Skip to contents

bboutools supports several extensions to the standard single-population monthly survival workflow. These include multi-population analysis, aggregate annual survival data, predictions for unobserved years, prior-only sampling and other features described below.

Multi-Population Analysis

bboutools can fit models to data from multiple populations simultaneously. The fitting functions auto-detect multiple populations from the PopulationName column.

Data format

Multi-population data has the same columns as single-population data, with different values in PopulationName.

surv_multi <- bboudata::bbousurv_multi[!is.na(bboudata::bbousurv_multi$Month), ]
head(surv_multi)
#> # A tibble: 6 × 6
#>   PopulationName  Year Month StartTotal MortalitiesCertain MortalitiesUncertain
#>   <chr>          <int> <int>      <int>              <int>                <int>
#> 1 A               2001     1         17                  0                    0
#> 2 A               2001     2         17                  0                    0
#> 3 A               2001     3         17                  0                    0
#> 4 A               2001     4         24                  0                    0
#> 5 A               2001     5         28                  0                    0
#> 6 A               2001     6         28                  0                    0

Populations can have different year ranges.

surv_multi |>
  group_by(PopulationName) |>
  summarise(min_year = min(Year), max_year = max(Year))
#> # A tibble: 3 × 3
#>   PopulationName min_year max_year
#>   <chr>             <int>    <int>
#> 1 A                  2001     2010
#> 2 B                  2005     2014
#> 3 C                  2003     2013

Fitting

No special argument is needed. The function bb_fit_survival() detects multiple populations and fits a single model with population-indexed parameters.

set.seed(101)
survival_multi <- bb_fit_survival(surv_multi, quiet = TRUE)

The tidy() output shows population-indexed parameters. The intercept b0, year trend bYear, annual effects bAnnual and month effects bMonth are estimated for each population. The standard deviations sAnnual and sMonth are shared across populations.

tidy(survival_multi, include_random_effects = FALSE)
#> # A tibble: 8 × 4
#>   term     estimate lower upper
#>   <term>      <dbl> <dbl> <dbl>
#> 1 b0[1]       4.51  3.7    5.38
#> 2 b0[2]       5.01  4.2    6.19
#> 3 b0[3]       5.03  4.12   6   
#> 4 bYear[1]    0     0      0   
#> 5 bYear[2]    0     0      0   
#> 6 bYear[3]    0     0      0   
#> 7 sAnnual     0.808 0.554  1.29
#> 8 sMonth      0.992 0.662  1.52

Predictions and plots

Prediction and plot functions work as with single-population models. Plots are automatically faceted by PopulationName.

pred_multi <- bb_predict_survival(survival_multi)
bb_plot_year_survival(pred_multi)

The default faceting can be overridden since all plot functions return ggplot2 objects.

bb_plot_year_survival(pred_multi) + facet_wrap(~PopulationName, ncol = 3)

The recruitment model supports the same multi-population structure.

rec_multi <- bboudata::bbourecruit_multi[
  !is.na(bboudata::bbourecruit_multi$Month),
]
set.seed(101)
recruitment_multi <- bb_fit_recruitment(rec_multi, quiet = TRUE)
pred_rec_multi <- bb_predict_recruitment(recruitment_multi)
bb_plot_year_recruitment(pred_rec_multi)

See the analytical methods article for details on the multi-population model specification.

Multi-population Maximum Likelihood models are not currently supported.

Aggregate Annual Survival Data

Survival data can be provided as one row per population per year, instead of monthly records. This is useful when monthly collar data is not available and only annual summaries exist.

Data format

Annual survival data uses the same columns as monthly data. Set Month to the year_start value (default 4L for April).

surv_annual <- bboudata::bbousurv_annual
surv_annual_c <- surv_annual[surv_annual$PopulationName == "C", ]
head(surv_annual_c)
#> # A tibble: 6 × 6
#>   PopulationName  Year Month StartTotal MortalitiesCertain MortalitiesUncertain
#>   <chr>          <int> <int>      <int>              <int>                <int>
#> 1 C               2003     4         10                  3                    0
#> 2 C               2004     4         15                  3                    0
#> 3 C               2005     4         20                  3                    0
#> 4 C               2006     4         23                  0                    0
#> 5 C               2007     4         26                 10                    0
#> 6 C               2010     4         29                  2                    0

Fitting

Annual data is auto-detected from the input. No special argument is needed.

set.seed(101)
survival_annual <- bb_fit_survival(surv_annual_c, quiet = TRUE)

When annual data is provided, the model has no month random effect (bMonth is fixed to zero) and the ^12 annualization used to convert monthly to annual survival is skipped.

Predictions

The function bb_predict_survival() works as normal for annual fits.

bb_predict_survival(survival_annual)
#> # A tibble: 9 × 6
#>   PopulationName CaribouYear Month estimate lower upper
#>   <fct>                <int> <int>    <dbl> <dbl> <dbl>
#> 1 C                     2003    NA    0.808 0.476 0.94 
#> 2 C                     2004    NA    0.835 0.659 0.946
#> 3 C                     2005    NA    0.868 0.713 0.958
#> 4 C                     2006    NA    0.953 0.87  0.995
#> 5 C                     2007    NA    0.684 0.509 0.835
#> 6 C                     2010    NA    0.92  0.813 0.98 
#> 7 C                     2011    NA    0.934 0.83  0.987
#> 8 C                     2012    NA    0.942 0.838 0.988
#> 9 C                     2013    NA    0.836 0.656 0.929

Calling bb_predict_survival(survival_annual, month = TRUE) will error, since monthly predictions are not meaningful for annual fits. Predictions from annual and monthly models are both on the annual survival scale and are directly comparable.

Predicting Unobserved Years

Predictions can be generated for years with no observed data. This is done by adding placeholder rows to the input data and setting allow_missing = TRUE.

Data format

A placeholder row has PopulationName and Year filled in, with all measurement columns set to NA_integer_.

surv_missing <- bboudata::bbousurv_missing
head(surv_missing)
#> # A tibble: 6 × 6
#>   PopulationName  Year Month StartTotal MortalitiesCertain MortalitiesUncertain
#>   <chr>          <int> <int>      <int>              <int>                <int>
#> 1 C               2003     3          0                  0                    0
#> 2 C               2003     4         10                  2                    0
#> 3 C               2003     5          5                  0                    0
#> 4 C               2003     6          5                  1                    0
#> 5 C               2003     7          4                  0                    0
#> 6 C               2003     8          3                  0                    0
# Placeholder rows have NA in the Month column
surv_missing[is.na(surv_missing$Month), ]
#> # A tibble: 4 × 6
#>   PopulationName  Year Month StartTotal MortalitiesCertain MortalitiesUncertain
#>   <chr>          <int> <int>      <int>              <int>                <int>
#> 1 C               2010    NA         NA                 NA                   NA
#> 2 C               2011    NA         NA                 NA                   NA
#> 3 C               2012    NA         NA                 NA                   NA
#> 4 C               2013    NA         NA                 NA                   NA

Fitting

set.seed(101)
survival_missing <- bb_fit_survival(
  surv_missing,
  allow_missing = TRUE,
  quiet = TRUE
)

Predictions

For unobserved years, the year random effects (bAnnual) are sampled from their prior distribution (Normal(0, sAnnual)) rather than being informed by data. This means predictions reflect the population-level mean with additional uncertainty from the year effect prior, giving wider credible intervals than observed years where the data constrains the year effect.

pred_missing <- bb_predict_survival(survival_missing)
bb_plot_year_survival(pred_missing)

This approach works for years within or beyond the range of observed data. It requires random year effects (not fixed effects), since the model must include the hierarchical year structure. Unobserved year predictions are only supported for Bayesian models.

Prior-Only Sampling

Sampling from prior distributions without observed data enables prior predictive checks. This is useful for verifying that priors produce ecologically plausible predictions before fitting to data, particularly when using disturbance-informed national priors.

Workflow

Provide a data frame where PopulationName and Year are filled in but all measurement columns are NA. Set allow_missing = TRUE.

# Build placeholder data
recruit_prior <- data.frame(
  PopulationName = "A",
  Year = 2020:2030,
  Month = NA_integer_,
  Day = NA_integer_,
  Cows = NA_integer_,
  Bulls = NA_integer_,
  UnknownAdults = NA_integer_,
  Yearlings = NA_integer_,
  Calves = NA_integer_
)

# Low disturbance scenario
priors_low <- bb_priors_recruitment_national(anthro = 10, fire_excl_anthro = 5)

set.seed(1)
fit_low <- bb_fit_recruitment(
  data = recruit_prior,
  niters = 1000,
  allow_missing = TRUE,
  year_trend = TRUE,
  priors = priors_low,
  quiet = TRUE
)

pred_low <- bb_predict_recruitment(fit_low)
bb_plot_year_recruitment(pred_low)

With year_trend = FALSE (default), all years are exchangeable and marginal predictions are identical across years. The year_trend = TRUE option allows temporal patterns from the bYear prior to appear in predictions.

