Why Use Wald Statistics in a Full Covariate Model?
A full covariate model puts the candidate effects of interest into one model at the same time. Its fitted coefficients describe each candidate's association with the modeled pharmacokinetic or pharmacodynamic parameter while the other included covariates are held in the model. This makes it possible to screen several effects from one fit, without repeatedly refitting a reduced model for every candidate.
A Wald statistic uses a coefficient estimate and its estimated uncertainty from that full fit. It is computationally convenient, but it is not the same calculation as a likelihood-ratio test comparing the full model with a reduced model. The two procedures use different information and can give different screening conclusions, particularly when the likelihood is not close to quadratic or estimates are imprecise.
Model and Wald Inference
Let \(y_{ij}\) denote the observation for individual \(i\) at observation occasion \(j\). A general NLME model relates these observations to individual parameters \(\theta_i\), which depend on fixed effects, covariates, and random effects. For illustration, a covariate model for one positive parameter may be written
where \(x_i\) and \(z_i\) are covariates, \(\beta_0\) is an intercept, \(\beta_1\) and \(\beta_2\) are covariate coefficients, and \(\eta_i\) is an individual random effect. The exponential link ensures \(\theta_i>0\). The error and random-effects distributions, covariance structure, and observation model must also be specified as part of the NLME model; the Wald calculation does not make those assumptions valid.
Suppose the full model is fitted by maximum likelihood or an appropriate likelihood-based method. For one coefficient \(\beta_k\), let \(\hat{\beta}_k\) be its estimate and \(SE(\hat{\beta}_k)\) its standard error, typically derived from the estimated covariance matrix of the fixed-effect estimates. Under the null hypothesis \(H_0:\beta_k=0\), the Wald statistic is
With adequate information and regularity conditions, the estimate is approximately normally distributed around the true coefficient, and \(Z_W\) is approximately standard normal under \(H_0\). The two-sided approximate p-value is \(2P(Z\geq |Z_W|)\) for \(Z\sim N(0,1)\). A \(100(1-\alpha)\%\) Wald confidence interval is
where \(z_{1-\alpha/2}\) is the corresponding standard-normal quantile. At a 5% two-sided level, this is approximately \(\hat{\beta}_k\pm1.96SE(\hat{\beta}_k)\). If zero is outside this interval, the corresponding two-sided Wald test rejects \(H_0\) at that level; if zero is inside, it does not. This equivalence holds because the test and interval use the same normal approximation and standard error.
The coefficient's interpretation depends on the covariate scale and parameter link. In the log-linear model above, a one-unit increase in \(x_i\), at fixed \(z_i\), multiplies the typical parameter value by \(\exp(\beta_1)\), with the random-effect contribution otherwise unchanged. Thus \(\beta_1\) is a log-ratio effect, not a change in the original parameter measured in its units.
Worked Wald Calculations
Worked Example: Screening a Continuous Weight Effect
Hypothetical model. Suppose a full NLME model describes apparent clearance \(CL_i\) using weight \(W_i\) in kilograms and a second covariate \(z_i\): \(\log(CL_i)=\beta_0+\beta_W(W_i-70)+\beta_z z_i+\eta_i\). The full fit gives \(\hat{\beta}_W=-0.0180\) per kg and \(SE(\hat{\beta}_W)=0.00700\) per kg. Assume the model converged, the estimate is interior to the parameter space, and its estimated covariance is usable for normal-approximation inference.
The conditional test is \(H_0:\beta_W=0\) against \(H_a:\beta_W\neq0\), with the other covariate retained. The statistic is
The two-sided normal-approximation p-value is approximately \(2P(Z\geq2.571)=0.0101\). At \(\alpha=0.05\), this is evidence against a zero weight coefficient, conditional on the other terms and model specification.
The 95% interval on the log-clearance scale is
The interval excludes zero. For a 10 kg increase, the estimated clearance ratio is \(\exp(10\hat{\beta}_W)=\exp(-0.180)=0.8353\). Transforming the endpoints after multiplying by 10 gives \([\exp(-0.3172),\exp(-0.04280)]\approx[0.7284,0.9581]\). Under this model, the estimated typical clearance is about 16.5% lower per 10 kg, with a 95% interval for the ratio from about 0.728 to 0.958, conditional on the other included covariate. This interpretation is model-based, not a causal claim.
Worked Example: A Wald Interval That Includes Zero
Hypothetical model. In another full model, a centered laboratory covariate \(L_i\) enters the log-volume parameter as \(\log(V_i)=\beta_0+\beta_L L_i+\eta_i\). Let \(L_i\) be measured in units of 10 laboratory-value units. Suppose \(\hat{\beta}_L=0.0600\) per unit of \(L_i\) and \(SE(\hat{\beta}_L)=0.0400\). Assume adequate convergence and a finite, stable covariance estimate.
The Wald statistic is \(Z_W=0.0600/0.0400=1.500\), giving a two-sided approximate p-value of \(2P(Z\geq1.500)=0.1336\). The 95% interval is
Because zero lies inside this interval, the Wald test does not reject a zero coefficient at the 5% level. The corresponding estimated volume ratio per 10-unit increase is \(\exp(0.0600)=1.062\), while the transformed interval is \([\exp(-0.0184),\exp(0.1384)]\approx[0.9818,1.148]\). The data and model are compatible with a small decrease or an increase in the typical volume ratio. Failure to reject does not establish that the covariate has no effect.
