Issue with rbms::flight_curve()

[SarahVray]

Hi @RetoSchmucki,
For some species the fitted counts computed by the function rbms::flight_curve() are not at all in line with the "real" counts.
With the code:

species.names <- unique(LUBMS_count[order(SPECIES), SPECIES])
for(i in 1:length(species.names)){ 
ts_season_count <- rbms::ts_monit_count_site(ts_season_visit, LUBMS_count, sp = species.names[i])
ts_flight_curve <- rbms::flight_curve(ts_season_count,
                                      NbrSample = 500,  
                                      MinVisit = 3,    
                                      MinOccur = 2,   
                                      MinNbrSite = 5,  
                                      MaxTrial = 4,    
                                      GamFamily = "nb", 
                                      SpeedGam = FALSE, 
                                      CompltSeason = TRUE, 
                                      SelectYear = NULL,   
                                      TimeUnit = "w",     
                                      MultiVisit = 'mean', 
                                      weekday = 3,        
                                      KeepModel = TRUE)}

Here is what I get for P. c-album:
Polygonia c album_ts_season_count.csv
As you can see, count = NA in week 16 for 2017.
Polygonia c album_pheno.csv
Here in the fitted counts, I have a sum of 6814 counts in week 16 for 2017.
And it gives me this curve:

I don't understand how I can get this huge peak in week 16 without having any real count? Maybe I am doing something wrong... Do you know where this issue could come from?
Thank you very much in advance for your help!
Sarah

[RetoSchmucki]

Hi @SarahVray
The abnormally high peak is related to missing data on that week. You might have some data before or after, and the GAM try to fit a function that can go up and down quickly and, therefore, the steep increase and decrease. I suspect something around week 13, 14 or 15, probably higher than in weeks 26 - 28.

[SarahVray]

Hi @RetoSchmucki,
Thank you for your reply. This is what I thought too but the counts are not so high around week 16 compared to weeks 25-28, only a few counts in weeks 14 and in 19-20, if we look in the pheno file for 2017:

<style> </style>

Row Labels	Sum of COUNT	Sum of FITTED
12	0	0
13	0	0
14	14	315.5967142
15	0	3779.034072
16	0	6814.130681
17	0	2993.912999
18	0	534.6665671
19	7	64.73782182
20	11	8.865034033
21	0	2.290161285
22	0	1.848683409
23	0	5.054775022
24	8	26.30491722
25	163	138.3477652
26	0	399.9537871
27	101	531.2740158
28	98	393.1229698
29	36	198.8809329
30	0	84.42085017
31	47	36.66912955
32	0	19.61616074
33	0	15.52873704
34	27	21.30448689
35	14	44.02468948
36	0	98.24799282
37	0	169.228053
38	0	160.7908898
39	25	60.23816853
40	0	0

I don't understand why the peak is so huge in week 16 compared to weeks 25-28.

[SarahVray]

I don't know actually which GAM family would be the best for this kind of count data so I tested the Poisson distribution instead of the negative-binomial but this time it is week 37 in 2014 which acts weird (but not in the same proportion) so it is maybe not the solution:

Although it does resolve the abnormally low collated index in 2017:

[SarahVray]

Hi @RetoSchmucki, I resolved this issue by changing the weighting I applied to the counts (standardization to 1km and not 10km), now the flight curve looks OK so the GAM probably didn't like the too high numbers. Thanks!