国語
# クラスサイズだけ投入
koku.res1 <- brm(koku ~ year +
cs.c +
year:cs.c +
(1 + year|id) + (1 + cs.c|sid),
data = koku.nrt.lt,
prior = NULL,
chains = 4,
iter = 5000,
warmup = 2500,
see = 1234,
)
## Compiling the C++ model
## Start sampling
##
## SAMPLING FOR MODEL 'fec4628b2169c0d0c95a80cd2e9c35b0' NOW (CHAIN 1).
## Chain 1:
## Chain 1: Gradient evaluation took 0.007761 seconds
## Chain 1: 1000 transitions using 10 leapfrog steps per transition would take 77.61 seconds.
## Chain 1: Adjust your expectations accordingly!
## Chain 1:
## Chain 1:
## Chain 1: Iteration: 1 / 5000 [ 0%] (Warmup)
## Chain 1: Iteration: 500 / 5000 [ 10%] (Warmup)
## Chain 1: Iteration: 1000 / 5000 [ 20%] (Warmup)
## Chain 1: Iteration: 1500 / 5000 [ 30%] (Warmup)
## Chain 1: Iteration: 2000 / 5000 [ 40%] (Warmup)
## Chain 1: Iteration: 2500 / 5000 [ 50%] (Warmup)
## Chain 1: Iteration: 2501 / 5000 [ 50%] (Sampling)
## Chain 1: Iteration: 3000 / 5000 [ 60%] (Sampling)
## Chain 1: Iteration: 3500 / 5000 [ 70%] (Sampling)
## Chain 1: Iteration: 4000 / 5000 [ 80%] (Sampling)
## Chain 1: Iteration: 4500 / 5000 [ 90%] (Sampling)
## Chain 1: Iteration: 5000 / 5000 [100%] (Sampling)
## Chain 1:
## Chain 1: Elapsed Time: 2074.24 seconds (Warm-up)
## Chain 1: 6657.08 seconds (Sampling)
## Chain 1: 8731.32 seconds (Total)
## Chain 1:
##
## SAMPLING FOR MODEL 'fec4628b2169c0d0c95a80cd2e9c35b0' NOW (CHAIN 2).
## Chain 2:
## Chain 2: Gradient evaluation took 0.003232 seconds
## Chain 2: 1000 transitions using 10 leapfrog steps per transition would take 32.32 seconds.
## Chain 2: Adjust your expectations accordingly!
## Chain 2:
## Chain 2:
## Chain 2: Iteration: 1 / 5000 [ 0%] (Warmup)
## Chain 2: Iteration: 500 / 5000 [ 10%] (Warmup)
## Chain 2: Iteration: 1000 / 5000 [ 20%] (Warmup)
## Chain 2: Iteration: 1500 / 5000 [ 30%] (Warmup)
## Chain 2: Iteration: 2000 / 5000 [ 40%] (Warmup)
## Chain 2: Iteration: 2500 / 5000 [ 50%] (Warmup)
## Chain 2: Iteration: 2501 / 5000 [ 50%] (Sampling)
## Chain 2: Iteration: 3000 / 5000 [ 60%] (Sampling)
## Chain 2: Iteration: 3500 / 5000 [ 70%] (Sampling)
## Chain 2: Iteration: 4000 / 5000 [ 80%] (Sampling)
## Chain 2: Iteration: 4500 / 5000 [ 90%] (Sampling)
## Chain 2: Iteration: 5000 / 5000 [100%] (Sampling)
## Chain 2:
## Chain 2: Elapsed Time: 2055.95 seconds (Warm-up)
## Chain 2: 6489.97 seconds (Sampling)
## Chain 2: 8545.92 seconds (Total)
## Chain 2:
##
## SAMPLING FOR MODEL 'fec4628b2169c0d0c95a80cd2e9c35b0' NOW (CHAIN 3).
## Chain 3:
## Chain 3: Gradient evaluation took 0.003079 seconds
## Chain 3: 1000 transitions using 10 leapfrog steps per transition would take 30.79 seconds.
## Chain 3: Adjust your expectations accordingly!
## Chain 3:
## Chain 3:
## Chain 3: Iteration: 1 / 5000 [ 0%] (Warmup)
## Chain 3: Iteration: 500 / 5000 [ 10%] (Warmup)
## Chain 3: Iteration: 1000 / 5000 [ 20%] (Warmup)
## Chain 3: Iteration: 1500 / 5000 [ 30%] (Warmup)
## Chain 3: Iteration: 2000 / 5000 [ 40%] (Warmup)
## Chain 3: Iteration: 2500 / 5000 [ 50%] (Warmup)
## Chain 3: Iteration: 2501 / 5000 [ 50%] (Sampling)
## Chain 3: Iteration: 3000 / 5000 [ 60%] (Sampling)
## Chain 3: Iteration: 3500 / 5000 [ 70%] (Sampling)
## Chain 3: Iteration: 4000 / 5000 [ 80%] (Sampling)
## Chain 3: Iteration: 4500 / 5000 [ 90%] (Sampling)
## Chain 3: Iteration: 5000 / 5000 [100%] (Sampling)
## Chain 3:
## Chain 3: Elapsed Time: 5210.32 seconds (Warm-up)
## Chain 3: 1697.16 seconds (Sampling)
## Chain 3: 6907.48 seconds (Total)
## Chain 3:
##
## SAMPLING FOR MODEL 'fec4628b2169c0d0c95a80cd2e9c35b0' NOW (CHAIN 4).
## Chain 4:
## Chain 4: Gradient evaluation took 0.002891 seconds
## Chain 4: 1000 transitions using 10 leapfrog steps per transition would take 28.91 seconds.
## Chain 4: Adjust your expectations accordingly!
## Chain 4:
## Chain 4:
## Chain 4: Iteration: 1 / 5000 [ 0%] (Warmup)
## Chain 4: Iteration: 500 / 5000 [ 10%] (Warmup)
## Chain 4: Iteration: 1000 / 5000 [ 20%] (Warmup)
## Chain 4: Iteration: 1500 / 5000 [ 30%] (Warmup)
## Chain 4: Iteration: 2000 / 5000 [ 40%] (Warmup)
## Chain 4: Iteration: 2500 / 5000 [ 50%] (Warmup)
## Chain 4: Iteration: 2501 / 5000 [ 50%] (Sampling)
## Chain 4: Iteration: 3000 / 5000 [ 60%] (Sampling)
## Chain 4: Iteration: 3500 / 5000 [ 70%] (Sampling)
## Chain 4: Iteration: 4000 / 5000 [ 80%] (Sampling)
## Chain 4: Iteration: 4500 / 5000 [ 90%] (Sampling)
## Chain 4: Iteration: 5000 / 5000 [100%] (Sampling)
## Chain 4:
## Chain 4: Elapsed Time: 1817.41 seconds (Warm-up)
## Chain 4: 61.7885 seconds (Sampling)
## Chain 4: 1879.2 seconds (Total)
## Chain 4:
## Warning: There were 2751 divergent transitions after warmup. Increasing adapt_delta above 0.8 may help. See
## http://mc-stan.org/misc/warnings.html#divergent-transitions-after-warmup
## Warning: There were 4758 transitions after warmup that exceeded the maximum treedepth. Increase max_treedepth above 10. See
## http://mc-stan.org/misc/warnings.html#maximum-treedepth-exceeded
## Warning: There were 1 chains where the estimated Bayesian Fraction of Missing Information was low. See
## http://mc-stan.org/misc/warnings.html#bfmi-low
## Warning: Examine the pairs() plot to diagnose sampling problems
## Warning: The largest R-hat is 1.6, indicating chains have not mixed.
## Running the chains for more iterations may help. See
## http://mc-stan.org/misc/warnings.html#r-hat
## Warning: Bulk Effective Samples Size (ESS) is too low, indicating posterior means and medians may be unreliable.
## Running the chains for more iterations may help. See
## http://mc-stan.org/misc/warnings.html#bulk-ess
## Warning: Tail Effective Samples Size (ESS) is too low, indicating posterior variances and tail quantiles may be unreliable.
