Open DeirdreLoughnan opened 10 months ago
@DeirdreLoughnan Ah, how lovely! This looks good to me.
@lizzieinvancouver the full joint model run fairly well, but produces a few kinda poor estimates, in particular the estimates for sigma_sp is off by 1.1, mu_phenosp is off by 4.2, mu_photosp off by 1, sigma_forcesp off by 0.6, but all within the 95 UI:
Parameter Test.data.values Estiamte X2.5 X97.5
1 mu_grand 10.0 10.4441841 7.6460133 13.3216813
2 sigma_sp 5.0 6.1673466 4.4243863 8.7681425
3 pop2 2.0 1.9311370 1.6461068 2.2102440
4 pop3 3.0 2.9893503 2.7161513 3.2691316
5 pop4 4.0 3.9709713 3.6889872 4.2405484
6 sigma_traity 1.0 1.0210205 0.9500146 1.0989096
7 mu_forcesp -10.0 -10.1009704 -11.6134198 -8.5305682
8 mu_chillsp -14.0 -14.3923631 -15.2717498 -13.5422928
9 mu_photosp -15.0 -14.0063419 -15.3118170 -12.6064038
10 mu_phenosp 80.0 75.8092541 64.3400733 87.7067775
11 sigma_forcesp 1.0 1.6469020 1.1691591 2.3764052
12 sigma_chillsp 1.0 0.9263713 0.6510505 1.3470959
13 sigma_photosp 1.0 0.7590249 0.1042704 1.6417310
14 sigma_phenosp 30.0 28.0666737 20.6209860 39.5719637
15 sigma_phenoy 3.0 3.0088378 2.9271198 3.0912800
16 beta_tf 0.3 0.3024894 0.1696252 0.4390084
17 beta_tc -0.4 -0.3732745 -0.4433245 -0.2978961
18 beta_tp -0.2 -0.2619117 -0.3773023 -0.1475535
The species level estimates from the trait part of the model are slightly worse than the phenology model estiamtes:
@DeirdreLoughnan This looks promising to me! I don't think you can compare trait and sp so well because you have different values for them (by 10X). That's not a complaint -- just a reminder that when you look at the above plots you have to take the scale into consideration (and other differences you have in sample size would matter too). The only value that looks bad to me is mu_photosp... it's not great.
I suggest you up the reps and species number and check that estimates IMPROVE (post the above again). As long as that is happening and Faith doesn't catch anything, then I think it looks good!
@lizzieinvancouver I agree, I just meant it as a sanity check that the species level model estimates were not too far off from the simulated values, not to compare the two parts of the model.
Increasing the number of species and replicates by a third does improve the model estimates, now mu_photosp is only off by 0.24.
Parameter Test.data.values Estimate X2.5 X97.5
1 mu_grand 10.0 10.1605632 8.6710522 11.90982879
2 sigma_sp 5.0 4.6938167 3.5964846 6.24729560
3 pop2 2.0 2.0663416 1.9102283 2.22842689
4 pop3 3.0 3.0535842 2.8991202 3.21057612
5 pop4 4.0 4.1630543 4.0033034 4.32502282
6 sigma_traity 1.0 1.0022984 0.9644023 1.04198805
7 mu_forcesp -10.0 -10.0310663 -11.0709772 -8.93302257
8 mu_chillsp -14.0 -14.5954942 -15.7084225 -13.45009479
9 mu_photosp -15.0 -15.2438077 -16.3171013 -14.27234427
10 mu_phenosp 80.0 87.2091821 75.7682370 98.99164329
11 sigma_forcesp 1.0 1.1657436 0.8815548 1.55382457
12 sigma_chillsp 1.0 1.0899414 0.7762494 1.49716057
13 sigma_photosp 1.0 1.0779306 0.8070464 1.44256121
14 sigma_phenosp 30.0 33.8818060 25.9555789 44.43130266
15 sigma_phenoy 3.0 3.0236932 2.9646841 3.08394045
16 beta_tf 0.3 0.2991830 0.2004204 0.39773297
17 beta_tc -0.4 -0.3412474 -0.4467813 -0.24280313
18 beta_tp -0.2 -0.1673943 -0.2543025 -0.07252499
Doubling the number of species and replicates also improves the estimates of the beta_trait cues but the mu_photosp is off by a bit more:
Parameter Test.data.values Estimate X2.5 X97.5
1 mu_grand 10.0 8.9267986 7.2861606 10.5361421
2 sigma_sp 5.0 5.1878103 4.1745135 6.5623880
3 pop2 2.0 2.1629220 2.0655901 2.2614118
4 pop3 3.0 3.0575121 2.9593887 3.1592143
5 pop4 4.0 4.0194503 3.9210913 4.1178029
6 sigma_traity 1.0 0.9992280 0.9754701 1.0243269
7 mu_forcesp -10.0 -9.8318857 -10.7635660 -8.8773510
8 mu_chillsp -14.0 -14.6297820 -15.3424317 -13.9343588
9 mu_photosp -15.0 -14.6205509 -15.1389477 -14.0788621
10 mu_phenosp 80.0 80.0859711 71.8293629 88.5917402
11 sigma_forcesp 1.0 1.1481956 0.7420146 1.6217848
12 sigma_chillsp 1.0 1.0471821 0.8261831 1.3415233
13 sigma_photosp 1.0 0.8500683 0.6752768 1.0811460
14 sigma_phenosp 30.0 26.5404128 21.2623168 33.2160940
15 sigma_phenoy 3.0 2.9947200 2.9581781 3.0304403
16 beta_tf 0.3 0.2950475 0.1989360 0.3891516
17 beta_tc -0.4 -0.3602279 -0.4285898 -0.2941492
18 beta_tp -0.2 -0.2375688 -0.2886307 -0.1876610
@DeirdreLoughnan That could just be natural variation in MCMC output -- on quick glance this looks good to me!
In light of recent issues with this trait model and the traitors model, I am redoing the test data for this project.
There are some key differences between the two modeling approaches that arise from me using my experiment data:
To start, I ran the trait only part of the model alone to see if we are finding the same issues as with the traitors trait only model. Overall I think this model is running much better than the traitors model
The species level intercepts for a sample size close to the real data looks fairly good to me (Nrep = 8, Npop = 4, Nspp = 50):![Screen Shot 2023-08-23 at 1 25 21 PM](https://github.com/DeirdreLoughnan/Treetraits/assets/16216421/8430c727-f923-43db-b831-42d5e7ffd2fb)
Next I will combine it with the phenology model used in my growth chamber study, but without the phylogeny.