[R-meta] Unimodal relationship?
lists at dewey.myzen.co.uk
Sat Mar 10 12:06:39 CET 2018
You have a number of options:
1 - if you genuinely believe it is a curved relationship then you could
include a quadratic term in the model as well as the linear one. You
either include I(CNr^2) in the formula or use poly(). The advantage of
this is that it is simple and people think they understand the
coefficient. The disadvantage is that polynomial fits are influenced by
the glbal data rather than by locally and they go off to plus and minus
infinity at the extremes. It is important to include the linear term as
well and test both linear and quadratic together.
2 - if you believe it is really two linear segments with a break-point
you can fit it as such. The advantage of this is that it may be more
credible in some applications, the disadvantage is that if you also try
to estimate the break-point it becomes more complicated and also more
data-driven. If you want to do this searching for "broken stick" should
On 09/03/2018 16:01, Cesar Terrer Moreno wrote:
> Dear Michael,
> Indeed, I mean a U-shaped curve. I want to:
> i) run model selection in the context of meta-analysis, but including
> this unimodal response, e.g.:
> glmulti(es ~ */CNr/* + Biome + ph + totN. + MAT + MAP + deltaco2,
> dat=ecmdat,level=2, fitfunction=rma.glmulti,
> crit="aicc", confsetsize=2^7)
> ii) and if it comes up as an important predictor, including this
> unimodal predictor in a meta-regression.
> rma(es, var, data=ecmdat , mods= ~ MAT + MAP + deltaco2 + */CNr/*)
> So my question is, how can I treat */CNr/* so the model understand CNr
> does not follow a linear relationship?
>> On 9 Mar 2018, at 16:38, Michael Dewey <lists at dewey.myzen.co.uk
>> <mailto:lists at dewey.myzen.co.uk>> wrote:
>> Dear Cesar
>> When you say a unimodal relationship do you mean a U-shaped one? There
>> are various options depending on exactly what you want to do.
>> On 09/03/2018 12:14, Cesar Terrer Moreno wrote:
>>> Dear all,
>>> I initially run a model selection analysis, and found 3 potentially
>>> important predictors: MAT, MAP and deltaCO2, and potentially also the
>>> interaction MAT:deltaco2:
>>> However, I suspect that the next predictor in importance, CNr, might
>>> be more important than here shown. The problem is that it may follow
>>> a nonlinear behaviour. In particular, based on the plot, it seems it
>>> may follow an unimodal relationship:
>>> <https://www.dropbox.com/s/ahpq07y8929muti/Rplot13.jpeg?dl=0> or at
>>> least interact with other predictors more clearly.
>>> How would you include a potentially unimodal shape in a model in
>>> This is the best model:
>>> rma(es, var, data=dat , mods= ~ 1 + MAT + MAP + deltaco2 +
>>> deltaco2:MAT, knha=TRUE))
>>> Please help me figure out how to include CNr in this model as a
>>> nonlinear relationship.
>>> [[alternative HTML version deleted]]
>>> R-sig-meta-analysis mailing list
>>> R-sig-meta-analysis at r-project.org
>>> <mailto:R-sig-meta-analysis at r-project.org>
More information about the R-sig-meta-analysis