## Running the chains for more iterations may help. See
## http://mc-stan.org/misc/warnings.html#tail-ess
# モデル1 + 初期値にだけ大卒率が影響
koku.res2<- brm(koku ~ year +
cs.c +
dai.c +
year:cs.c +
(1 + year|id) + (1 + cs.c * dai.c|sid),
data = koku.nrt.lt,
prior = NULL,
chains = 4,
iter = 5000,
warmup = 2500,
see = 1234
)
## Compiling the C++ model
## Start sampling
##
## SAMPLING FOR MODEL 'e95c3961c8fbd2991e14129d68c8ae77' NOW (CHAIN 1).
## Chain 1:
## Chain 1: Gradient evaluation took 0.008297 seconds
## Chain 1: 1000 transitions using 10 leapfrog steps per transition would take 82.97 seconds.
## Chain 1: Adjust your expectations accordingly!
## Chain 1:
## Chain 1:
## Chain 1: Iteration: 1 / 5000 [ 0%] (Warmup)
## Chain 1: Iteration: 500 / 5000 [ 10%] (Warmup)
## Chain 1: Iteration: 1000 / 5000 [ 20%] (Warmup)
## Chain 1: Iteration: 1500 / 5000 [ 30%] (Warmup)
## Chain 1: Iteration: 2000 / 5000 [ 40%] (Warmup)
## Chain 1: Iteration: 2500 / 5000 [ 50%] (Warmup)
## Chain 1: Iteration: 2501 / 5000 [ 50%] (Sampling)
## Chain 1: Iteration: 3000 / 5000 [ 60%] (Sampling)
## Chain 1: Iteration: 3500 / 5000 [ 70%] (Sampling)
## Chain 1: Iteration: 4000 / 5000 [ 80%] (Sampling)
## Chain 1: Iteration: 4500 / 5000 [ 90%] (Sampling)
## Chain 1: Iteration: 5000 / 5000 [100%] (Sampling)
## Chain 1:
## Chain 1: Elapsed Time: 3230.83 seconds (Warm-up)
## Chain 1: 2452.03 seconds (Sampling)
## Chain 1: 5682.86 seconds (Total)
## Chain 1:
##
## SAMPLING FOR MODEL 'e95c3961c8fbd2991e14129d68c8ae77' NOW (CHAIN 2).
## Chain 2:
## Chain 2: Gradient evaluation took 0.004487 seconds
## Chain 2: 1000 transitions using 10 leapfrog steps per transition would take 44.87 seconds.
## Chain 2: Adjust your expectations accordingly!
## Chain 2:
## Chain 2:
## Chain 2: Iteration: 1 / 5000 [ 0%] (Warmup)
## Chain 2: Iteration: 500 / 5000 [ 10%] (Warmup)
## Chain 2: Iteration: 1000 / 5000 [ 20%] (Warmup)
## Chain 2: Iteration: 1500 / 5000 [ 30%] (Warmup)
## Chain 2: Iteration: 2000 / 5000 [ 40%] (Warmup)
## Chain 2: Iteration: 2500 / 5000 [ 50%] (Warmup)
## Chain 2: Iteration: 2501 / 5000 [ 50%] (Sampling)
## Chain 2: Iteration: 3000 / 5000 [ 60%] (Sampling)
## Chain 2: Iteration: 3500 / 5000 [ 70%] (Sampling)
## Chain 2: Iteration: 4000 / 5000 [ 80%] (Sampling)
## Chain 2: Iteration: 4500 / 5000 [ 90%] (Sampling)
## Chain 2: Iteration: 5000 / 5000 [100%] (Sampling)
## Chain 2:
## Chain 2: Elapsed Time: 3072.26 seconds (Warm-up)
## Chain 2: 2441.02 seconds (Sampling)
## Chain 2: 5513.27 seconds (Total)
## Chain 2:
##
## SAMPLING FOR MODEL 'e95c3961c8fbd2991e14129d68c8ae77' NOW (CHAIN 3).
## Chain 3:
## Chain 3: Gradient evaluation took 0.004274 seconds
## Chain 3: 1000 transitions using 10 leapfrog steps per transition would take 42.74 seconds.
## Chain 3: Adjust your expectations accordingly!
## Chain 3:
## Chain 3:
## Chain 3: Iteration: 1 / 5000 [ 0%] (Warmup)
## Chain 3: Iteration: 500 / 5000 [ 10%] (Warmup)
## Chain 3: Iteration: 1000 / 5000 [ 20%] (Warmup)
## Chain 3: Iteration: 1500 / 5000 [ 30%] (Warmup)
## Chain 3: Iteration: 2000 / 5000 [ 40%] (Warmup)
## Chain 3: Iteration: 2500 / 5000 [ 50%] (Warmup)
## Chain 3: Iteration: 2501 / 5000 [ 50%] (Sampling)
## Chain 3: Iteration: 3000 / 5000 [ 60%] (Sampling)
## Chain 3: Iteration: 3500 / 5000 [ 70%] (Sampling)
## Chain 3: Iteration: 4000 / 5000 [ 80%] (Sampling)
## Chain 3: Iteration: 4500 / 5000 [ 90%] (Sampling)
## Chain 3: Iteration: 5000 / 5000 [100%] (Sampling)
## Chain 3:
## Chain 3: Elapsed Time: 4159.54 seconds (Warm-up)
## Chain 3: 2453.63 seconds (Sampling)
## Chain 3: 6613.17 seconds (Total)
## Chain 3:
##
## SAMPLING FOR MODEL 'e95c3961c8fbd2991e14129d68c8ae77' NOW (CHAIN 4).
## Chain 4:
## Chain 4: Gradient evaluation took 0.004156 seconds
## Chain 4: 1000 transitions using 10 leapfrog steps per transition would take 41.56 seconds.
## Chain 4: Adjust your expectations accordingly!
## Chain 4:
## Chain 4:
## Chain 4: Iteration: 1 / 5000 [ 0%] (Warmup)
## Chain 4: Iteration: 500 / 5000 [ 10%] (Warmup)
## Chain 4: Iteration: 1000 / 5000 [ 20%] (Warmup)
## Chain 4: Iteration: 1500 / 5000 [ 30%] (Warmup)
## Chain 4: Iteration: 2000 / 5000 [ 40%] (Warmup)
## Chain 4: Iteration: 2500 / 5000 [ 50%] (Warmup)
## Chain 4: Iteration: 2501 / 5000 [ 50%] (Sampling)
## Chain 4: Iteration: 3000 / 5000 [ 60%] (Sampling)
## Chain 4: Iteration: 3500 / 5000 [ 70%] (Sampling)
## Chain 4: Iteration: 4000 / 5000 [ 80%] (Sampling)
## Chain 4: Iteration: 4500 / 5000 [ 90%] (Sampling)
## Chain 4: Iteration: 5000 / 5000 [100%] (Sampling)
## Chain 4:
## Chain 4: Elapsed Time: 2935.41 seconds (Warm-up)
## Chain 4: 2432.18 seconds (Sampling)
## Chain 4: 5367.59 seconds (Total)
## Chain 4:
## Warning: There were 62 divergent transitions after warmup. Increasing adapt_delta above 0.8 may help. See
## http://mc-stan.org/misc/warnings.html#divergent-transitions-after-warmup
## Warning: Examine the pairs() plot to diagnose sampling problems
# モデル1 + 初期値と推移に大卒率が影響
koku.res3 <- brm(koku ~ year +
cs.c +
dai.c +
year:cs.c +
year:dai.c +
(1 + year|id) + (1 + cs.c * dai.c|sid),
data = koku.nrt.lt,
prior = NULL,
chains = 4,
iter = 5000,
warmup = 2500,
see = 1234
)