Prior-only sampling is only supported for Bayesian models. See the priors article for more information on national priors and data rescaling.

Growth Predictions with Mismatched Fits

The functions bb_predict_growth() and bb_predict_population_change() auto-filter to the intersection of populations and years when survival and recruitment fits cover different ranges. An informative message reports what was excluded.

growth <- bb_predict_growth(survival_multi, recruitment_multi)
#> Filtering to shared population and year combinations. CaribouYears in survival only: 2000, 2001.

If there are no shared population-year combinations, an empty tibble is returned.

Raw MCMC Samples

All bb_predict_*() functions have bb_predict_*_samples() counterparts that return raw MCMC samples instead of summarized estimates. These are useful for custom uncertainty analyses.

x <- bb_predict_survival_samples(survival_annual)
names(x)
#> [1] "samples" "data"
head(x$samples)
#> , , 1
#> 
#>           [,1]      [,2]      [,3]      [,4]      [,5]      [,6]      [,7]
#> [1,] 0.8100703 0.8328225 0.8990700 0.7512339 0.8211276 0.8807257 0.8743078
#> [2,] 0.9238838 0.8245992 0.7577845 0.8189265 0.7930301 0.5303166 0.8641353
#> [3,] 0.5880067 0.7304969 0.8157142 0.7416115 0.8141208 0.8081028 0.8096746
#>           [,8]      [,9]     [,10]     [,11]     [,12]     [,13]     [,14]
#> [1,] 0.7846362 0.8198930 0.8315293 0.5411449 0.9003593 0.8051385 0.5582257
#> [2,] 0.8532382 0.7692350 0.7014759 0.7047861 0.7555578 0.8272819 0.8933403
#> [3,] 0.7360473 0.7054802 0.6285257 0.8122481 0.4219754 0.8295529 0.9042276
#>          [,15]     [,16]     [,17]     [,18]     [,19]     [,20]     [,21]
#> [1,] 0.8668522 0.7635840 0.7534132 0.6752304 0.8134217 0.8952287 0.5693488
#> [2,] 0.7951685 0.7876833 0.8633726 0.6656884 0.7629868 0.8916369 0.9071309
#> [3,] 0.8774068 0.6856384 0.8419345 0.8145904 0.7597264 0.8516626 0.8139380
#>          [,22]     [,23]     [,24]     [,25]     [,26]     [,27]     [,28]
#> [1,] 0.7468098 0.9421215 0.8766044 0.8685347 0.9836947 0.8522096 0.8564739
#> [2,] 0.8468165 0.8969447 0.9185559 0.5508445 0.7904945 0.6405212 0.7738015
#> [3,] 0.7725931 0.8291788 0.9017913 0.7267522 0.6992850 0.8723192 0.6542679
#>          [,29]     [,30]     [,31]     [,32]     [,33]     [,34]     [,35]
#> [1,] 0.7383450 0.8271721 0.8875011 0.4262374 0.8141064 0.7730571 0.8730336
#> [2,] 0.8647817 0.5905949 0.8075227 0.6710467 0.5344299 0.5477466 0.7558449
#> [3,] 0.7091833 0.8378856 0.8569975 0.7276413 0.8126181 0.8898369 0.6980845
#>          [,36]     [,37]     [,38]     [,39]     [,40]     [,41]     [,42]
#> [1,] 0.6195000 0.7456662 0.9196607 0.5829161 0.7037264 0.6435583 0.8963820
#> [2,] 0.8663753 0.7643604 0.7744783 0.8153313 0.7726081 0.4196870 0.8058030
#> [3,] 0.7850245 0.8293675 0.7508207 0.6216705 0.8888044 0.6416695 0.9497972
#>          [,43]     [,44]     [,45]     [,46]     [,47]     [,48]     [,49]
#> [1,] 0.7841771 0.9263563 0.8170395 0.7463283 0.8619600 0.7372546 0.9107587
#> [2,] 0.7620853 0.8463511 0.8143873 0.8939234 0.8547188 0.6750149 0.8600775
#> [3,] 0.8586423 0.7714185 0.8008535 0.8710972 0.7598697 0.9590951 0.8618687
#>          [,50]     [,51]     [,52]     [,53]     [,54]     [,55]     [,56]
#> [1,] 0.8369435 0.5463371 0.8573501 0.6526534 0.8107696 0.7969534 0.6873653
#> [2,] 0.8723852 0.7900491 0.9331585 0.3366803 0.8379062 0.7484743 0.9425491
#> [3,] 0.9212059 0.5329239 0.8361467 0.6845912 0.7484464 0.8481787 0.9256950
#>          [,57]     [,58]     [,59]     [,60]     [,61]     [,62]     [,63]
#> [1,] 0.8586997 0.9099490 0.8280727 0.6255021 0.8498267 0.9380145 0.6376911
#> [2,] 0.9047329 0.8836313 0.9354116 0.6722869 0.7217318 0.8040429 0.7941548
#> [3,] 0.7610538 0.8678330 0.9436548 0.7334547 0.7736738 0.7865973 0.8702353
#>          [,64]     [,65]     [,66]     [,67]     [,68]     [,69]     [,70]
#> [1,] 0.8017054 0.3338572 0.7926214 0.6760746 0.8591673 0.8028355 0.5164189
#> [2,] 0.9225951 0.8954302 0.7325501 0.7628435 0.3820068 0.7503877 0.7765615
#> [3,] 0.8500770 0.8448014 0.7671730 0.8370517 0.8342128 0.5888692 0.8395882
#>          [,71]     [,72]     [,73]     [,74]     [,75]     [,76]     [,77]
#> [1,] 0.8919600 0.7349450 0.8396727 0.7284589 0.7823173 0.7870130 0.5672428
#> [2,] 0.6503320 0.8492333 0.6059178 0.9074754 0.8750516 0.9032515 0.3159498
#> [3,] 0.6280472 0.5117538 0.7130307 0.8417812 0.8406776 0.8805779 0.8120334
#>          [,78]     [,79]     [,80]     [,81]     [,82]     [,83]     [,84]
#> [1,] 0.7793890 0.8554130 0.7337111 0.7793924 0.6603114 0.8905571 0.8080208
#> [2,] 0.6396056 0.5940490 0.6281076 0.6278698 0.6628142 0.6022987 0.6752844
#> [3,] 0.8791846 0.8722574 0.8479166 0.7488208 0.9283555 0.8964176 0.8695403
#>          [,85]     [,86]     [,87]     [,88]     [,89]     [,90]     [,91]
#> [1,] 0.8270506 0.8549400 0.8277871 0.9456186 0.7619716 0.7874446 0.8042569
#> [2,] 0.8367249 0.8960159 0.8320300 0.8768826 0.8919049 0.7439260 0.7343759
#> [3,] 0.7119678 0.9062449 0.4431451 0.7892155 0.8451588 0.5397931 0.8709981
#>          [,92]     [,93]     [,94]     [,95]     [,96]     [,97]     [,98]
#> [1,] 0.8885190 0.9317869 0.8037340 0.6021400 0.6559141 0.8691621 0.8176755
#> [2,] 0.7682111 0.7800728 0.8126976 0.6990013 0.7020015 0.7480573 0.7909865
#> [3,] 0.8083182 0.7916754 0.6965616 0.8992922 0.5588646 0.8855327 0.9319988
#>          [,99]    [,100]
#> [1,] 0.9331796 0.8836836
#> [2,] 0.7395588 0.8707166
#> [3,] 0.8467371 0.9660237
#> 
#> , , 2
#> 
#>           [,1]      [,2]      [,3]      [,4]      [,5]      [,6]      [,7]
#> [1,] 0.8293867 0.7311051 0.9336984 0.8639506 0.8398274 0.7271077 0.8916298
#> [2,] 0.8873797 0.8681714 0.8369724 0.9074485 0.8392773 0.8997494 0.8266006
#> [3,] 0.6746982 0.7292853 0.7841578 0.8187604 0.9006604 0.7195872 0.7995201
#>           [,8]      [,9]     [,10]     [,11]     [,12]     [,13]     [,14]
#> [1,] 0.8780739 0.9059842 0.8520349 0.8787033 0.8177614 0.9179396 0.6573988
