1: \begin{abstract}

2: This article considers estimation of constant and time-varying

3: coefficients in nonlinear ordinary differential equation (ODE)

4: models where analytic closed-form solutions are not available. The

5: numerical solution-based nonlinear least squares (NLS) estimator is

6: investigated in this study. A numerical algorithm such as the Runge--Kutta

7: method is used to approximate the ODE solution. The asymptotic

8: properties are established for the proposed estimators

9: considering both numerical error and measurement error. The

10: B-spline is used to approximate the time-varying

11: coefficients, and the corresponding asymptotic theories in this case

12: are investigated under the framework of the sieve approach. Our

13: results show that if the maximum\vspace*{1pt} step size of the $p$-order

14: numerical algorithm goes to zero at a rate faster than

15: $n^{-1/(p\wedge4)}$, the numerical error is negligible compared to

16: the measurement error. This result provides a theoretical guidance in

17: selection of the step size for numerical evaluations of ODEs.

18: Moreover, we have shown that the numerical solution-based NLS

19: estimator and the sieve NLS estimator are strongly consistent. The

20: sieve estimator of constant parameters is asymptotically normal with

21: the same asymptotic co-variance as that of the case where the true

22: ODE solution is exactly known, while the estimator of the

23: time-varying parameter has the optimal convergence rate under some

24: regularity conditions. The theoretical results are also developed

25: for the case when the step size of the ODE numerical solver does not

26: go to zero fast enough or the numerical error is comparable to the

27: measurement error. We illustrate our approach with both simulation

28: studies and clinical data on HIV viral dynamics.

29: \end{abstract}