## Compiling the C++ model
## recompiling to avoid crashing R session
## Start sampling
##
## SAMPLING FOR MODEL 'e95c3961c8fbd2991e14129d68c8ae77' NOW (CHAIN 1).
## Chain 1:
## Chain 1: Gradient evaluation took 0.00856 seconds
## Chain 1: 1000 transitions using 10 leapfrog steps per transition would take 85.6 seconds.
## Chain 1: Adjust your expectations accordingly!
## Chain 1:
## Chain 1:
## Chain 1: Iteration: 1 / 5000 [ 0%] (Warmup)
## Chain 1: Iteration: 500 / 5000 [ 10%] (Warmup)
## Chain 1: Iteration: 1000 / 5000 [ 20%] (Warmup)
## Chain 1: Iteration: 1500 / 5000 [ 30%] (Warmup)
## Chain 1: Iteration: 2000 / 5000 [ 40%] (Warmup)
## Chain 1: Iteration: 2500 / 5000 [ 50%] (Warmup)
## Chain 1: Iteration: 2501 / 5000 [ 50%] (Sampling)
## Chain 1: Iteration: 3000 / 5000 [ 60%] (Sampling)
## Chain 1: Iteration: 3500 / 5000 [ 70%] (Sampling)
## Chain 1: Iteration: 4000 / 5000 [ 80%] (Sampling)
## Chain 1: Iteration: 4500 / 5000 [ 90%] (Sampling)
## Chain 1: Iteration: 5000 / 5000 [100%] (Sampling)
## Chain 1:
## Chain 1: Elapsed Time: 3356.17 seconds (Warm-up)
## Chain 1: 2493.17 seconds (Sampling)
## Chain 1: 5849.34 seconds (Total)
## Chain 1:
##
## SAMPLING FOR MODEL 'e95c3961c8fbd2991e14129d68c8ae77' NOW (CHAIN 2).
## Chain 2:
## Chain 2: Gradient evaluation took 0.004232 seconds
## Chain 2: 1000 transitions using 10 leapfrog steps per transition would take 42.32 seconds.
## Chain 2: Adjust your expectations accordingly!
## Chain 2:
## Chain 2:
## Chain 2: Iteration: 1 / 5000 [ 0%] (Warmup)
## Chain 2: Iteration: 500 / 5000 [ 10%] (Warmup)
## Chain 2: Iteration: 1000 / 5000 [ 20%] (Warmup)
## Chain 2: Iteration: 1500 / 5000 [ 30%] (Warmup)
## Chain 2: Iteration: 2000 / 5000 [ 40%] (Warmup)
## Chain 2: Iteration: 2500 / 5000 [ 50%] (Warmup)
## Chain 2: Iteration: 2501 / 5000 [ 50%] (Sampling)
## Chain 2: Iteration: 3000 / 5000 [ 60%] (Sampling)
## Chain 2: Iteration: 3500 / 5000 [ 70%] (Sampling)
## Chain 2: Iteration: 4000 / 5000 [ 80%] (Sampling)
## Chain 2: Iteration: 4500 / 5000 [ 90%] (Sampling)
## Chain 2: Iteration: 5000 / 5000 [100%] (Sampling)
## Chain 2:
## Chain 2: Elapsed Time: 3400.14 seconds (Warm-up)
## Chain 2: 2509.23 seconds (Sampling)
## Chain 2: 5909.36 seconds (Total)
## Chain 2:
##
## SAMPLING FOR MODEL 'e95c3961c8fbd2991e14129d68c8ae77' NOW (CHAIN 3).
## Chain 3:
## Chain 3: Gradient evaluation took 0.004244 seconds
## Chain 3: 1000 transitions using 10 leapfrog steps per transition would take 42.44 seconds.
## Chain 3: Adjust your expectations accordingly!
## Chain 3:
## Chain 3:
## Chain 3: Iteration: 1 / 5000 [ 0%] (Warmup)
## Chain 3: Iteration: 500 / 5000 [ 10%] (Warmup)
## Chain 3: Iteration: 1000 / 5000 [ 20%] (Warmup)
## Chain 3: Iteration: 1500 / 5000 [ 30%] (Warmup)
## Chain 3: Iteration: 2000 / 5000 [ 40%] (Warmup)
## Chain 3: Iteration: 2500 / 5000 [ 50%] (Warmup)
## Chain 3: Iteration: 2501 / 5000 [ 50%] (Sampling)
## Chain 3: Iteration: 3000 / 5000 [ 60%] (Sampling)
## Chain 3: Iteration: 3500 / 5000 [ 70%] (Sampling)
## Chain 3: Iteration: 4000 / 5000 [ 80%] (Sampling)
## Chain 3: Iteration: 4500 / 5000 [ 90%] (Sampling)
## Chain 3: Iteration: 5000 / 5000 [100%] (Sampling)
## Chain 3:
## Chain 3: Elapsed Time: 3223.92 seconds (Warm-up)
## Chain 3: 9868 seconds (Sampling)
## Chain 3: 13091.9 seconds (Total)
## Chain 3:
##
## SAMPLING FOR MODEL 'e95c3961c8fbd2991e14129d68c8ae77' NOW (CHAIN 4).
## Chain 4:
## Chain 4: Gradient evaluation took 0.004323 seconds
## Chain 4: 1000 transitions using 10 leapfrog steps per transition would take 43.23 seconds.
## Chain 4: Adjust your expectations accordingly!
## Chain 4:
## Chain 4:
## Chain 4: Iteration: 1 / 5000 [ 0%] (Warmup)
## Chain 4: Iteration: 500 / 5000 [ 10%] (Warmup)
## Chain 4: Iteration: 1000 / 5000 [ 20%] (Warmup)
## Chain 4: Iteration: 1500 / 5000 [ 30%] (Warmup)
## Chain 4: Iteration: 2000 / 5000 [ 40%] (Warmup)
## Chain 4: Iteration: 2500 / 5000 [ 50%] (Warmup)
## Chain 4: Iteration: 2501 / 5000 [ 50%] (Sampling)
## Chain 4: Iteration: 3000 / 5000 [ 60%] (Sampling)
## Chain 4: Iteration: 3500 / 5000 [ 70%] (Sampling)
## Chain 4: Iteration: 4000 / 5000 [ 80%] (Sampling)
## Chain 4: Iteration: 4500 / 5000 [ 90%] (Sampling)
## Chain 4: Iteration: 5000 / 5000 [100%] (Sampling)
## Chain 4:
## Chain 4: Elapsed Time: 3198.24 seconds (Warm-up)
## Chain 4: 2474.9 seconds (Sampling)
## Chain 4: 5673.14 seconds (Total)
## Chain 4:
## Warning: There were 214 divergent transitions after warmup. Increasing adapt_delta above 0.8 may help. See
## http://mc-stan.org/misc/warnings.html#divergent-transitions-after-warmup
## Warning: There were 2341 transitions after warmup that exceeded the maximum treedepth. Increase max_treedepth above 10. See
## http://mc-stan.org/misc/warnings.html#maximum-treedepth-exceeded
## Warning: Examine the pairs() plot to diagnose sampling problems
# 初期値に大卒率とクラスサイズの交互作用が影響
koku.res4 <- brm(koku ~ year +
cs.c +
dai.c +
dai.c:cs.c +
year:cs.c +
year:dai.c +
(1 + year|id) + (1 + cs.c * dai.c|sid),
data = koku.nrt.lt,
prior = NULL,
chains = 4,
iter = 5000,
warmup = 2500,
see = 1234
)
