For the implementation of the unconstrained quadratic spline smoother
\(\tilde\varphi_n\) (see quad_spline_est
), based on the knot mesh
\(\{t_j = x_{[j n/k_n]}: j=1,\ldots,k_n-1\}\),
the user has to employ the option method="u"
.
Since the number \(k_n\) determines the complexity of the spline approximation,
its choice may be viewed as model selection via the minimization of the following Akaike (option type="AIC"
)
or Bayesian (option type="BIC"
) information criteria:
$$
A\tilde{I}C(k) = \log \left( \sum_{i=1}^{n} (\tilde \varphi_n(x_i)- y_i) \right) + (k+2)/n,$$
$$B\tilde{I}C(k) = \log \left( \sum_{i=1}^{n} (\tilde \varphi_n(x_i) - y_i) \right) + \log n \cdot (k+2)/2n.$$
For the implementation of the monotone (option method="m"
) quadratic spline smoother \(\hat\varphi_n\) (see quad_spline_est
),
the authors first suggest using the set of knots \(\{ t_j = {\mathcal{X}_{[j \mathcal{N}/k_n]}},~j=1,\ldots,k_n-1 \}\)
among the FDH points \((\mathcal{X}_{\ell},\mathcal{Y}_{\ell})\), \(\ell=1,\ldots,\mathcal{N}\)
(function quad_spline_est
).
Then, they propose to choose \(k_n\) by minimizing the following AIC (option type="AIC"
) or BIC (option type="BIC"
) information criteria:
$$
A\hat{I}C(k) = \log \left( \sum_{i=1}^{n} (\hat \varphi_n(x_i)- y_i) \right) + (k+2)/n,$$
$$B\hat{I}C(k) = \log \left( \sum_{i=1}^{n} (\hat \varphi_n(x_i) - y_i) \right) + \log n \cdot (k+2)/2n.$$
A small number of knots is typically needed as elucidated by the asymptotic theory.
For the implementation of the monotone and concave (option method="mc"
) spline estimator \(\hat\varphi^{\star}_n\),
just apply the same scheme as above by replacing the FDH points \((\mathcal{X}_{\ell},\mathcal{Y}_{\ell})\)
with the DEA points \((\mathcal{X}^*_{\ell},\mathcal{Y}^*_{\ell})\) (see dea_est
).