#> [2,] 0.7024160 0.8093073 0.7444704 0.8036637 0.8537575 0.7247583 0.7485156
#> [3,] 0.9030053 0.9504198 0.8388340 0.9017147 0.5775758 0.8226344 0.8807530
#>          [,15]     [,16]     [,17]     [,18]     [,19]     [,20]     [,21]
#> [1,] 0.8369408 0.8280352 0.8234930 0.8498158 0.8161856 0.8586348 0.7500366
#> [2,] 0.9305410 0.7270518 0.8425565 0.8683086 0.7226480 0.9081365 0.9340098
#> [3,] 0.8132049 0.6791910 0.8150912 0.8198546 0.9469997 0.7346267 0.9129579
#>          [,22]     [,23]     [,24]     [,25]     [,26]     [,27]     [,28]
#> [1,] 0.8855860 0.9425874 0.8476553 0.9548032 0.8107868 0.7916221 0.9491966
#> [2,] 0.9204855 0.8912619 0.9157908 0.7269239 0.8454117 0.9194586 0.8153364
#> [3,] 0.8555252 0.7055170 0.7695387 0.8932254 0.7748247 0.7690427 0.7173810
#>          [,29]     [,30]     [,31]     [,32]     [,33]     [,34]     [,35]
#> [1,] 0.8011630 0.7547088 0.9222943 0.8130967 0.7896388 0.8296530 0.8346123
#> [2,] 0.8100982 0.8591612 0.7211850 0.5980312 0.7941119 0.7787029 0.8634473
#> [3,] 0.8065386 0.9007199 0.8531560 0.7934420 0.7035387 0.7312174 0.8194325
#>          [,36]     [,37]     [,38]     [,39]     [,40]     [,41]     [,42]
#> [1,] 0.8460653 0.8172861 0.7647252 0.6831386 0.8778357 0.8528962 0.9296982
#> [2,] 0.8544311 0.8910292 0.9386935 0.8280755 0.9304470 0.8806999 0.9814696
#> [3,] 0.7954272 0.9149802 0.8333634 0.8084516 0.7736039 0.6703391 0.8534657
#>          [,43]     [,44]     [,45]     [,46]     [,47]     [,48]     [,49]
#> [1,] 0.7811261 0.8948840 0.8589858 0.8845901 0.7623304 0.9238319 0.8563279
#> [2,] 0.7525711 0.8410882 0.8704187 0.7901223 0.8445063 0.8115907 0.8803641
#> [3,] 0.8009595 0.7123385 0.8674800 0.8461334 0.9057415 0.8418792 0.8630210
#>          [,50]     [,51]     [,52]     [,53]     [,54]     [,55]     [,56]
#> [1,] 0.6774996 0.8507111 0.8375339 0.8852542 0.8939923 0.8588805 0.7689726
#> [2,] 0.8168257 0.8089686 0.8283668 0.8082810 0.8585731 0.8446993 0.7768397
#> [3,] 0.8748812 0.6155297 0.5301974 0.9572894 0.8225471 0.9194190 0.8544440
#>          [,57]     [,58]     [,59]     [,60]     [,61]     [,62]     [,63]
#> [1,] 0.8462371 0.7533872 0.8803979 0.8625208 0.9303752 0.7874693 0.8197173
#> [2,] 0.8906101 0.9044991 0.8083001 0.9108421 0.8200310 0.8859758 0.8022308
#> [3,] 0.7579061 0.9053719 0.7452506 0.8748790 0.8720810 0.8812737 0.8327380
#>          [,64]     [,65]     [,66]     [,67]     [,68]     [,69]     [,70]
#> [1,] 0.9533242 0.7683925 0.9169126 0.7574053 0.9224265 0.9330959 0.8090889
#> [2,] 0.8750712 0.7188973 0.8041502 0.6621094 0.6725660 0.7730831 0.9320745
#> [3,] 0.9174009 0.7602372 0.8157698 0.8170600 0.9108447 0.8063568 0.8496216
#>          [,71]     [,72]     [,73]     [,74]     [,75]     [,76]     [,77]
#> [1,] 0.8148795 0.7898443 0.8823886 0.8117362 0.8195257 0.8577727 0.9272104
#> [2,] 0.8076728 0.9029436 0.9609612 0.9326824 0.8510755 0.8659497 0.6617831
#> [3,] 0.7947315 0.8673826 0.7856433 0.8822647 0.8186052 0.9148140 0.8845194
#>          [,78]     [,79]     [,80]     [,81]     [,82]     [,83]     [,84]
#> [1,] 0.9304156 0.8542400 0.8887304 0.9083874 0.8427650 0.9394655 0.8794863
#> [2,] 0.7662508 0.8600049 0.9448754 0.7508172 0.8361367 0.6655703 0.6935361
#> [3,] 0.8525972 0.8116240 0.7776032 0.8091174 0.8676534 0.6505736 0.6970490
#>          [,85]     [,86]     [,87]     [,88]     [,89]     [,90]     [,91]
#> [1,] 0.9101032 0.6372443 0.8818551 0.8348509 0.7680798 0.7874888 0.7567610
#> [2,] 0.6789243 0.8887513 0.8472573 0.9083839 0.7610991 0.9404909 0.8704755
#> [3,] 0.8266824 0.9102422 0.9131952 0.9066883 0.7732072 0.7812563 0.6919923
#>          [,92]     [,93]     [,94]     [,95]     [,96]     [,97]     [,98]
#> [1,] 0.8950414 0.7728260 0.7932243 0.8319341 0.7280164 0.8531637 0.8521593
#> [2,] 0.7147997 0.7641965 0.7766775 0.7307407 0.6540238 0.8853885 0.7961907
#> [3,] 0.7006307 0.8681243 0.6867075 0.8204617 0.7570312 0.8580996 0.8934793
#>          [,99]    [,100]
#> [1,] 0.7804450 0.9200551
#> [2,] 0.7794402 0.7983124
#> [3,] 0.8410643 0.8252591
#> 
#> , , 3
#> 
#>           [,1]      [,2]      [,3]      [,4]      [,5]      [,6]      [,7]
#> [1,] 0.8951500 0.9107236 0.8041130 0.9535924 0.8282742 0.8502156 0.7566478
#> [2,] 0.8158645 0.6816880 0.8400046 0.8374145 0.7698017 0.8369123 0.8594950
#> [3,] 0.9081653 0.8264055 0.8811966 0.8177609 0.8986505 0.9483199 0.8151739
#>           [,8]      [,9]     [,10]     [,11]     [,12]     [,13]     [,14]
#> [1,] 0.8437907 0.8467502 0.8896888 0.9083972 0.8083448 0.8848419 0.8900466
#> [2,] 0.7658273 0.9217423 0.8453742 0.8551429 0.9099919 0.8003665 0.9596853
#> [3,] 0.9338432 0.8043243 0.8982171 0.8996577 0.8654528 0.8723985 0.9446332
#>          [,15]     [,16]     [,17]     [,18]     [,19]     [,20]     [,21]
#> [1,] 0.8997366 0.8346078 0.8830437 0.8909072 0.9112686 0.7400774 0.8921982
#> [2,] 0.8852777 0.8769669 0.8357654 0.7681840 0.8412224 0.9001914 0.8812081
#> [3,] 0.8737531 0.9190535 0.8698283 0.8814223 0.7759807 0.7344760 0.9165530
#>          [,22]     [,23]     [,24]     [,25]     [,26]     [,27]     [,28]
#> [1,] 0.8608584 0.9632909 0.8735285 0.8986076 0.9014225 0.8323764 0.9402781
#> [2,] 0.8836648 0.8685428 0.8689913 0.8407546 0.9559023 0.9316712 0.8940689
#> [3,] 0.8531180 0.8454558 0.8486527 0.8450087 0.7811114 0.9621437 0.9074657
#>          [,29]     [,30]     [,31]     [,32]     [,33]     [,34]     [,35]
#> [1,] 0.7852975 0.6465497 0.7409607 0.8846483 0.9479749 0.8249267 0.8655759
#> [2,] 0.7967953 0.8701877 0.8081394 0.7255014 0.8964806 0.8810424 0.9533713
#> [3,] 0.8587880 0.8298049 0.8478950 0.8160541 0.7796005 0.8161459 0.9221744
#>          [,36]     [,37]     [,38]     [,39]     [,40]     [,41]     [,42]
#> [1,] 0.8370325 0.8712001 0.8230425 0.7608727 0.8201963 0.8312390 0.8093533
#> [2,] 0.8752257 0.9054110 0.7347317 0.9395892 0.8830635 0.8448732 0.9197452
#> [3,] 0.9034119 0.8916923 0.8983554 0.7945879 0.7817754 0.8486407 0.8444349