## Compiling the C++ model
## recompiling to avoid crashing R session
## Start sampling
##
## SAMPLING FOR MODEL 'e95c3961c8fbd2991e14129d68c8ae77' NOW (CHAIN 1).
## Chain 1:
## Chain 1: Gradient evaluation took 0.005086 seconds
## Chain 1: 1000 transitions using 10 leapfrog steps per transition would take 50.86 seconds.
## Chain 1: Adjust your expectations accordingly!
## Chain 1:
## Chain 1:
## Chain 1: Iteration: 1 / 5000 [ 0%] (Warmup)
## Chain 1: Iteration: 500 / 5000 [ 10%] (Warmup)
## Chain 1: Iteration: 1000 / 5000 [ 20%] (Warmup)
## Chain 1: Iteration: 1500 / 5000 [ 30%] (Warmup)
## Chain 1: Iteration: 2000 / 5000 [ 40%] (Warmup)
## Chain 1: Iteration: 2500 / 5000 [ 50%] (Warmup)
## Chain 1: Iteration: 2501 / 5000 [ 50%] (Sampling)
## Chain 1: Iteration: 3000 / 5000 [ 60%] (Sampling)
## Chain 1: Iteration: 3500 / 5000 [ 70%] (Sampling)
## Chain 1: Iteration: 4000 / 5000 [ 80%] (Sampling)
## Chain 1: Iteration: 4500 / 5000 [ 90%] (Sampling)
## Chain 1: Iteration: 5000 / 5000 [100%] (Sampling)
## Chain 1:
## Chain 1: Elapsed Time: 3649.34 seconds (Warm-up)
## Chain 1: 9673.2 seconds (Sampling)
## Chain 1: 13322.5 seconds (Total)
## Chain 1:
##
## SAMPLING FOR MODEL 'e95c3961c8fbd2991e14129d68c8ae77' NOW (CHAIN 2).
## Chain 2:
## Chain 2: Gradient evaluation took 0.004401 seconds
## Chain 2: 1000 transitions using 10 leapfrog steps per transition would take 44.01 seconds.
## Chain 2: Adjust your expectations accordingly!
## Chain 2:
## Chain 2:
## Chain 2: Iteration: 1 / 5000 [ 0%] (Warmup)
## Chain 2: Iteration: 500 / 5000 [ 10%] (Warmup)
## Chain 2: Iteration: 1000 / 5000 [ 20%] (Warmup)
## Chain 2: Iteration: 1500 / 5000 [ 30%] (Warmup)
## Chain 2: Iteration: 2000 / 5000 [ 40%] (Warmup)
## Chain 2: Iteration: 2500 / 5000 [ 50%] (Warmup)
## Chain 2: Iteration: 2501 / 5000 [ 50%] (Sampling)
## Chain 2: Iteration: 3000 / 5000 [ 60%] (Sampling)
## Chain 2: Iteration: 3500 / 5000 [ 70%] (Sampling)
## Chain 2: Iteration: 4000 / 5000 [ 80%] (Sampling)
## Chain 2: Iteration: 4500 / 5000 [ 90%] (Sampling)
## Chain 2: Iteration: 5000 / 5000 [100%] (Sampling)
## Chain 2:
## Chain 2: Elapsed Time: 3247.52 seconds (Warm-up)
## Chain 2: 9526.89 seconds (Sampling)
## Chain 2: 12774.4 seconds (Total)
## Chain 2:
##
## SAMPLING FOR MODEL 'e95c3961c8fbd2991e14129d68c8ae77' NOW (CHAIN 3).
## Chain 3:
## Chain 3: Gradient evaluation took 0.004162 seconds
## Chain 3: 1000 transitions using 10 leapfrog steps per transition would take 41.62 seconds.
## Chain 3: Adjust your expectations accordingly!
## Chain 3:
## Chain 3:
## Chain 3: Iteration: 1 / 5000 [ 0%] (Warmup)
## Chain 3: Iteration: 500 / 5000 [ 10%] (Warmup)
## Chain 3: Iteration: 1000 / 5000 [ 20%] (Warmup)
## Chain 3: Iteration: 1500 / 5000 [ 30%] (Warmup)
## Chain 3: Iteration: 2000 / 5000 [ 40%] (Warmup)
## Chain 3: Iteration: 2500 / 5000 [ 50%] (Warmup)
## Chain 3: Iteration: 2501 / 5000 [ 50%] (Sampling)
## Chain 3: Iteration: 3000 / 5000 [ 60%] (Sampling)
## Chain 3: Iteration: 3500 / 5000 [ 70%] (Sampling)
## Chain 3: Iteration: 4000 / 5000 [ 80%] (Sampling)
## Chain 3: Iteration: 4500 / 5000 [ 90%] (Sampling)
## Chain 3: Iteration: 5000 / 5000 [100%] (Sampling)
## Chain 3:
## Chain 3: Elapsed Time: 3403.45 seconds (Warm-up)
## Chain 3: 2556.46 seconds (Sampling)
## Chain 3: 5959.91 seconds (Total)
## Chain 3:
##
## SAMPLING FOR MODEL 'e95c3961c8fbd2991e14129d68c8ae77' NOW (CHAIN 4).
## Chain 4:
## Chain 4: Gradient evaluation took 0.004416 seconds
## Chain 4: 1000 transitions using 10 leapfrog steps per transition would take 44.16 seconds.
## Chain 4: Adjust your expectations accordingly!
## Chain 4:
## Chain 4:
## Chain 4: Iteration: 1 / 5000 [ 0%] (Warmup)
## Chain 4: Iteration: 500 / 5000 [ 10%] (Warmup)
## Chain 4: Iteration: 1000 / 5000 [ 20%] (Warmup)
## Chain 4: Iteration: 1500 / 5000 [ 30%] (Warmup)
## Chain 4: Iteration: 2000 / 5000 [ 40%] (Warmup)
## Chain 4: Iteration: 2500 / 5000 [ 50%] (Warmup)
## Chain 4: Iteration: 2501 / 5000 [ 50%] (Sampling)
## Chain 4: Iteration: 3000 / 5000 [ 60%] (Sampling)
## Chain 4: Iteration: 3500 / 5000 [ 70%] (Sampling)
## Chain 4: Iteration: 4000 / 5000 [ 80%] (Sampling)
## Chain 4: Iteration: 4500 / 5000 [ 90%] (Sampling)
## Chain 4: Iteration: 5000 / 5000 [100%] (Sampling)
## Chain 4:
## Chain 4: Elapsed Time: 3550.77 seconds (Warm-up)
## Chain 4: 2529.38 seconds (Sampling)
## Chain 4: 6080.16 seconds (Total)
## Chain 4:
## Warning: There were 752 divergent transitions after warmup. Increasing adapt_delta above 0.8 may help. See
## http://mc-stan.org/misc/warnings.html#divergent-transitions-after-warmup
## Warning: There were 4266 transitions after warmup that exceeded the maximum treedepth. Increase max_treedepth above 10. See
## http://mc-stan.org/misc/warnings.html#maximum-treedepth-exceeded
## Warning: Examine the pairs() plot to diagnose sampling problems
# 推移に大卒率とクラスサイズの交互作用が影響
koku.res5 <- brm(koku ~ year +
cs.c +
dai.c +
year:cs.c +
year:dai.c +
year:cs.c:dai.c +
(1 + year|id) + (1 + cs.c * dai.c|sid),
data = koku.nrt.lt,
prior = NULL,
chains = 4,
iter = 5000,
warmup = 2500,
see = 1234
)
