood found at each grid point. Plot both the profile log-likelihood and the likelihood-ratio statistic against the parameter value. Also inspect the nuisance estimates along the curve: abrupt jumps may indicate competing local optima or a change in the fitted regime, rather than a smooth biological trade-off.
Common warning signs include a nearly flat profile, a confidence set that reaches a parameter boundary, strongly unequal lower and upper limits, disconnected regions below the cutoff, and profiles that change substantially with starting values or grid refinement. A flat or unbounded profile indicates that the data and model do not constrain the parameter adequately over the explored range. A disconnected set can be a real feature of a nonstandard likelihood, but it should also trigger checks for optimization failure, parameter constraints, and model implementation errors.
A Wald interval uses the estimated parameter and its local standard error, often in the form estimate plus or minus a normal quantile times the standard error. It is computationally convenient, but it relies on a locally quadratic log-likelihood and can be misleading when the profile is curved or asymmetric. The profile method uses the likelihood away from the maximum, but it requires repeated optimization and still depends on the model and the validity of the likelihood-ratio reference distribution. Bootstrap methods provide another way to assess uncertainty and can help evaluate calibration, although they also require a suitable model and successful refitting. These approaches are complementary rather than interchangeable.
Practical Workflow
State whether the target is a PD parameter such as EC50 or Emax, or a PK parameter such as CL. For joint inference, use the full joint likelihood.
Use a grid wide enough to show both sides of the likelihood maximum, respecting the parameter’s units and valid range.
At each fixed value, maximize over the remaining parameters and check convergence, constraints, and sensitivity to starting values.
Compute \(q(\psi)=2\{\ell(\hat{\psi},\hat{\lambda})-\ell_p(\psi)\}\). Compare it with an appropriate reference cutoff, while checking whether regularity conditions are plausible.
Plot the profile and track nuisance estimates. Report flatness, asymmetry, boundaries, or numerical instability rather than relying only on a point estimate and standard error.
Check Your Understanding
Answer each question, then compare with the solutions below.
- At a fixed parameter value, what is optimized when constructing a profile likelihood?
- A joint maximum has log-likelihood \(-30.0\), and the profile log-likelihood at a fixed value is \(-32.0\). Calculate \(q\) and compare it with 3.84.
- Why can a sequential EC50 profile be narrower than a joint EC50 profile?
- What trade-off may occur when observed concentrations are all small relative to EC50 in an Emax model?
- Name two diagnostics that should be checked before interpreting a flat or disconnected profile.
Solutions. (1) All nuisance parameters are re-optimized while the parameter of interest is held fixed. (2) \(q=2\{-30.0-(-32.0)\}=4.0\), which exceeds 3.84; under regularity conditions this fixed value is outside the approximate 95% likelihood-ratio set. (3) The sequential fit conditions on PK estimates instead of allowing PK parameters to adjust, so it can omit PK uncertainty and PK–PD trade-offs. (4) The data may mainly identify \(E_{\max}/EC_{50}\), allowing different Emax and EC50 values to produce similar predictions. (5) Check optimizer convergence and multiple starting values; also check grid coverage, parameter constraints, and whether nuisance estimates jump abruptly.