#>          [,43]     [,44]     [,45]     [,46]     [,47]     [,48]     [,49]
#> [1,] 0.8339872 0.8142735 0.8117632 0.9089905 0.8985137 0.8720813 0.9367906
#> [2,] 0.9237725 0.8875320 0.8795596 0.8678278 0.8850304 0.9313207 0.9002167
#> [3,] 0.8460263 0.8877801 0.9202429 0.8860233 0.9482301 0.8694775 0.9015154
#>          [,50]     [,51]     [,52]     [,53]     [,54]     [,55]     [,56]
#> [1,] 0.9038232 0.7704136 0.9672230 0.8317099 0.8438566 0.8492747 0.9253166
#> [2,] 0.8257471 0.9310629 0.8223734 0.7387574 0.8768370 0.8679235 0.8725061
#> [3,] 0.8943857 0.8317384 0.9320637 0.8721511 0.7836798 0.9526074 0.8910265
#>          [,57]     [,58]     [,59]     [,60]     [,61]     [,62]     [,63]
#> [1,] 0.9418538 0.8527747 0.9572568 0.8342982 0.8832001 0.9116158 0.8555157
#> [2,] 0.8682481 0.8570410 0.8504432 0.9260021 0.8930438 0.8017223 0.8623001
#> [3,] 0.7616608 0.7890077 0.8853412 0.8336445 0.7973976 0.7289099 0.9256022
#>          [,64]     [,65]     [,66]     [,67]     [,68]     [,69]     [,70]
#> [1,] 0.8998781 0.8521724 0.8540725 0.8436454 0.9051283 0.7038300 0.9238687
#> [2,] 0.8758444 0.9234001 0.8586364 0.8680910 0.8381334 0.8492218 0.8090387
#> [3,] 0.9378822 0.8942558 0.8304243 0.8665814 0.9172459 0.8762036 0.8440560
#>          [,71]     [,72]     [,73]     [,74]     [,75]     [,76]     [,77]
#> [1,] 0.6967482 0.6852197 0.8071427 0.8196453 0.7910830 0.9224122 0.8084232
#> [2,] 0.8314074 0.9497612 0.9835318 0.8398754 0.8953904 0.7649320 0.9483206
#> [3,] 0.9238435 0.9361182 0.7925583 0.8060915 0.8636291 0.8964907 0.7419340
#>          [,78]     [,79]     [,80]     [,81]     [,82]     [,83]     [,84]
#> [1,] 0.7418837 0.9423231 0.6776345 0.8829712 0.9016426 0.8170246 0.8642042
#> [2,] 0.8476703 0.9566127 0.8774936 0.7615699 0.7481179 0.9118719 0.9276245
#> [3,] 0.8795283 0.7228932 0.7800052 0.7867621 0.6671192 0.8392579 0.9219939
#>          [,85]     [,86]     [,87]     [,88]     [,89]     [,90]     [,91]
#> [1,] 0.8550667 0.9118695 0.8733794 0.8649852 0.9129793 0.9592934 0.9416891
#> [2,] 0.6904283 0.8373427 0.8670555 0.8837160 0.9172207 0.8281723 0.9204021
#> [3,] 0.8828813 0.9098252 0.8320826 0.8724895 0.8104590 0.9107188 0.8351000
#>          [,92]     [,93]     [,94]     [,95]     [,96]     [,97]     [,98]
#> [1,] 0.8140195 0.8443556 0.9589558 0.7692932 0.8195157 0.9046333 0.8266810
#> [2,] 0.8826794 0.8385556 0.9201355 0.9290506 0.8887590 0.7441011 0.8841965
#> [3,] 0.8229106 0.7947541 0.9208993 0.8903287 0.9289014 0.8963132 0.9589371
#>          [,99]    [,100]
#> [1,] 0.8294706 0.8615571
#> [2,] 0.9239445 0.7753189
#> [3,] 0.8901697 0.7418943
#> 
#> , , 4
#> 
#>           [,1]      [,2]      [,3]      [,4]      [,5]      [,6]      [,7]
#> [1,] 0.9089690 0.9309081 0.9644331 0.9571147 0.9401363 0.9414034 0.9544975
#> [2,] 0.9424814 0.9900445 0.8710137 0.9057138 0.9898660 0.9872637 0.9663425
#> [3,] 0.8959032 0.9679070 0.9468310 0.8696679 0.8994770 0.9567815 0.8693561
#>           [,8]      [,9]     [,10]     [,11]     [,12]     [,13]     [,14]
#> [1,] 0.8792322 0.9056194 0.9285127 0.9768863 0.9682306 0.9416506 0.9949206
#> [2,] 0.9980131 0.9800743 0.9865527 0.9757480 0.9593726 0.9640241 0.9562883
#> [3,] 0.9665850 0.8945389 0.8777745 0.9622079 0.9641054 0.9520830 0.9899677
#>          [,15]     [,16]     [,17]     [,18]     [,19]     [,20]     [,21]
#> [1,] 0.9917820 0.9808039 0.9469516 0.9881520 0.9201191 0.9976296 0.9890382
#> [2,] 0.9830041 0.9885728 0.9856041 0.9715634 0.9502849 0.9869182 0.9340241
#> [3,] 0.9437709 0.9898003 0.8859758 0.9897768 0.9583505 0.9429517 0.9891588
#>          [,22]     [,23]     [,24]     [,25]     [,26]     [,27]     [,28]
#> [1,] 0.9476390 0.9685256 0.9000553 0.9285531 0.9637215 0.9615567 0.9202328
#> [2,] 0.9155428 0.9326172 0.9883384 0.9437181 0.9865486 0.9463216 0.8742665
#> [3,] 0.8863727 0.9683025 0.9832454 0.9328134 0.9372012 0.9413873 0.9888043
#>          [,29]     [,30]     [,31]     [,32]     [,33]     [,34]     [,35]
#> [1,] 0.9462198 0.9941639 0.9783791 0.9997485 0.9934411 0.9282244 0.9404766
#> [2,] 0.9891563 0.9644889 0.9813245 0.9753894 0.9335825 0.9418396 0.9673037
#> [3,] 0.9836974 0.9593712 0.8664540 0.8887586 0.8850441 0.9432629 0.9914758
#>          [,36]     [,37]     [,38]     [,39]     [,40]     [,41]     [,42]
#> [1,] 0.9592070 0.9469280 0.9703422 0.8970686 0.9394089 0.9769556 0.9628378
#> [2,] 0.9839406 0.8876998 0.9704830 0.9240271 0.9590213 0.9972190 0.9725683
#> [3,] 0.9678457 0.9754055 0.9240306 0.9479248 0.9789813 0.9761877 0.9799692
#>          [,43]     [,44]     [,45]     [,46]     [,47]     [,48]     [,49]
#> [1,] 0.9753528 0.9299283 0.9402095 0.8771257 0.9474215 0.9532325 0.9895956
#> [2,] 0.9227137 0.7781500 0.9490275 0.9036037 0.9943251 0.9910080 0.9855305
#> [3,] 0.9656864 0.9221832 0.9918919 0.8457753 0.9631675 0.9678011 0.8775056
#>          [,50]     [,51]     [,52]     [,53]     [,54]     [,55]     [,56]
#> [1,] 0.9918270 0.9776934 0.9862256 0.9500635 0.9593121 0.9225767 0.9280386
#> [2,] 0.9209401 0.9590488 0.9419600 0.9947637 0.9291663 0.9335311 0.9493182
#> [3,] 0.9464140 0.9748515 0.9877541 0.9591589 0.8949126 0.9940973 0.9259708
#>          [,57]     [,58]     [,59]     [,60]     [,61]     [,62]     [,63]
#> [1,] 0.9620578 0.9177859 0.9842447 0.9789307 0.9349837 0.9209877 0.9089577
#> [2,] 0.9347969 0.9822073 0.9930509 0.9468077 0.8800447 0.9123120 0.8682172
#> [3,] 0.9518699 0.9342874 0.9533654 0.9692109 0.8890076 0.9891635 0.9020053
#>          [,64]     [,65]     [,66]     [,67]     [,68]     [,69]     [,70]
#> [1,] 0.9817461 0.9680085 0.9361217 0.9060937 0.9444437 0.8822589 0.9523049
#> [2,] 0.9628950 0.9348175 0.8713308 0.9479601 0.9669241 0.9691106 0.9973809
#> [3,] 0.9695820 0.9097436 0.8706679 0.9438554 0.9361859 0.9405045 0.9825850
#>          [,71]     [,72]     [,73]     [,74]     [,75]     [,76]     [,77]
#> [1,] 0.9940155 0.9708007 0.9770024 0.8993536 0.9423513 0.9865406 0.9003348
#> [2,] 0.9091172 0.8696724 0.9611408 0.9002874 0.8397609 0.9813197 0.9805989