## Compiling the C++ model
## recompiling to avoid crashing R session
## Start sampling
##
## SAMPLING FOR MODEL 'e95c3961c8fbd2991e14129d68c8ae77' NOW (CHAIN 1).
## Chain 1:
## Chain 1: Gradient evaluation took 0.004631 seconds
## Chain 1: 1000 transitions using 10 leapfrog steps per transition would take 46.31 seconds.
## Chain 1: Adjust your expectations accordingly!
## Chain 1:
## Chain 1:
## Chain 1: Iteration: 1 / 5000 [ 0%] (Warmup)
## Chain 1: Iteration: 500 / 5000 [ 10%] (Warmup)
## Chain 1: Iteration: 1000 / 5000 [ 20%] (Warmup)
## Chain 1: Iteration: 1500 / 5000 [ 30%] (Warmup)
## Chain 1: Iteration: 2000 / 5000 [ 40%] (Warmup)
## Chain 1: Iteration: 2500 / 5000 [ 50%] (Warmup)
## Chain 1: Iteration: 2501 / 5000 [ 50%] (Sampling)
## Chain 1: Iteration: 3000 / 5000 [ 60%] (Sampling)
## Chain 1: Iteration: 3500 / 5000 [ 70%] (Sampling)
## Chain 1: Iteration: 4000 / 5000 [ 80%] (Sampling)
## Chain 1: Iteration: 4500 / 5000 [ 90%] (Sampling)
## Chain 1: Iteration: 5000 / 5000 [100%] (Sampling)
## Chain 1:
## Chain 1: Elapsed Time: 3424.84 seconds (Warm-up)
## Chain 1: 9548.24 seconds (Sampling)
## Chain 1: 12973.1 seconds (Total)
## Chain 1:
##
## SAMPLING FOR MODEL 'e95c3961c8fbd2991e14129d68c8ae77' NOW (CHAIN 2).
## Chain 2:
## Chain 2: Gradient evaluation took 0.004151 seconds
## Chain 2: 1000 transitions using 10 leapfrog steps per transition would take 41.51 seconds.
## Chain 2: Adjust your expectations accordingly!
## Chain 2:
## Chain 2:
## Chain 2: Iteration: 1 / 5000 [ 0%] (Warmup)
## Chain 2: Iteration: 500 / 5000 [ 10%] (Warmup)
## Chain 2: Iteration: 1000 / 5000 [ 20%] (Warmup)
## Chain 2: Iteration: 1500 / 5000 [ 30%] (Warmup)
## Chain 2: Iteration: 2000 / 5000 [ 40%] (Warmup)
## Chain 2: Iteration: 2500 / 5000 [ 50%] (Warmup)
## Chain 2: Iteration: 2501 / 5000 [ 50%] (Sampling)
## Chain 2: Iteration: 3000 / 5000 [ 60%] (Sampling)
## Chain 2: Iteration: 3500 / 5000 [ 70%] (Sampling)
## Chain 2: Iteration: 4000 / 5000 [ 80%] (Sampling)
## Chain 2: Iteration: 4500 / 5000 [ 90%] (Sampling)
## Chain 2: Iteration: 5000 / 5000 [100%] (Sampling)
## Chain 2:
## Chain 2: Elapsed Time: 3473.97 seconds (Warm-up)
## Chain 2: 2552.15 seconds (Sampling)
## Chain 2: 6026.12 seconds (Total)
## Chain 2:
##
## SAMPLING FOR MODEL 'e95c3961c8fbd2991e14129d68c8ae77' NOW (CHAIN 3).
## Chain 3:
## Chain 3: Gradient evaluation took 0.004293 seconds
## Chain 3: 1000 transitions using 10 leapfrog steps per transition would take 42.93 seconds.
## Chain 3: Adjust your expectations accordingly!
## Chain 3:
## Chain 3:
## Chain 3: Iteration: 1 / 5000 [ 0%] (Warmup)
## Chain 3: Iteration: 500 / 5000 [ 10%] (Warmup)
## Chain 3: Iteration: 1000 / 5000 [ 20%] (Warmup)
## Chain 3: Iteration: 1500 / 5000 [ 30%] (Warmup)
## Chain 3: Iteration: 2000 / 5000 [ 40%] (Warmup)
## Chain 3: Iteration: 2500 / 5000 [ 50%] (Warmup)
## Chain 3: Iteration: 2501 / 5000 [ 50%] (Sampling)
## Chain 3: Iteration: 3000 / 5000 [ 60%] (Sampling)
## Chain 3: Iteration: 3500 / 5000 [ 70%] (Sampling)
## Chain 3: Iteration: 4000 / 5000 [ 80%] (Sampling)
## Chain 3: Iteration: 4500 / 5000 [ 90%] (Sampling)
## Chain 3: Iteration: 5000 / 5000 [100%] (Sampling)
## Chain 3:
## Chain 3: Elapsed Time: 3392.71 seconds (Warm-up)
## Chain 3: 5050.24 seconds (Sampling)
## Chain 3: 8442.95 seconds (Total)
## Chain 3:
##
## SAMPLING FOR MODEL 'e95c3961c8fbd2991e14129d68c8ae77' NOW (CHAIN 4).
## Chain 4:
## Chain 4: Gradient evaluation took 0.004326 seconds
## Chain 4: 1000 transitions using 10 leapfrog steps per transition would take 43.26 seconds.
## Chain 4: Adjust your expectations accordingly!
## Chain 4:
## Chain 4:
## Chain 4: Iteration: 1 / 5000 [ 0%] (Warmup)
## Chain 4: Iteration: 500 / 5000 [ 10%] (Warmup)
## Chain 4: Iteration: 1000 / 5000 [ 20%] (Warmup)
## Chain 4: Iteration: 1500 / 5000 [ 30%] (Warmup)
## Chain 4: Iteration: 2000 / 5000 [ 40%] (Warmup)
## Chain 4: Iteration: 2500 / 5000 [ 50%] (Warmup)
## Chain 4: Iteration: 2501 / 5000 [ 50%] (Sampling)
## Chain 4: Iteration: 3000 / 5000 [ 60%] (Sampling)
## Chain 4: Iteration: 3500 / 5000 [ 70%] (Sampling)
## Chain 4: Iteration: 4000 / 5000 [ 80%] (Sampling)
## Chain 4: Iteration: 4500 / 5000 [ 90%] (Sampling)
## Chain 4: Iteration: 5000 / 5000 [100%] (Sampling)
## Chain 4:
## Chain 4: Elapsed Time: 3386.78 seconds (Warm-up)
## Chain 4: 2507.78 seconds (Sampling)
## Chain 4: 5894.57 seconds (Total)
## Chain 4:
## Warning: There were 415 divergent transitions after warmup. Increasing adapt_delta above 0.8 may help. See
## http://mc-stan.org/misc/warnings.html#divergent-transitions-after-warmup
## Warning: There were 2241 transitions after warmup that exceeded the maximum treedepth. Increase max_treedepth above 10. See
## http://mc-stan.org/misc/warnings.html#maximum-treedepth-exceeded
## Warning: Examine the pairs() plot to diagnose sampling problems
# 初期値と推移に大卒率とクラスサイズの交互作用が影響
koku.res6 <- brm(koku ~ year +
dai.c +
cs.c +
dai.c:cs.c +
year:cs.c +
year:dai.c +
year:cs.c:dai.c +
(1 + year|id) + (1 + cs.c * dai.c|sid),
data = koku.nrt.lt,
prior = NULL,
chains = 4,
iter = 5000,
warmup = 2500,
see = 1234
)
