Is there any consequence of using 'fixPar' when interpreting results?

[Dekike]

Hi Brett,

Thanks for creating the package momentuHMM and all the time you spend replying users' doubts.

I am assessing the effect of different environmental co-variables in the activity behaviour of a marine fish species using acoustic telemetry. I am running a HMM by individual using fitHMM() since I have only a few individuals and I think that the patterns can be different among them. So my data stream is the average activity of my fish at 15-min time bins (meanAcc). I am just defining two states (Active and Resting) after observing the histograms for meanAcc for all the individuals (two clear frequency peaks in meanAcc).

My doubt comes when I plot the histogram of the activity values and the estimated state-dependent probability distributions using plot() in the same way you do it in the provided examples. If I do not use fixPar(), I get estimated state-dependent probability distributions that do not seem proper to me (at least visually). I have tried different initial parameters but the state-dependent probability distributions end always in the same fit independently of the initial values. However, when I use fixPar I really get good state-dependent probability distributions.

This is the fit when I do not use fixPar (independently of the initial parameters):

This is the fit when I do use fixPar:

Thus, my question is: is it a problem to do not leave to the algorithm to estimate the parameters for the state-dependent probability distributions? I wonder if in this case, for example, it would be advisable to use fixPar or not.

Thanks for your time Brett.

[bmcclintock]

I'm not sure I follow you here, but it sounds like you want to use fixPar to fix the meanAcc probability distribution parameters to values that yield a desired result. This is not valid from a statistical standpoint. Assuming the freely estimated model is converging to the global maximum, the data are suggesting that the state-dependent distributions are perhaps not as distinct as you might have hoped and/or the lack of fit is attributable to some additional structure in the data that is not well explained by the model. Pseudo-residuals can be helpful for diagnosing lack of fit (see the plotPR function).

These are complex models, and there is a lot that needs to be understood in order to draw reliable inferences from them. I'd suggest carefully going through the moveHMM vignettes, momentuHMM vignette, this recent paper and supplemental tutorial, and other more general resources in the statistical literature (e.g. Zucchini et al. 2016) before proceeding further. If you don't have enough free time to invest in gaining a better understanding of these fundamentals, it might be worth considering collaborating with a statistician.

[Dekike]

Hi Brett,

I completely understood. I won't use fixPart. I will try reducing data since there is substantial autocorrelation for some individuals and a colleague said to me that one option to reduce autocorrelation is to remove part of the data (e.g, 1 out of 4 time-bins). I wonder if that could help to a better fit now that you talked about Pseudo-residuals.

Thanks for your time.