#> [3,] 0.9412135 0.9952430 0.9632325 0.8798998 0.9429024 0.9477355 0.9165908
#>          [,78]     [,79]     [,80]     [,81]     [,82]     [,83]     [,84]
#> [1,] 0.9888168 0.9804203 0.9831528 0.9380842 0.9407274 0.8748807 0.8979882
#> [2,] 0.9407115 0.9470308 0.9889822 0.9442447 0.9380449 0.9845150 0.9850293
#> [3,] 0.8924063 0.9228027 0.9043470 0.9311780 0.9064411 0.9864148 0.9522677
#>          [,85]     [,86]     [,87]     [,88]     [,89]     [,90]     [,91]
#> [1,] 0.8929748 0.9323078 0.9742206 0.9298528 0.9728353 0.9628285 0.9264786
#> [2,] 0.9862019 0.9769168 0.9174295 0.9468010 0.9299539 0.9692835 0.9781554
#> [3,] 0.9751259 0.9251663 0.9328571 0.9951218 0.9074691 0.9615958 0.9981840
#>          [,92]     [,93]     [,94]     [,95]     [,96]     [,97]     [,98]
#> [1,] 0.8922777 0.9565770 0.9580062 0.9938361 0.9785258 0.9452929 0.9297318
#> [2,] 0.9818139 0.9613096 0.9619260 0.9349790 0.9953192 0.9409398 0.9523714
#> [3,] 0.9357228 0.9296157 0.9880128 0.9768364 0.9317509 0.9830330 0.9436084
#>          [,99]    [,100]
#> [1,] 0.9171872 0.9725403
#> [2,] 0.9947128 0.9943845
#> [3,] 0.9883932 0.9977893
#> 
#> , , 5
#> 
#>           [,1]      [,2]      [,3]      [,4]      [,5]      [,6]      [,7]
#> [1,] 0.6671095 0.6812501 0.6165308 0.7456585 0.6712411 0.7612416 0.7961667
#> [2,] 0.6546328 0.6168316 0.8263333 0.7931230 0.7246530 0.7677807 0.6844154
#> [3,] 0.6571816 0.7510101 0.6328577 0.7057502 0.8651662 0.7099198 0.6583721
#>           [,8]      [,9]     [,10]     [,11]     [,12]     [,13]     [,14]
#> [1,] 0.7507710 0.7358820 0.7946316 0.5471628 0.5633488 0.5531390 0.6189648
#> [2,] 0.5224143 0.4340525 0.5788350 0.6455682 0.8301687 0.7382807 0.6826356
#> [3,] 0.4926304 0.4655878 0.7136711 0.7644236 0.8039323 0.5508265 0.7204025
#>          [,15]     [,16]     [,17]     [,18]     [,19]     [,20]     [,21]
#> [1,] 0.5414877 0.5631744 0.7728278 0.5940068 0.6960996 0.5886364 0.7793416
#> [2,] 0.6903370 0.5801808 0.6810815 0.7757038 0.6715547 0.7301240 0.8032695
#> [3,] 0.5272018 0.6182046 0.7348106 0.6330788 0.7725534 0.6680271 0.7658489
#>          [,22]     [,23]     [,24]     [,25]     [,26]     [,27]     [,28]
#> [1,] 0.6742642 0.7760595 0.6850285 0.7499466 0.6504591 0.7674408 0.6772221
#> [2,] 0.7881444 0.7289241 0.8091614 0.6332621 0.7473403 0.7603140 0.7414029
#> [3,] 0.8317684 0.7251032 0.5577614 0.7299714 0.8306654 0.7226062 0.5086108
#>          [,29]     [,30]     [,31]     [,32]     [,33]     [,34]     [,35]
#> [1,] 0.7351033 0.7551530 0.6812300 0.5307066 0.8065109 0.6056142 0.5028731
#> [2,] 0.6391419 0.6623537 0.6811274 0.6683590 0.6568731 0.7244582 0.7856582
#> [3,] 0.7209618 0.7002189 0.7399624 0.7339780 0.5552515 0.7710559 0.6072423
#>          [,36]     [,37]     [,38]     [,39]     [,40]     [,41]     [,42]
#> [1,] 0.7158237 0.5566633 0.7284331 0.6114879 0.6367075 0.6686362 0.6770445
#> [2,] 0.7289401 0.6018055 0.6305629 0.5966601 0.7940129 0.5712624 0.5593245
#> [3,] 0.6951081 0.7291568 0.6399098 0.7390586 0.7519389 0.6103855 0.6798107
#>          [,43]     [,44]     [,45]     [,46]     [,47]     [,48]     [,49]
#> [1,] 0.6453133 0.7021876 0.6033238 0.7173359 0.7001436 0.6189915 0.6871396
#> [2,] 0.8407509 0.7440646 0.5957717 0.8004952 0.6621070 0.6550656 0.7000494
#> [3,] 0.6213363 0.7112919 0.7746692 0.7855921 0.7345934 0.7460167 0.7088462
#>          [,50]     [,51]     [,52]     [,53]     [,54]     [,55]     [,56]
#> [1,] 0.5294796 0.7049402 0.6717924 0.6732593 0.6396639 0.8112256 0.5301826
#> [2,] 0.5711359 0.5273657 0.7673464 0.5871536 0.6959793 0.7394463 0.7448794
#> [3,] 0.6335882 0.5688859 0.7439476 0.5008162 0.7830447 0.7597466 0.7197316
#>          [,57]     [,58]     [,59]     [,60]     [,61]     [,62]     [,63]
#> [1,] 0.6831922 0.7276090 0.6390612 0.7810301 0.7132677 0.7022783 0.6077038
#> [2,] 0.8610622 0.6271756 0.5934255 0.5340220 0.6707719 0.6710633 0.8060965
#> [3,] 0.6665331 0.6403394 0.6718444 0.4730214 0.6013065 0.7671460 0.6953483
#>          [,64]     [,65]     [,66]     [,67]     [,68]     [,69]     [,70]
#> [1,] 0.5587883 0.5614271 0.7907951 0.5041350 0.8065552 0.6324437 0.5488417
#> [2,] 0.6681204 0.5464159 0.6433686 0.6112448 0.7289903 0.6950075 0.6219289
#> [3,] 0.7757986 0.5813366 0.7586613 0.7382617 0.6485643 0.7247057 0.7988612
#>          [,71]     [,72]     [,73]     [,74]     [,75]     [,76]     [,77]
#> [1,] 0.5514447 0.5556435 0.6997799 0.7332493 0.6045459 0.6767609 0.6032130
#> [2,] 0.5664374 0.6355065 0.7480061 0.7555426 0.7460287 0.6920013 0.5835536
#> [3,] 0.6171324 0.5868211 0.7016501 0.8345871 0.7639214 0.8421009 0.6330981
#>          [,78]     [,79]     [,80]     [,81]     [,82]     [,83]     [,84]
#> [1,] 0.6972555 0.6955739 0.5442876 0.7048465 0.7417070 0.8306495 0.7377526
#> [2,] 0.5829835 0.6194462 0.6724657 0.6935904 0.6520671 0.6597753 0.5578402
#> [3,] 0.8468380 0.7283226 0.8349085 0.5883561 0.6720385 0.5416208 0.7162320
#>          [,85]     [,86]     [,87]     [,88]     [,89]     [,90]     [,91]
#> [1,] 0.8362683 0.5088571 0.5935783 0.7139427 0.6146512 0.6716758 0.7563561
#> [2,] 0.7183921 0.8235125 0.7520186 0.8828585 0.6687283 0.5697420 0.7958146
#> [3,] 0.5692797 0.8277074 0.5881323 0.5975786 0.6643719 0.6429476 0.5646034
#>          [,92]     [,93]     [,94]     [,95]     [,96]     [,97]     [,98]
#> [1,] 0.6984414 0.7440319 0.7195688 0.5582543 0.5393894 0.7895238 0.7301392
#> [2,] 0.7828808 0.6050589 0.6848768 0.5998422 0.7233138 0.5086818 0.5681289
#> [3,] 0.6409463 0.6412171 0.6784021 0.7116610 0.5260135 0.7021447 0.7134651
#>          [,99]    [,100]
#> [1,] 0.7288353 0.7275041
#> [2,] 0.6718102 0.6257651
#> [3,] 0.7123201 0.6906958
#> 
#> , , 6
#> 
#>           [,1]      [,2]      [,3]      [,4]      [,5]      [,6]      [,7]
#> [1,] 0.8774637 0.9435257 0.9742930 0.9541549 0.9184630 0.9069609 0.9648386
#> [2,] 0.9656812 0.9235622 0.8642720 0.9085583 0.9137575 0.9928852 0.9656771
#> [3,] 0.9303682 0.9166397 0.8641745 0.8960762 0.9123001 0.8311360 0.9068971