## Compiling the C++ model
## recompiling to avoid crashing R session
## Start sampling
##
## SAMPLING FOR MODEL 'e95c3961c8fbd2991e14129d68c8ae77' NOW (CHAIN 1).
## Chain 1:
## Chain 1: Gradient evaluation took 0.006233 seconds
## Chain 1: 1000 transitions using 10 leapfrog steps per transition would take 62.33 seconds.
## Chain 1: Adjust your expectations accordingly!
## Chain 1:
## Chain 1:
## Chain 1: Iteration: 1 / 5000 [ 0%] (Warmup)
## Chain 1: Iteration: 500 / 5000 [ 10%] (Warmup)
## Chain 1: Iteration: 1000 / 5000 [ 20%] (Warmup)
## Chain 1: Iteration: 1500 / 5000 [ 30%] (Warmup)
## Chain 1: Iteration: 2000 / 5000 [ 40%] (Warmup)
## Chain 1: Iteration: 2500 / 5000 [ 50%] (Warmup)
## Chain 1: Iteration: 2501 / 5000 [ 50%] (Sampling)
## Chain 1: Iteration: 3000 / 5000 [ 60%] (Sampling)
## Chain 1: Iteration: 3500 / 5000 [ 70%] (Sampling)
## Chain 1: Iteration: 4000 / 5000 [ 80%] (Sampling)
## Chain 1: Iteration: 4500 / 5000 [ 90%] (Sampling)
## Chain 1: Iteration: 5000 / 5000 [100%] (Sampling)
## Chain 1:
## Chain 1: Elapsed Time: 4183.83 seconds (Warm-up)
## Chain 1: 2616.81 seconds (Sampling)
## Chain 1: 6800.65 seconds (Total)
## Chain 1:
##
## SAMPLING FOR MODEL 'e95c3961c8fbd2991e14129d68c8ae77' NOW (CHAIN 2).
## Chain 2:
## Chain 2: Gradient evaluation took 0.004334 seconds
## Chain 2: 1000 transitions using 10 leapfrog steps per transition would take 43.34 seconds.
## Chain 2: Adjust your expectations accordingly!
## Chain 2:
## Chain 2:
## Chain 2: Iteration: 1 / 5000 [ 0%] (Warmup)
## Chain 2: Iteration: 500 / 5000 [ 10%] (Warmup)
## Chain 2: Iteration: 1000 / 5000 [ 20%] (Warmup)
## Chain 2: Iteration: 1500 / 5000 [ 30%] (Warmup)
## Chain 2: Iteration: 2000 / 5000 [ 40%] (Warmup)
## Chain 2: Iteration: 2500 / 5000 [ 50%] (Warmup)
## Chain 2: Iteration: 2501 / 5000 [ 50%] (Sampling)
## Chain 2: Iteration: 3000 / 5000 [ 60%] (Sampling)
## Chain 2: Iteration: 3500 / 5000 [ 70%] (Sampling)
## Chain 2: Iteration: 4000 / 5000 [ 80%] (Sampling)
## Chain 2: Iteration: 4500 / 5000 [ 90%] (Sampling)
## Chain 2: Iteration: 5000 / 5000 [100%] (Sampling)
## Chain 2:
## Chain 2: Elapsed Time: 3725.76 seconds (Warm-up)
## Chain 2: 2606.24 seconds (Sampling)
## Chain 2: 6332 seconds (Total)
## Chain 2:
##
## SAMPLING FOR MODEL 'e95c3961c8fbd2991e14129d68c8ae77' NOW (CHAIN 3).
## Chain 3:
## Chain 3: Gradient evaluation took 0.004272 seconds
## Chain 3: 1000 transitions using 10 leapfrog steps per transition would take 42.72 seconds.
## Chain 3: Adjust your expectations accordingly!
## Chain 3:
## Chain 3:
## Chain 3: Iteration: 1 / 5000 [ 0%] (Warmup)
## Chain 3: Iteration: 500 / 5000 [ 10%] (Warmup)
## Chain 3: Iteration: 1000 / 5000 [ 20%] (Warmup)
## Chain 3: Iteration: 1500 / 5000 [ 30%] (Warmup)
## Chain 3: Iteration: 2000 / 5000 [ 40%] (Warmup)
## Chain 3: Iteration: 2500 / 5000 [ 50%] (Warmup)
## Chain 3: Iteration: 2501 / 5000 [ 50%] (Sampling)
## Chain 3: Iteration: 3000 / 5000 [ 60%] (Sampling)
## Chain 3: Iteration: 3500 / 5000 [ 70%] (Sampling)
## Chain 3: Iteration: 4000 / 5000 [ 80%] (Sampling)
## Chain 3: Iteration: 4500 / 5000 [ 90%] (Sampling)
## Chain 3: Iteration: 5000 / 5000 [100%] (Sampling)
## Chain 3:
## Chain 3: Elapsed Time: 3355.07 seconds (Warm-up)
## Chain 3: 2602.38 seconds (Sampling)
## Chain 3: 5957.44 seconds (Total)
## Chain 3:
##
## SAMPLING FOR MODEL 'e95c3961c8fbd2991e14129d68c8ae77' NOW (CHAIN 4).
## Chain 4:
## Chain 4: Gradient evaluation took 0.004613 seconds
## Chain 4: 1000 transitions using 10 leapfrog steps per transition would take 46.13 seconds.
## Chain 4: Adjust your expectations accordingly!
## Chain 4:
## Chain 4:
## Chain 4: Iteration: 1 / 5000 [ 0%] (Warmup)
## Chain 4: Iteration: 500 / 5000 [ 10%] (Warmup)
## Chain 4: Iteration: 1000 / 5000 [ 20%] (Warmup)
## Chain 4: Iteration: 1500 / 5000 [ 30%] (Warmup)
## Chain 4: Iteration: 2000 / 5000 [ 40%] (Warmup)
## Chain 4: Iteration: 2500 / 5000 [ 50%] (Warmup)
## Chain 4: Iteration: 2501 / 5000 [ 50%] (Sampling)
## Chain 4: Iteration: 3000 / 5000 [ 60%] (Sampling)
## Chain 4: Iteration: 3500 / 5000 [ 70%] (Sampling)
## Chain 4: Iteration: 4000 / 5000 [ 80%] (Sampling)
## Chain 4: Iteration: 4500 / 5000 [ 90%] (Sampling)
## Chain 4: Iteration: 5000 / 5000 [100%] (Sampling)
## Chain 4:
## Chain 4: Elapsed Time: 3579.41 seconds (Warm-up)
## Chain 4: 5007.04 seconds (Sampling)
## Chain 4: 8586.44 seconds (Total)
## Chain 4:
## Warning: There were 187 divergent transitions after warmup. Increasing adapt_delta above 0.8 may help. See
## http://mc-stan.org/misc/warnings.html#divergent-transitions-after-warmup
## Warning: Examine the pairs() plot to diagnose sampling problems
print(koku.res1, digits = 3)
## Warning: The model has not converged (some Rhats are > 1.1). Do not analyse the results!
## We recommend running more iterations and/or setting stronger priors.
## Warning: There were 2751 divergent transitions after warmup. Increasing adapt_delta above 0.8 may help.
## See http://mc-stan.org/misc/warnings.html#divergent-transitions-after-warmup
## Family: gaussian
## Links: mu = identity; sigma = identity
## Formula: koku ~ year + cs.c + year:cs.c + (1 + year | id) + (1 + cs.c | sid)
## Data: koku.nrt.lt (Number of observations: 20718)
## Samples: 4 chains, each with iter = 5000; warmup = 2500; thin = 1;
## total post-warmup samples = 10000
##
## Group-Level Effects:
## ~id (Number of levels: 3453)
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhat
## sd(Intercept) 6.379 1.711 2.822 7.553 2 8.778
## sd(year) 1.148 0.509 0.808 2.036 2 24.094
## cor(Intercept,year) -0.087 0.118 -0.318 0.034 2 4.327
##
## ~sid (Number of levels: 102)
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhat
## sd(Intercept) 1.856 0.430 1.153 2.548 3 1.932
## sd(cs.c) 0.072 0.055 0.005 0.206 15 1.099
## cor(Intercept,cs.c) 0.014 0.521 -0.738 0.931 5 1.327
##
## Population-Level Effects:
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhat
## Intercept 53.717 0.246 53.343 54.241 7 1.178
## year -0.212 0.027 -0.253 -0.157 8 1.158
## cs.c -0.088 0.038 -0.151 -0.005 11 1.124
## year:cs.c -0.007 0.005 -0.017 0.001 6 1.244
##
## Family Specific Parameters:
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhat
## sigma 4.695 0.777 4.200 6.076 2 33.454
##
## Samples were drawn using sampling(NUTS). For each parameter, Eff.Sample
## is a crude measure of effective sample size, and Rhat is the potential
## scale reduction factor on split chains (at convergence, Rhat = 1).
print(koku.res2, digits = 3)