Why Wald and Likelihood-Ratio Results Can Differ
A likelihood-ratio (LR) test compares the maximized likelihoods of two nested models: the full model and a reduced model in which the tested coefficient or coefficients are constrained to their null values. For one coefficient, its statistic is
Under \(H_0\), and under regularity conditions, \(\Lambda\) is approximately chi-square distributed with degrees of freedom equal to the number of independent restrictions. Unlike the Wald test, the LR test requires fitting the reduced model. It measures the change in maximized likelihood after imposing the restriction, rather than measuring the distance of the full-model estimate from zero in units of its estimated standard error.
Worked Example: Comparing a Wald Test With a Refitted LR Test
Hypothetical fit results. A full model has \(\hat{\beta}=0.0840\) and \(SE(\hat{\beta})=0.0400\) for one covariate coefficient. Its maximized log-likelihood is \(-812.40\). The otherwise identical reduced model, with \(\beta=0\), is refitted to the same data and has log-likelihood \(-814.05\). Assume both fits converge to appropriate maxima, use the same observations and likelihood specification, and differ only by this one restriction.
The Wald statistic is \(Z_W=0.0840/0.0400=2.100\), with a two-sided approximate p-value of \(0.0357\). Its 95% interval is \(0.0840\pm1.96(0.0400)=[0.00560,0.1624]\), which excludes zero.
The LR statistic is
With one restriction, the reference distribution is approximately chi-square with \(df=1\); the corresponding p-value is approximately \(0.0693\). The LR test therefore does not reject at the 5% level, unlike the Wald test. Neither arithmetic result is contradictory: the two tests use different approximations to the evidence near the null. In a sufficiently regular, large-sample setting they often agree closely, but curvature, weak information, or nuisance-parameter changes can make them differ. The discrepancy is a reason to inspect model behavior and uncertainty, not a reason to report only the more favorable result.
Assumptions, Diagnostics, and Screening Practice
Wald screening is most defensible when the full model is well specified for the intended analysis, the estimation has converged reliably, the coefficient is identifiable, and the sampling distribution of the estimate is reasonably approximated by a normal distribution. The covariance matrix must be finite and adequately estimated. The relevant asymptotic approximations can be poor with sparse covariate support, strong collinearity, weakly informed parameters, or estimates near parameter boundaries. Earlier material on covariate support and collinearity is important when judging these conditions.
Confirm convergence, plausible parameter estimates, a usable covariance matrix, and no obvious estimation warnings before interpreting standard errors or Wald statistics.
State the coefficient, its covariate scale, the null value, and which other model terms remain in the full model. Check that the coefficient corresponds to the scientific contrast intended.
Give the estimate, standard error, Wald statistic, approximate p-value, and confidence interval. Interpret effects on the parameter scale used by the model, transforming both interval limits if needed.
Consider the magnitude and precision of the effect as well as the threshold. If comparing with an LR test, refit a valid nested reduced model on the same data and likelihood basis.
A small Wald p-value is not a measure of clinical importance, and a large p-value is not proof of no meaningful effect. If many candidate coefficients are tested, ordinary per-coefficient p-values do not automatically account for the multiplicity or the model-selection process. Selection can also make subsequent estimates and intervals appear more decisive than they are. These tests support screening within the specified model; they do not by themselves establish a final model or causal interpretation.
Wald tests are attractive because the full model supplies all coefficient estimates and standard errors in one fit. Their limitations are the reliance on local, approximately quadratic likelihood behavior and the possibility of unstable uncertainty estimates. LR tests are often preferred for nested comparisons because they use refitted likelihoods, but they require successful comparable fits and also rely on regularity conditions for their reference distribution. Neither procedure replaces checking the covariate data, model structure, fit, or scientific plausibility.
Common Pitfall
Correction: The tests use different calculations and approximations, so disagreement is possible. Verify that the estimates and fits are reliable, check that the LR models are genuinely nested and comparable, and interpret the effect estimate and interval alongside both results. A screening threshold is a decision rule, not proof that one test is correct or that an effect is important.
Check Your Understanding
Answer each question using the definitions and calculations above. Brief solutions follow.
- A coefficient estimate is \(0.12\) with standard error \(0.06\). Calculate the Wald statistic and state the approximate two-sided p-value interpretation.
- For that coefficient, calculate a 95% Wald confidence interval. Does it include zero?
- Why is a conditional Wald test not a test conditional on known individual random effects?
- A full model has log-likelihood \(-350.0\) and a one-coefficient reduced model has log-likelihood \(-351.2\). Calculate the LR statistic and identify its approximate reference degrees of freedom.
- Give two reasons the Wald and LR conclusions might differ in an NLME covariate screen.
Solutions. (1) \(Z_W=0.12/0.06=2.00\); the two-sided normal-approximation p-value is about 0.0455. If the model and approximation are adequate, this is evidence against a zero coefficient at the 5% level, conditional on other included terms. (2) \(0.12\pm1.96(0.06)=0.12\pm0.1176=[0.00240,0.2376]\), which narrowly excludes zero. (3) “Conditional” refers to testing the coefficient with the other specified covariates retained in the model. Random effects are still modeled as random and accounted for through the NLME likelihood and estimation. (4) \(\Lambda=2\{-350.0-(-351.2)\}=2.40\); one restriction gives an approximate chi-square reference distribution with \(df=1\). (5) For example, the likelihood may be poorly approximated by a quadratic near its maximum, or the coefficient and its standard error may be unstable because of weak information or collinearity. Different behavior of nuisance-parameter estimates after refitting can also contribute.