#>           [,8]      [,9]     [,10]     [,11]     [,12]     [,13]     [,14]
#> [1,] 0.9363716 0.8863568 0.9799754 0.9512223 0.8787542 0.9448033 0.9158743
#> [2,] 0.8847019 0.8800783 0.8742949 0.8483893 0.9334238 0.9278590 0.9796172
#> [3,] 0.9411757 0.9501868 0.9524554 0.9163599 0.8627831 0.9059547 0.8944133
#>          [,15]     [,16]     [,17]     [,18]     [,19]     [,20]     [,21]
#> [1,] 0.9875868 0.9623192 0.9195782 0.9445119 0.9372357 0.9036342 0.9855416
#> [2,] 0.9869347 0.9783148 0.9297032 0.8760284 0.9023937 0.9443494 0.8971263
#> [3,] 0.8678678 0.9414826 0.9397115 0.9330218 0.9535470 0.9224547 0.9580124
#>          [,22]     [,23]     [,24]     [,25]     [,26]     [,27]     [,28]
#> [1,] 0.9231674 0.9119583 0.8378782 0.8906240 0.9677807 0.9476033 0.8857524
#> [2,] 0.8800305 0.9151528 0.9450238 0.9081844 0.9633026 0.8722711 0.9795903
#> [3,] 0.9300644 0.9455913 0.9061034 0.8817211 0.9563840 0.9177498 0.9015450
#>          [,29]     [,30]     [,31]     [,32]     [,33]     [,34]     [,35]
#> [1,] 0.9123961 0.9435924 0.9437463 0.9127685 0.9746222 0.8736142 0.9512114
#> [2,] 0.9634216 0.9243432 0.9147717 0.9250358 0.8824086 0.9275999 0.8600339
#> [3,] 0.9576664 0.8049304 0.8744541 0.8949807 0.8901005 0.8852164 0.8431278
#>          [,36]     [,37]     [,38]     [,39]     [,40]     [,41]     [,42]
#> [1,] 0.8827683 0.9036546 0.9265322 0.7575169 0.8433736 0.9307281 0.9061297
#> [2,] 0.9152881 0.8444689 0.9336117 0.8892357 0.7985430 0.9498746 0.8876310
#> [3,] 0.9125826 0.8922251 0.9203888 0.9095640 0.9295597 0.9103223 0.9205120
#>          [,43]     [,44]     [,45]     [,46]     [,47]     [,48]     [,49]
#> [1,] 0.9184067 0.9227848 0.9220462 0.8796718 0.9401379 0.8418050 0.9148885
#> [2,] 0.9233403 0.8399377 0.8279218 0.8662649 0.8186101 0.9522839 0.9043936
#> [3,] 0.8391596 0.9327296 0.9254897 0.9204011 0.9477040 0.9519321 0.9637756
#>          [,50]     [,51]     [,52]     [,53]     [,54]     [,55]     [,56]
#> [1,] 0.9606758 0.9804430 0.9256341 0.9328278 0.9253853 0.8438637 0.9546775
#> [2,] 0.8498303 0.9628688 0.8935900 0.9745565 0.9723035 0.8291525 0.9013922
#> [3,] 0.8860205 0.8412955 0.8850740 0.9557716 0.8917526 0.8918212 0.9672968
#>          [,57]     [,58]     [,59]     [,60]     [,61]     [,62]     [,63]
#> [1,] 0.9268515 0.9048080 0.9087740 0.9508157 0.9196868 0.8533188 0.8856820
#> [2,] 0.8859631 0.8604637 0.9114622 0.9520191 0.8896642 0.8967713 0.8290831
#> [3,] 0.9375366 0.9232103 0.9897010 0.9544767 0.8706617 0.9061595 0.7992172
#>          [,64]     [,65]     [,66]     [,67]     [,68]     [,69]     [,70]
#> [1,] 0.9269197 0.9597012 0.8929113 0.8909279 0.9422515 0.9070876 0.9049640
#> [2,] 0.9107860 0.9255279 0.8891768 0.8821033 0.8900039 0.9483055 0.9641357
#> [3,] 0.9128312 0.8099323 0.8522201 0.9152620 0.9689940 0.9041041 0.9449431
#>          [,71]     [,72]     [,73]     [,74]     [,75]     [,76]     [,77]
#> [1,] 0.9819132 0.9303556 0.9315815 0.9233970 0.7931444 0.8729086 0.9479404
#> [2,] 0.9163478 0.9776620 0.9487592 0.8876668 0.9453934 0.9695936 0.9146315
#> [3,] 0.8880726 0.9283894 0.9121191 0.8780728 0.8508840 0.8959349 0.9122572
#>          [,78]     [,79]     [,80]     [,81]     [,82]     [,83]     [,84]
#> [1,] 0.8605262 0.9523610 0.9583941 0.9105013 0.8906926 0.9032082 0.8762416
#> [2,] 0.9729016 0.9735164 0.9329891 0.8572450 0.9655663 0.8342395 0.9008041
#> [3,] 0.9334851 0.9078010 0.8633391 0.8959383 0.8908605 0.8770546 0.9442697
#>          [,85]     [,86]     [,87]     [,88]     [,89]     [,90]     [,91]
#> [1,] 0.8999717 0.9297698 0.9118684 0.9204946 0.9298328 0.9521244 0.9249482
#> [2,] 0.9081741 0.9554772 0.9243742 0.9551361 0.9678479 0.9547715 0.9668574
#> [3,] 0.9363046 0.9171186 0.9450914 0.8862662 0.9423493 0.9073945 0.9313349
#>          [,92]     [,93]     [,94]     [,95]     [,96]     [,97]     [,98]
#> [1,] 0.9649643 0.7589829 0.9269947 0.9051939 0.9208981 0.9007689 0.9359923
#> [2,] 0.9505972 0.9120728 0.8126948 0.9308690 0.9753057 0.9324367 0.9474794
#> [3,] 0.8132555 0.9121029 0.9332242 0.9391113 0.9275095 0.9555618 0.9478237
#>          [,99]    [,100]
#> [1,] 0.9415020 0.9364308
#> [2,] 0.9715197 0.9662184
#> [3,] 0.9239959 0.9349233
#> 
#> , , 7
#> 
#>           [,1]      [,2]      [,3]      [,4]      [,5]      [,6]      [,7]
#> [1,] 0.9645739 0.9322528 0.9425896 0.8486645 0.9222936 0.9126693 0.9585741
#> [2,] 0.9637566 0.8821965 0.8763269 0.9388315 0.9130556 0.9304043 0.9284445
#> [3,] 0.9373253 0.9255850 0.9387933 0.8373553 0.8951305 0.9343669 0.9079141
#>           [,8]      [,9]     [,10]     [,11]     [,12]     [,13]     [,14]
#> [1,] 0.8760294 0.8381794 0.8571092 0.9455737 0.8078467 0.8961563 0.7741971
#> [2,] 0.9798982 0.8956846 0.8854252 0.9343308 0.9380593 0.9630397 0.8916004
#> [3,] 0.9012537 0.9786743 0.9264966 0.8899519 0.9695132 0.9382321 0.9857760
#>          [,15]     [,16]     [,17]     [,18]     [,19]     [,20]     [,21]
#> [1,] 0.8448348 0.9333258 0.9332659 0.9011018 0.9462865 0.9685760 0.9588464
#> [2,] 0.9748999 0.9224157 0.9303996 0.8678272 0.8916098 0.9494222 0.9623555
#> [3,] 0.9639626 0.9599164 0.9639919 0.9411709 0.9094913 0.9449408 0.9848797
#>          [,22]     [,23]     [,24]     [,25]     [,26]     [,27]     [,28]
#> [1,] 0.8828008 0.9774622 0.9126558 0.9517202 0.9627553 0.9448009 0.9340237
#> [2,] 0.9367757 0.8735623 0.9454647 0.9369554 0.9162171 0.8624129 0.9725974
#> [3,] 0.9498502 0.8087983 0.8612448 0.9219865 0.9365277 0.9251879 0.9176900
#>          [,29]     [,30]     [,31]     [,32]     [,33]     [,34]     [,35]
#> [1,] 0.9723416 0.9678703 0.8891970 0.9684429 0.9276250 0.8412346 0.9814492
#> [2,] 0.8708086 0.8263400 0.8913099 0.9874697 0.9668455 0.8890854 0.9809257
#> [3,] 0.9815235 0.9514918 0.9106108 0.9065266 0.8918026 0.9493763 0.9576119
#>          [,36]     [,37]     [,38]     [,39]     [,40]     [,41]     [,42]
#> [1,] 0.9381647 0.9101386 0.9177256 0.9025891 0.9527395 0.9391068 0.9590238