## Warning: There were 62 divergent transitions after warmup. Increasing adapt_delta above 0.8 may help.
## See http://mc-stan.org/misc/warnings.html#divergent-transitions-after-warmup
## Family: gaussian
## Links: mu = identity; sigma = identity
## Formula: koku ~ year + cs.c + dai.c + year:cs.c + (1 + year | id) + (1 + cs.c * dai.c | sid)
## Data: koku.nrt.lt (Number of observations: 20718)
## Samples: 4 chains, each with iter = 5000; warmup = 2500; thin = 1;
## total post-warmup samples = 10000
##
## Group-Level Effects:
## ~id (Number of levels: 3453)
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhat
## sd(Intercept) 7.349 0.107 7.140 7.562 4351 1.001
## sd(year) 0.855 0.026 0.804 0.905 3359 1.001
## cor(Intercept,year) -0.022 0.030 -0.080 0.039 7128 1.000
##
## ~sid (Number of levels: 102)
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample
## sd(Intercept) 1.673 0.291 1.081 2.223 1422
## sd(cs.c) 0.090 0.065 0.004 0.239 1212
## sd(dai.c) 7.521 6.323 0.321 23.931 4685
## sd(cs.c:dai.c) 2.511 1.870 0.106 7.021 2032
## cor(Intercept,cs.c) 0.083 0.379 -0.672 0.789 4851
## cor(Intercept,dai.c) -0.039 0.434 -0.807 0.780 11347
## cor(cs.c,dai.c) -0.046 0.446 -0.825 0.793 10731
## cor(Intercept,cs.c:dai.c) -0.228 0.417 -0.862 0.679 5112
## cor(cs.c,cs.c:dai.c) 0.001 0.438 -0.795 0.792 7239
## cor(dai.c,cs.c:dai.c) -0.013 0.447 -0.809 0.803 7737
## Rhat
## sd(Intercept) 1.002
## sd(cs.c) 1.001
## sd(dai.c) 1.001
## sd(cs.c:dai.c) 1.002
## cor(Intercept,cs.c) 1.000
## cor(Intercept,dai.c) 1.000
## cor(cs.c,dai.c) 1.001
## cor(Intercept,cs.c:dai.c) 1.000
## cor(cs.c,cs.c:dai.c) 1.000
## cor(dai.c,cs.c:dai.c) 1.000
##
## Population-Level Effects:
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhat
## Intercept 53.787 0.246 53.304 54.276 5962 1.000
## year -0.205 0.026 -0.258 -0.153 14065 1.000
## cs.c -0.078 0.045 -0.167 0.010 6158 1.003
## dai.c -1.697 9.377 -20.071 16.642 5634 1.001
## year:cs.c -0.009 0.004 -0.017 0.000 14379 1.000
##
## Family Specific Parameters:
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhat
## sigma 4.247 0.025 4.197 4.297 5727 1.000
##
## Samples were drawn using sampling(NUTS). For each parameter, Eff.Sample
## is a crude measure of effective sample size, and Rhat is the potential
## scale reduction factor on split chains (at convergence, Rhat = 1).
print(koku.res3, digits = 3)
## Warning: There were 214 divergent transitions after warmup. Increasing adapt_delta above 0.8 may help.
## See http://mc-stan.org/misc/warnings.html#divergent-transitions-after-warmup
## Family: gaussian
## Links: mu = identity; sigma = identity
## Formula: koku ~ year + cs.c + dai.c + year:cs.c + year:dai.c + (1 + year | id) + (1 + cs.c * dai.c | sid)
## Data: koku.nrt.lt (Number of observations: 20718)
## Samples: 4 chains, each with iter = 5000; warmup = 2500; thin = 1;
## total post-warmup samples = 10000
##
## Group-Level Effects:
## ~id (Number of levels: 3453)
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhat
## sd(Intercept) 7.347 0.107 7.139 7.560 4147 1.000
## sd(year) 0.851 0.026 0.800 0.902 3360 1.000
## cor(Intercept,year) -0.019 0.031 -0.079 0.042 7375 1.000
##
## ~sid (Number of levels: 102)
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample
## sd(Intercept) 1.676 0.284 1.106 2.227 1922
## sd(cs.c) 0.089 0.063 0.004 0.233 1666
## sd(dai.c) 7.636 6.456 0.300 24.084 5085
## sd(cs.c:dai.c) 2.473 1.840 0.103 6.779 2708
## cor(Intercept,cs.c) 0.096 0.382 -0.685 0.797 4579
## cor(Intercept,dai.c) -0.041 0.436 -0.815 0.782 11481
## cor(cs.c,dai.c) -0.043 0.447 -0.825 0.796 10899
## cor(Intercept,cs.c:dai.c) -0.224 0.416 -0.865 0.680 6754
## cor(cs.c,cs.c:dai.c) -0.013 0.447 -0.826 0.808 8273
## cor(dai.c,cs.c:dai.c) -0.006 0.448 -0.819 0.804 8620
## Rhat
## sd(Intercept) 1.002
## sd(cs.c) 1.001
## sd(dai.c) 1.000
## sd(cs.c:dai.c) 1.000
## cor(Intercept,cs.c) 1.001
## cor(Intercept,dai.c) 1.000
## cor(cs.c,dai.c) 1.000
## cor(Intercept,cs.c:dai.c) 1.000
## cor(cs.c,cs.c:dai.c) 1.001
## cor(dai.c,cs.c:dai.c) 1.000
##
## Population-Level Effects:
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhat
## Intercept 53.817 0.244 53.328 54.289 8044 1.000
## year -0.224 0.027 -0.277 -0.170 12300 1.000
## cs.c -0.069 0.046 -0.160 0.021 6970 1.000
## dai.c -5.821 9.634 -24.617 12.760 7877 1.000
## year:cs.c -0.015 0.005 -0.024 -0.006 13730 1.000
## year:dai.c 2.722 0.808 1.127 4.282 13347 1.000
##
## Family Specific Parameters:
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhat
## sigma 4.247 0.025 4.198 4.297 5063 1.000
##
## Samples were drawn using sampling(NUTS). For each parameter, Eff.Sample
## is a crude measure of effective sample size, and Rhat is the potential
## scale reduction factor on split chains (at convergence, Rhat = 1).
print(koku.res4, digits = 3)
## Warning: There were 752 divergent transitions after warmup. Increasing adapt_delta above 0.8 may help.
## See http://mc-stan.org/misc/warnings.html#divergent-transitions-after-warmup
## Family: gaussian
## Links: mu = identity; sigma = identity
## Formula: koku ~ year + cs.c + dai.c + dai.c:cs.c + year:cs.c + year:dai.c + (1 + year | id) + (1 + cs.c * dai.c | sid)
