Why Wald Intervals Need a Diagnostic Check
A fitted QSP model can produce a parameter estimate and an approximate standard error, but those two numbers do not guarantee that a symmetric confidence interval is informative. A Wald interval is built from the curvature of the likelihood near its maximum. It is most useful when that local curvature describes the likelihood well over the range covered by the interval. Strong nonlinearity, compensating parameter changes, or a nearby biological or mathematical bound can break that approximation.
The calculation is still valuable: it provides a fast, transparent summary of local uncertainty and makes use of the covariance among fitted parameters. The key is to calculate it correctly, state what it assumes, and check whether its approximation is plausible before interpreting the result.
In a QSP application, the likelihood may be formed from measurements whose model predictions come from an ODE system. The Wald calculation applies to the fitted parameter likelihood; it does not require the ODE solution itself to be linear in the parameters. It does require the likelihood around the estimate to be sufficiently well approximated by a quadratic, and the estimate to be in a regular interior region of parameter space.
Derivation and Conditions
The justification is a local approximation. Write \(U(\theta)=\nabla\ell(\theta)\) for the score. At an interior maximum, \(U(\hat{\theta})=0\). A first-order Taylor expansion of the score around the true parameter \(\theta\), together with the approximation of the Hessian by negative information, gives
Solving this approximation gives \(\hat{\theta}-\theta\approx I(\theta)^{-1}U(\theta)\). Under regularity conditions, the score is approximately distributed as \(N(0,I(\theta))\) in repeated samples. Therefore the estimator is approximately distributed as \(N(\theta,I(\theta)^{-1})\). Replacing the unknown information by its value at the fitted estimate yields the estimated covariance matrix used for Wald intervals. In practice, either the observed or expected information may be used, as appropriate for the fitted likelihood and software; the choice should be reported.
This argument relies on a correctly specified likelihood, a sufficiently informative data set, a smooth likelihood near the estimate, an interior true parameter, and nonsingular information. The normal approximation is generally an asymptotic one: a finite data set can have a likelihood shape that is far from quadratic even when the Hessian exists. The information matrix should be positive definite at a well-behaved interior maximum. A nearly singular or numerically unstable matrix is a warning that the local covariance calculation may be unreliable.
Calculating the Intervals and Reading the Covariance
For a vector of estimates, the diagonal entries of the inverse information matrix are estimated marginal variances. Their square roots are standard errors in the units of the corresponding parameters. Off-diagonal entries are estimated covariances; they matter for joint uncertainty and reveal whether parameter estimates tend to move together or in opposite directions.
Worked Example: Wald Intervals With Strong Parameter Correlation
Suppose a hypothetical fit estimates two unconstrained, dimensionless parameters, \(\theta_1\) and \(\theta_2\), as \(1.20\) and \(0.80\). The observed information at the maximum is
The determinant is \(100(100)-98(98)=396\), so the inverse is
where the displayed decimal entries are rounded. Each marginal standard error is \(\sqrt{100/396}=0.5025\). Using \(z_{0.975}=1.96\), the 95% Wald interval for \(\theta_1\) is
The interval for \(\theta_2\) is \(0.80\pm0.9849=[-0.1849,\ 1.7849]\). The estimated correlation is calculated from the unrounded inverse entries:
Thus, the estimates are strongly negatively correlated. The information matrix has eigenvalues \(198\) and \(2\): information is much lower in the direction proportional to \((1,-1)\) than in the direction proportional to \((1,1)\). This is consistent with a likelihood ridge along which an increase in one parameter can be offset by a decrease in the other. The marginal intervals describe uncertainty one parameter at a time; they do not show the full joint shape. The intervals are meaningful as approximate intervals only if the local normal approximation is adequate and the parameter coordinates are indeed unconstrained.
A confidence interval is a procedure with a repeated-sampling coverage target, not a probability statement about a fixed parameter after the data have been observed. For example, a 95% Wald procedure has approximately 95% coverage under its assumptions. This does not mean there is a 95% probability that the particular fixed parameter lies in the calculated interval.
Failure Mode: Nonlinear Likelihoods
A Hessian describes curvature at one point. It cannot, by itself, show whether the likelihood remains approximately quadratic across the interval. If the likelihood is skewed or changes curvature rapidly, a symmetric Wald interval can give a poor account of uncertainty. This is particularly visible when the interval extends into a region where the parameter or the likelihood behaves very differently from its behavior near the estimate.
Worked Example: A Skewed Binomial Likelihood
In hypothetical data, 1 of 20 independent subjects has a binary response, with common response probability \(p\). The likelihood is proportional to \(p(1-p)^{19}\), and its maximum is at \(\hat{p}=1/20=0.05\). The observed information for \(p\) at this estimate is
The estimated standard error is \(1/\sqrt{421.1}=0.04873\). The nominal 95% Wald interval is
The interval includes negative probabilities, which are impossible. The difficulty is not an arithmetic error: with only one response, the likelihood is asymmetric near its maximum, and a symmetric normal approximation is poor over this range. The interval should not be silently clipped at zero and then treated as having its original 95% coverage. An interval method that respects the parameter space, or a likelihood-based analysis, is more appropriate to investigate this case.