#> [2,] 0.9026736 0.9830764 0.9588413 0.9528358 0.9154946 0.9874988 0.9655704
#> [3,] 0.9786474 0.9338194 0.8607432 0.9213877 0.9311743 0.9473074 0.8831369
#>          [,43]     [,44]     [,45]     [,46]     [,47]     [,48]     [,49]
#> [1,] 0.9582789 0.8434107 0.8854759 0.8712748 0.9819431 0.9485475 0.9411627
#> [2,] 0.9546007 0.8658209 0.9371945 0.9664809 0.9438513 0.9539794 0.9568095
#> [3,] 0.9368368 0.9601101 0.9352407 0.9724244 0.9481943 0.8870444 0.9158328
#>          [,50]     [,51]     [,52]     [,53]     [,54]     [,55]     [,56]
#> [1,] 0.9052575 0.9012352 0.9895975 0.8169951 0.8703574 0.9708938 0.9414226
#> [2,] 0.9617542 0.9446832 0.9516572 0.9079114 0.9148130 0.9659978 0.9048435
#> [3,] 0.9565424 0.9028174 0.9848038 0.9797605 0.8956335 0.9328737 0.9244866
#>          [,57]     [,58]     [,59]     [,60]     [,61]     [,62]     [,63]
#> [1,] 0.9493911 0.9119979 0.8990668 0.8573640 0.8593225 0.9064845 0.9623808
#> [2,] 0.8709952 0.9532094 0.9527654 0.9044063 0.8562688 0.9212191 0.8300988
#> [3,] 0.9063944 0.9646905 0.8938169 0.9666584 0.9545899 0.8898929 0.9280399
#>          [,64]     [,65]     [,66]     [,67]     [,68]     [,69]     [,70]
#> [1,] 0.9981783 0.9822198 0.9096239 0.9096195 0.9435464 0.9702264 0.9496254
#> [2,] 0.9603160 0.9670991 0.8728689 0.9044150 0.9357385 0.9745804 0.8751744
#> [3,] 0.9017803 0.8931089 0.8368984 0.9687403 0.9354378 0.9741099 0.9628787
#>          [,71]     [,72]     [,73]     [,74]     [,75]     [,76]     [,77]
#> [1,] 0.9659969 0.9111810 0.9751242 0.8296613 0.9620209 0.9818858 0.9295818
#> [2,] 0.9540458 0.9469797 0.8858938 0.9236796 0.9176519 0.9725841 0.8956254
#> [3,] 0.9748169 0.9855551 0.9499142 0.9345432 0.9015995 0.8740409 0.9114413
#>          [,78]     [,79]     [,80]     [,81]     [,82]     [,83]     [,84]
#> [1,] 0.9719852 0.9880046 0.9415586 0.9044113 0.9271083 0.8933244 0.9383434
#> [2,] 0.8978511 0.9038331 0.9755731 0.9152888 0.9429125 0.9562232 0.9822152
#> [3,] 0.9259048 0.9009423 0.8972609 0.8821015 0.8751183 0.9426180 0.9824655
#>          [,85]     [,86]     [,87]     [,88]     [,89]     [,90]     [,91]
#> [1,] 0.9623492 0.9451876 0.8660130 0.9758427 0.9611442 0.9256910 0.9550531
#> [2,] 0.9086011 0.8960849 0.8893016 0.9162775 0.9140786 0.8283209 0.9498448
#> [3,] 0.9628582 0.9190228 0.9257098 0.9056749 0.9619679 0.9569462 0.9814054
#>          [,92]     [,93]     [,94]     [,95]     [,96]     [,97]     [,98]
#> [1,] 0.8895393 0.9307159 0.9871694 0.9337828 0.9589471 0.8809662 0.9347076
#> [2,] 0.9537990 0.9192553 0.9623175 0.9558911 0.9684062 0.9519286 0.9222397
#> [3,] 0.8740190 0.9024484 0.9785091 0.9277725 0.9448681 0.9698068 0.9625117
#>          [,99]    [,100]
#> [1,] 0.9284908 0.9667299
#> [2,] 0.9030375 0.9888026
#> [3,] 0.8299797 0.9913099
#> 
#> , , 8
#> 
#>           [,1]      [,2]      [,3]      [,4]      [,5]      [,6]      [,7]
#> [1,] 0.9140512 0.9533715 0.9169355 0.9707764 0.9065994 0.9185528 0.9627537
#> [2,] 0.9612550 0.9544251 0.8509722 0.9053884 0.8574859 0.9153868 0.9716090
#> [3,] 0.9705201 0.9539478 0.8715788 0.8726978 0.9247342 0.9652367 0.9163872
#>           [,8]      [,9]     [,10]     [,11]     [,12]     [,13]     [,14]
#> [1,] 0.8231886 0.9123468 0.8670474 0.9686806 0.9567706 0.9580868 0.8597760
#> [2,] 0.9961243 0.9229234 0.9280289 0.9251095 0.8930060 0.9487068 0.8707039
#> [3,] 0.9784829 0.9265272 0.9695893 0.9722788 0.9113388 0.9419114 0.9606137
#>          [,15]     [,16]     [,17]     [,18]     [,19]     [,20]     [,21]
#> [1,] 0.9460450 0.8771769 0.8647815 0.8821031 0.9561068 0.9467810 0.9909943
#> [2,] 0.9714642 0.9713415 0.9108372 0.9658606 0.8959678 0.9648864 0.9722245
#> [3,] 0.9598057 0.9525873 0.9688049 0.9718722 0.8689552 0.9702241 0.9837382
#>          [,22]     [,23]     [,24]     [,25]     [,26]     [,27]     [,28]
#> [1,] 0.9021436 0.9481290 0.9540721 0.9082525 0.9056174 0.9068052 0.9896584
#> [2,] 0.8900223 0.8788782 0.8996848 0.9538371 0.9191036 0.9604362 0.9535889
#> [3,] 0.9361997 0.9127656 0.9295911 0.9167222 0.9544290 0.9249190 0.9898161
#>          [,29]     [,30]     [,31]     [,32]     [,33]     [,34]     [,35]
#> [1,] 0.9267503 0.9879554 0.8073215 0.9558312 0.9861797 0.8705416 0.9420153
#> [2,] 0.9791989 0.9576900 0.9034444 0.9416455 0.8453675 0.8832467 0.9334117
#> [3,] 0.9312678 0.9566474 0.8948859 0.8418940 0.8947599 0.9129361 0.8799207
#>          [,36]     [,37]     [,38]     [,39]     [,40]     [,41]     [,42]
#> [1,] 0.9548553 0.9620193 0.9304166 0.9460245 0.9455646 0.9759707 0.9801810
#> [2,] 0.9807428 0.9189967 0.9583398 0.9216044 0.9467636 0.9726643 0.9693034
#> [3,] 0.9178386 0.8829306 0.9571516 0.9261494 0.9487032 0.9707214 0.9640126
#>          [,43]     [,44]     [,45]     [,46]     [,47]     [,48]     [,49]
#> [1,] 0.9574251 0.9185264 0.9152546 0.8970203 0.9450966 0.9518286 0.9411072
#> [2,] 0.9624323 0.8906234 0.9158943 0.9392230 0.9840254 0.9375289 0.9630075
#> [3,] 0.9785978 0.9775645 0.9034280 0.8565488 0.9505742 0.9129292 0.9358879
#>          [,50]     [,51]     [,52]     [,53]     [,54]     [,55]     [,56]
#> [1,] 0.9591460 0.9752616 0.9704916 0.9351941 0.9529573 0.9387729 0.9898356
#> [2,] 0.9198723 0.9039384 0.9615837 0.9531276 0.9622563 0.9130520 0.9423408
#> [3,] 0.9835424 0.9824180 0.9486915 0.9737796 0.9076391 0.9871117 0.9282582
#>          [,57]     [,58]     [,59]     [,60]     [,61]     [,62]     [,63]
#> [1,] 0.9272125 0.9560330 0.9720127 0.9750098 0.8911744 0.9701941 0.9844705
#> [2,] 0.9320594 0.9330625 0.9572392 0.8707374 0.8351459 0.9281507 0.7729045
#> [3,] 0.9449681 0.9748161 0.9509369 0.9709949 0.9194317 0.9554867 0.9229994
#>          [,64]     [,65]     [,66]     [,67]     [,68]     [,69]     [,70]
#> [1,] 0.9011947 0.9854394 0.9410922 0.9378756 0.9372372 0.8466708 0.9420102
#> [2,] 0.8665849 0.9739968 0.9576988 0.9726971 0.9295765 0.9399872 0.9272892
#> [3,] 0.9430016 0.9595773 0.8407731 0.8688068 0.9726927 0.9773066 0.9409305
