From version 2.9.0, the model formulae provided to `secr.fit`

may
include smooth terms as specified for the mgcv function ‘gam’,
with some restrictions. Smooth terms may be used for both density and
detection parameters.

The specification of smooth terms is explained in
`formula.gam`

. Only a subset of options are relevant
to ‘secr’. Penalized splines are not available. The smooth function
may be ‘s’ or ‘te’.

The ‘wiggliness’ of the curve is controlled by the argument k, which in this implementation is set by the user. The argument ‘fx’ should be set to TRUE.

See also the example in secr-densitysurfaces.pdf.

Regression splines are a very flexible way to represent non-linear
responses in generalized additive models (e.g., mgcv, Wood
2006). Borchers and Kidney (in prep.) have shown how they may be used
to model 2-dimensional trend in density in secrgam, an **R**
package that extends secr. Their approach is to use mgcv
to construct regression spline basis functions from mask x- and
y-coordinates, and possibly additional mask covariates, and to pass
these as covariates to secr. The idea of using mgcv to
construct the basis functions is applied within secr from version 2.9.

Smooth semi-parametric responses are also useful for modelling variation in detection parameters such as g0 and sigma over time, or in response to individual or detector-level covariates, when (1) a linear or other parametric response is arbitrary or implausible, and (2) sampling spans a range of times or levels of the covariate(s).

For a concrete example, consider a population sampled monthly for a year (i.e., 12 ‘sessions’). If home range size varies seasonally then the parameter sigma may vary in a more-or-less sinusoidal fashion. A linear trend is obviously inadequate, and a quadratic is not much better. However, a sine curve is hard to fit (we would need to estimate its phase, amplitude, mean and spatial scale) and assumes the increase and decrease phases are equally steep. An extreme solution is to treat month as a factor and estimate a separate parameter for each level (month). A smooth (semi-parametric) curve may capture the main features of seasonal variation with fewer parameters.

This implementation of smooth models results in large fitted objects, on account of the need to store setup information from mgcv. It is also vulnerable to future changes in mgcv.

Expect that the implementation will change in later versions of
secr, and that smooth models fitted in the this version will not
necessarily be compatible with `predict`

and
`predictDsurface`

in later versions.

Setting the intercept of a smooth to zero is not a canned option in mgcv, and is not offered in secr. It may be achieved by placing a knot at zero and hacking the matrix of basis functions to drop the corresponding column, plus some more jiggling.

Borchers, D. L. and Kidney, D. J. (2014) Flexible density surface estimation using regression splines with spatially explicit capture-recapture data. Technical Report, Univ. St Andrews.

Wood, S. N. (2006) *Generalized Additive Models: An Introduction with
R.* Chapman and Hall/CRC.

`formula.gam`

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 | ```
## Not run:
## smooth density surface
possum.model.sxy <- secr.fit(possumCH, mask = possummask,
model = D ~ s(x,y, k = 6, fx = TRUE)), trace = FALSE)
fittedsurface <- predictDsurface(possum.model.sxy)
par(mar = c(1,1,1,6))
plot(fittedsurface)
plot(fittedsurface, plottype = 'contour', add = TRUE)
par(mar = c(5,4,4,2) + 0.1) ## reset to default
## Now try smooth on g0
## For the smooth we use 'Session' which is coded numerically (0:4)
## rather than the factor 'session' ('2005', '2006', '2007', '2008',
## '2009')
ovenbird.model.g0 <- secr.fit(ovenCH, mask = ovenmask,
model = g0 ~ session, trace = FALSE)
ovenbird.model.sg0 <- secr.fit(ovenCH, mask = ovenmask,
model = g0 ~ s(Session, k = 3, fx = TRUE), trace = FALSE)
AIC(ovenbird.model.g0, ovenbird.model.sg0)
## Or over occasions within a session...
fit.sT3 <- secr.fit(captdata, model = g0 ~ s(T, k = 3, fx = TRUE),
trace = FALSE)
pred <- predict(fit.sT3, newdata = data.frame(T = 0:4))
plot(sapply(pred, '[', 'g0', 'estimate'))
## End(Not run)
``` |

Questions? Problems? Suggestions? Tweet to @rdrrHQ or email at ian@mutexlabs.com.

All documentation is copyright its authors; we didn't write any of that.