Failure Mode: Proximity to a Parameter Bound
Many QSP parameters have restrictions such as a nonnegative rate or a fraction restricted to \([0,1]\). A standard Wald interval extends equally far in both directions from the estimate. If the estimate is close to a bound, the interval may include impossible values, or its symmetry may misrepresent uncertainty even if it stays within the allowed range. A constrained estimate at the boundary is especially problematic: the regular interior normal approximation used in the derivation does not apply in its usual form.
Worked Example: A Nonnegative Degradation Rate
Suppose a hypothetical model estimates a degradation rate \(k\), constrained by \(k\geq0\), as \(\hat{k}=0.080\ \mathrm{day}^{-1}\), with an information-based standard error of \(0.070\ \mathrm{day}^{-1}\). The 95% Wald interval is
The negative lower endpoint is outside the parameter space. This is a direct signal not to interpret the interval as a plausible range for the degradation rate. Simply replacing the lower endpoint by zero does not establish the coverage of the modified interval. The estimate is close enough to the bound relative to its standard error that the likelihood shape and the constraint need to be examined using a method suited to that situation.
Diagnostics, Alternatives, and Limitations
Before reporting a Wald interval, check whether the numerical and statistical conditions behind it are plausible. A matrix that is not positive definite at an alleged interior maximum may indicate an optimization problem, a saddle point, or a poorly identified direction. A very large condition number or near-unit parameter correlations indicate that small changes in the data or numerical derivatives may substantially affect the covariance estimate. Also compare each interval with parameter bounds and with the scale on which the parameter is modeled. A symmetric interval on a log parameter becomes asymmetric after exponentiation to the original scale; this can be appropriate when the log-scale approximation is credible, but a transformation does not repair every likelihood-shape problem.
The inverse information gives a local approximation, not a complete description of uncertainty. It may omit important asymmetry, multiple modes, boundary effects, model uncertainty, or practical limitations of the data. In addition, a narrow interval for an individual parameter does not automatically guarantee precise predictions for every model output, and a broad parameter interval does not necessarily imply equally broad uncertainty in every prediction.
| Approach | What it uses | Useful diagnostic or limitation |
|---|---|---|
| Wald interval | Estimate and local standard error from inverse information | Fast and simple, but symmetric and sensitive to local quadratic approximation |
| Interval on a transformed parameter scale | Wald calculation after a transformation such as a logarithm | Can respect a bound after transforming back, but depends on a suitable approximation on the transformed scale |
| Likelihood-based interval | How the likelihood changes away from its maximum | Can represent asymmetry better; requires evaluating the likelihood beyond its local curvature |
| Bootstrap interval | Variation in estimates across simulated or resampled data sets | Can reveal skewness, but depends on a suitable resampling or simulation scheme and stable refitting |
These approaches answer related but not identical practical needs. Wald intervals are convenient summaries when the fit is well behaved and the likelihood is locally close to quadratic. When diagnostics show strong nonlinearity, high correlation, or proximity to a bound, the interval should be treated as a warning or rough local summary rather than as a definitive uncertainty range. The alternative should match the model, the data-generating assumptions, and the scientific quantity being reported.
Common Mistakes
- Using the information matrix instead of its inverse for variances. The diagonal entries of the inverse information matrix estimate variances; their square roots are standard errors.
- Ignoring off-diagonal entries. A covariance matrix with large off-diagonal terms signals dependence among estimates and can reveal a weakly determined parameter combination.
- Reporting an impossible endpoint as a valid range. A negative rate or probability outside \([0,1]\) is evidence that the ordinary Wald approximation is unsuitable, not a value to interpret biologically.
- Clipping an interval to a bound without justification. Changing an endpoint changes the interval procedure; its nominal coverage is not automatically preserved.
- Calling the interval a probability statement about the parameter. The confidence level describes long-run coverage of the procedure under its assumptions.
Check Your Understanding
Answer each question, then compare your reasoning with the short solutions.
- If an information matrix is invertible, which entries of its inverse provide the estimated marginal variances?
- A Wald interval for a fraction is \([-0.03,0.18]\). What does the negative endpoint suggest, and why is clipping it to zero not automatically justified?
- What does a parameter-estimate correlation of \(-0.98\) indicate about the fitted estimates?
- State the repeated-sampling interpretation of a nominal 95% confidence procedure.
Solutions. (1) The diagonal entries of the inverse information matrix are the estimated marginal variances; take their square roots to obtain standard errors. (2) The interval extends outside the allowed parameter space, suggesting that the normal approximation may be poor or that the estimate is too close to a bound for an ordinary Wald interval. Clipping changes the procedure, so its coverage is not guaranteed to remain 95%. (3) The estimates tend to move in opposite directions across repeated samples or local perturbations, often reflecting a weakly determined compensating combination; the correlation alone does not identify its biological cause. (4) Under the method’s assumptions, intervals constructed by the procedure in repeated samples would contain the fixed parameter approximately 95% of the time.