#>          [,71]     [,72]     [,73]     [,74]     [,75]     [,76]     [,77]
#> [1,] 0.9907771 0.8264574 0.9536716 0.8425610 0.7196740 0.9882754 0.9041809
#> [2,] 0.9704689 0.9686764 0.9465840 0.9084125 0.9353628 0.9821913 0.8820183
#> [3,] 0.9193727 0.8235792 0.9360355 0.9083081 0.8712078 0.9232547 0.8869572
#>          [,78]     [,79]     [,80]     [,81]     [,82]     [,83]     [,84]
#> [1,] 0.9818437 0.9640323 0.9637448 0.9584508 0.9655171 0.9261241 0.9579162
#> [2,] 0.9706356 0.8939615 0.9421560 0.9688546 0.9675902 0.9380451 0.9743880
#> [3,] 0.9265097 0.8890403 0.8953560 0.9359014 0.8575220 0.9134331 0.9849591
#>          [,85]     [,86]     [,87]     [,88]     [,89]     [,90]     [,91]
#> [1,] 0.9365574 0.9513760 0.8909563 0.9545208 0.9573306 0.9790783 0.9567911
#> [2,] 0.9873985 0.7754663 0.9191787 0.8874262 0.9482178 0.9499990 0.9629180
#> [3,] 0.9235984 0.9431474 0.9790072 0.8747650 0.9697794 0.9092209 0.8983449
#>          [,92]     [,93]     [,94]     [,95]     [,96]     [,97]     [,98]
#> [1,] 0.9548155 0.9587900 0.9540930 0.9237652 0.9635879 0.9616431 0.9395403
#> [2,] 0.9507541 0.9502997 0.9281425 0.9548647 0.9729385 0.9139935 0.9592807
#> [3,] 0.9281473 0.9646086 0.9746222 0.9262199 0.8824479 0.8704673 0.9226471
#>          [,99]    [,100]
#> [1,] 0.9523783 0.9160150
#> [2,] 0.9758517 0.9375031
#> [3,] 0.9688189 0.9977511
#> 
#> , , 9
#> 
#>           [,1]      [,2]      [,3]      [,4]      [,5]      [,6]      [,7]
#> [1,] 0.7938533 0.8549439 0.7784774 0.8734194 0.8978808 0.8247742 0.7972926
#> [2,] 0.8462314 0.7826510 0.8017332 0.7127540 0.7853141 0.7452720 0.7835279
#> [3,] 0.9218941 0.7572890 0.7317937 0.6413919 0.8718792 0.5977998 0.8397932
#>           [,8]      [,9]     [,10]     [,11]     [,12]     [,13]     [,14]
#> [1,] 0.8237689 0.7358544 0.7457572 0.7373950 0.6941152 0.7364569 0.8254512
#> [2,] 0.8548136 0.8848987 0.8313935 0.6832983 0.8959018 0.8497792 0.8871225
#> [3,] 0.7367469 0.6163312 0.9067038 0.8776114 0.8590328 0.7098745 0.8387398
#>          [,15]     [,16]     [,17]     [,18]     [,19]     [,20]     [,21]
#> [1,] 0.7906247 0.8132770 0.8465491 0.7212451 0.9204899 0.9241257 0.8613355
#> [2,] 0.6764636 0.7663901 0.8480811 0.8360312 0.8831518 0.8119473 0.8801671
#> [3,] 0.8141265 0.7968397 0.7550005 0.9362160 0.8538087 0.8923362 0.8487513
#>          [,22]     [,23]     [,24]     [,25]     [,26]     [,27]     [,28]
#> [1,] 0.8327602 0.9086323 0.7671230 0.7736386 0.7831387 0.8831976 0.8121436
#> [2,] 0.8731775 0.8683434 0.8568767 0.8065296 0.8006989 0.8478434 0.8895711
#> [3,] 0.8381646 0.8717860 0.7540676 0.8950240 0.8748109 0.8085619 0.5793277
#>          [,29]     [,30]     [,31]     [,32]     [,33]     [,34]     [,35]
#> [1,] 0.7602998 0.7661915 0.8504427 0.7039186 0.8783553 0.7875429 0.8245903
#> [2,] 0.8442044 0.6666791 0.8615191 0.5494191 0.7367923 0.8117114 0.9047683
#> [3,] 0.8037018 0.7243977 0.8234119 0.7507431 0.8746208 0.7945596 0.8898007
#>          [,36]     [,37]     [,38]     [,39]     [,40]     [,41]     [,42]
#> [1,] 0.8478127 0.8350714 0.7461278 0.8749962 0.9006622 0.8715197 0.8576128
#> [2,] 0.8539452 0.7529556 0.8491505 0.8901330 0.7706552 0.7308012 0.9072255
#> [3,] 0.8516317 0.8866792 0.8626835 0.8702216 0.8032107 0.7812276 0.8586946
#>          [,43]     [,44]     [,45]     [,46]     [,47]     [,48]     [,49]
#> [1,] 0.8856969 0.8029884 0.8320865 0.8599035 0.9071328 0.7791583 0.9012204
#> [2,] 0.9134217 0.7590166 0.6573730 0.7592103 0.8297618 0.8671074 0.8200488
#> [3,] 0.8621788 0.8577782 0.8235589 0.9010934 0.6548298 0.8556545 0.8621439
#>          [,50]     [,51]     [,52]     [,53]     [,54]     [,55]     [,56]
#> [1,] 0.8584898 0.8523540 0.8709411 0.6939374 0.8630973 0.8979130 0.8163741
#> [2,] 0.7522199 0.8121527 0.8086270 0.9024712 0.8681842 0.8794400 0.8308841
#> [3,] 0.9315385 0.8431145 0.9079878 0.8004416 0.9304955 0.7257929 0.8139157
#>          [,57]     [,58]     [,59]     [,60]     [,61]     [,62]     [,63]
#> [1,] 0.9041312 0.8640401 0.8237663 0.7366199 0.7589144 0.8153377 0.8611264
#> [2,] 0.8511599 0.8197654 0.7410964 0.7900933 0.8061639 0.8366339 0.7741647
#> [3,] 0.8509041 0.8221839 0.9248542 0.7902958 0.8847277 0.9110830 0.8565645
#>          [,64]     [,65]     [,66]     [,67]     [,68]     [,69]     [,70]
#> [1,] 0.8413870 0.8970149 0.9133863 0.7282819 0.7933622 0.7362362 0.9301488
#> [2,] 0.8513747 0.7233123 0.9107475 0.7704442 0.8397544 0.8632485 0.8561210
#> [3,] 0.7837294 0.7904017 0.8162915 0.8857422 0.9080901 0.8074167 0.7711385
#>          [,71]     [,72]     [,73]     [,74]     [,75]     [,76]     [,77]
#> [1,] 0.9747573 0.7620010 0.9274822 0.8660743 0.7014816 0.9266317 0.8666414
#> [2,] 0.8908499 0.8300596 0.7247137 0.8490967 0.8661328 0.8880221 0.7710720
#> [3,] 0.8403084 0.8651940 0.8026629 0.9455567 0.8401909 0.8438140 0.8443590
#>          [,78]     [,79]     [,80]     [,81]     [,82]     [,83]     [,84]
#> [1,] 0.7232290 0.8717683 0.8807466 0.8522313 0.7052523 0.8216101 0.8058502
#> [2,] 0.7572046 0.7160394 0.8589392 0.7640931 0.7389044 0.8738370 0.8276735
#> [3,] 0.9123281 0.8085360 0.7018717 0.8710399 0.7696624 0.8469406 0.7917169
#>          [,85]     [,86]     [,87]     [,88]     [,89]     [,90]     [,91]
#> [1,] 0.8677556 0.7142477 0.8965073 0.8934136 0.6595409 0.7985321 0.7228234
#> [2,] 0.7829975 0.9407627 0.7378135 0.8935988 0.8398117 0.8775598 0.9343043
#> [3,] 0.7874635 0.9117923 0.7529472 0.7942174 0.8885974 0.6642434 0.8382638
#>          [,92]     [,93]     [,94]     [,95]     [,96]     [,97]     [,98]
#> [1,] 0.7953004 0.8722216 0.9235810 0.9157081 0.7949208 0.8226142 0.7891980
#> [2,] 0.8213757 0.8945421 0.8765407 0.7143446 0.8081002 0.9129338 0.8730736
#> [3,] 0.8649682 0.6140551 0.8081932 0.8876260 0.6589954 0.9164699 0.8525857
#>          [,99]    [,100]
#> [1,] 0.8502412 0.6870411
#> [2,] 0.7349623 0.6204449
#> [3,] 0.9206119 0.7490953

The bb_predict_*_trend_samples() functions are also available for models fit with year_trend = TRUE.

References