## Data: koku.nrt.lt (Number of observations: 20718)
## Samples: 4 chains, each with iter = 5000; warmup = 2500; thin = 1;
## total post-warmup samples = 10000
##
## Group-Level Effects:
## ~id (Number of levels: 3453)
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhat
## sd(Intercept) 7.352 0.106 7.143 7.564 4292 1.000
## sd(year) 0.852 0.026 0.800 0.901 3033 1.001
## cor(Intercept,year) -0.020 0.031 -0.080 0.041 7028 1.000
##
## ~sid (Number of levels: 102)
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample
## sd(Intercept) 1.675 0.293 1.065 2.242 1628
## sd(cs.c) 0.092 0.064 0.004 0.233 1503
## sd(dai.c) 7.685 6.444 0.274 24.122 5374
## sd(cs.c:dai.c) 2.599 1.875 0.114 6.918 2382
## cor(Intercept,cs.c) 0.099 0.380 -0.656 0.797 4124
## cor(Intercept,dai.c) -0.050 0.433 -0.817 0.777 9510
## cor(cs.c,dai.c) -0.046 0.448 -0.830 0.798 9915
## cor(Intercept,cs.c:dai.c) -0.224 0.415 -0.866 0.668 5305
## cor(cs.c,cs.c:dai.c) -0.007 0.439 -0.803 0.804 7164
## cor(dai.c,cs.c:dai.c) -0.012 0.448 -0.811 0.807 6691
## Rhat
## sd(Intercept) 1.002
## sd(cs.c) 1.001
## sd(dai.c) 1.000
## sd(cs.c:dai.c) 1.001
## cor(Intercept,cs.c) 1.000
## cor(Intercept,dai.c) 1.000
## cor(cs.c,dai.c) 1.000
## cor(Intercept,cs.c:dai.c) 1.000
## cor(cs.c,cs.c:dai.c) 1.000
## cor(dai.c,cs.c:dai.c) 1.000
##
## Population-Level Effects:
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhat
## Intercept 53.790 0.269 53.265 54.325 6839 1.001
## year -0.222 0.027 -0.274 -0.170 10863 1.000
## cs.c -0.067 0.046 -0.158 0.021 6379 1.000
## dai.c -6.632 10.617 -28.023 13.824 6561 1.001
## cs.c:dai.c 0.387 1.709 -2.929 3.819 7047 1.000
## year:cs.c -0.015 0.005 -0.024 -0.006 11354 1.000
## year:dai.c 2.717 0.818 1.120 4.323 10599 1.000
##
## Family Specific Parameters:
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhat
## sigma 4.247 0.025 4.197 4.297 4899 1.001
##
## Samples were drawn using sampling(NUTS). For each parameter, Eff.Sample
## is a crude measure of effective sample size, and Rhat is the potential
## scale reduction factor on split chains (at convergence, Rhat = 1).
print(koku.res5, digits = 3)
## Warning: There were 415 divergent transitions after warmup. Increasing adapt_delta above 0.8 may help.
## See http://mc-stan.org/misc/warnings.html#divergent-transitions-after-warmup
## Family: gaussian
## Links: mu = identity; sigma = identity
## Formula: koku ~ year + cs.c + dai.c + year:cs.c + year:dai.c + year:cs.c:dai.c + (1 + year | id) + (1 + cs.c * dai.c | sid)
## Data: koku.nrt.lt (Number of observations: 20718)
## Samples: 4 chains, each with iter = 5000; warmup = 2500; thin = 1;
## total post-warmup samples = 10000
##
## Group-Level Effects:
## ~id (Number of levels: 3453)
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhat
## sd(Intercept) 7.350 0.106 7.147 7.562 4110 1.001
## sd(year) 0.852 0.026 0.802 0.901 3455 1.000
## cor(Intercept,year) -0.020 0.030 -0.078 0.040 7591 1.000
##
## ~sid (Number of levels: 102)
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample
## sd(Intercept) 1.672 0.278 1.116 2.216 1879
## sd(cs.c) 0.089 0.063 0.003 0.230 1544
## sd(dai.c) 7.740 6.493 0.276 24.358 5567
## sd(cs.c:dai.c) 2.510 1.897 0.091 7.055 1658
## cor(Intercept,cs.c) 0.099 0.381 -0.662 0.810 4609
## cor(Intercept,dai.c) -0.036 0.432 -0.812 0.778 8512
## cor(cs.c,dai.c) -0.053 0.447 -0.839 0.797 10392
## cor(Intercept,cs.c:dai.c) -0.223 0.414 -0.856 0.673 5893
## cor(cs.c,cs.c:dai.c) -0.018 0.439 -0.814 0.796 7526
## cor(dai.c,cs.c:dai.c) -0.012 0.452 -0.835 0.812 6714
## Rhat
## sd(Intercept) 1.003
## sd(cs.c) 1.001
## sd(dai.c) 1.001
## sd(cs.c:dai.c) 1.003
## cor(Intercept,cs.c) 1.001
## cor(Intercept,dai.c) 1.000
## cor(cs.c,dai.c) 1.000
## cor(Intercept,cs.c:dai.c) 1.001
## cor(cs.c,cs.c:dai.c) 1.000
## cor(dai.c,cs.c:dai.c) 1.000
##
## Population-Level Effects:
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhat
## Intercept 53.814 0.247 53.332 54.297 6513 1.000
## year -0.228 0.029 -0.283 -0.171 8745 1.000
## cs.c -0.067 0.045 -0.158 0.021 7051 1.000
## dai.c -5.687 9.593 -24.736 12.778 7624 1.000
## year:cs.c -0.015 0.005 -0.024 -0.006 11718 1.000
## year:dai.c 2.301 1.051 0.217 4.346 10211 1.001
## year:cs.c:dai.c 0.100 0.161 -0.218 0.415 10034 1.000
##
## Family Specific Parameters:
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhat
## sigma 4.246 0.026 4.196 4.297 4579 1.000
##
## Samples were drawn using sampling(NUTS). For each parameter, Eff.Sample
## is a crude measure of effective sample size, and Rhat is the potential
## scale reduction factor on split chains (at convergence, Rhat = 1).
print(koku.res6, digits = 3)
## Warning: There were 187 divergent transitions after warmup. Increasing adapt_delta above 0.8 may help.
## See http://mc-stan.org/misc/warnings.html#divergent-transitions-after-warmup
## Family: gaussian
## Links: mu = identity; sigma = identity
## Formula: koku ~ year + dai.c + cs.c + dai.c:cs.c + year:cs.c + year:dai.c + year:cs.c:dai.c + (1 + year | id) + (1 + cs.c * dai.c | sid)
## Data: koku.nrt.lt (Number of observations: 20718)
## Samples: 4 chains, each with iter = 5000; warmup = 2500; thin = 1;
## total post-warmup samples = 10000
##
## Group-Level Effects:
## ~id (Number of levels: 3453)
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhat
## sd(Intercept) 7.351 0.106 7.148 7.559 4287 1.000
## sd(year) 0.852 0.026 0.801 0.902 3361 1.002
## cor(Intercept,year) -0.020 0.030 -0.078 0.040 6580 1.000
##
## ~sid (Number of levels: 102)
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample
## sd(Intercept) 1.681 0.285 1.111 2.238 1395
## sd(cs.c) 0.087 0.062 0.003 0.227 1206
## sd(dai.c) 7.651 6.316 0.302 23.567 4676
## sd(cs.c:dai.c) 2.642 1.941 0.107 7.131 1796
## cor(Intercept,cs.c) 0.098 0.382 -0.672 0.794 4363
## cor(Intercept,dai.c) -0.049 0.430 -0.807 0.774 9417
## cor(cs.c,dai.c) -0.049 0.443 -0.823 0.775 10805
## cor(Intercept,cs.c:dai.c) -0.219 0.414 -0.856 0.682 5354
## cor(cs.c,cs.c:dai.c) -0.009 0.437 -0.796 0.796 7119
## cor(dai.c,cs.c:dai.c) -0.005 0.444 -0.798 0.803 6988
## Rhat
## sd(Intercept) 1.005
## sd(cs.c) 1.003
## sd(dai.c) 1.001
## sd(cs.c:dai.c) 1.002
## cor(Intercept,cs.c) 1.001
## cor(Intercept,dai.c) 1.000
## cor(cs.c,dai.c) 1.000
## cor(Intercept,cs.c:dai.c) 1.000
## cor(cs.c,cs.c:dai.c) 1.000
## cor(dai.c,cs.c:dai.c) 1.001
##
## Population-Level Effects:
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhat
## Intercept 53.803 0.270 53.278 54.337 6355 1.000
## year -0.229 0.028 -0.283 -0.173 11823 1.000
## dai.c -6.483 10.747 -27.710 14.981 5493 1.000
## cs.c -0.065 0.046 -0.156 0.024 5410 1.000
## dai.c:cs.c 0.236 1.792 -3.300 3.720 4992 1.000
## year:cs.c -0.015 0.005 -0.024 -0.005 12588 1.000
## year:dai.c 2.327 1.086 0.211 4.448 12636 1.000
## year:dai.c:cs.c 0.099 0.165 -0.221 0.424 11296 1.000
##
## Family Specific Parameters:
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhat
## sigma 4.247 0.026 4.196 4.298 5082 1.001
##
## Samples were drawn using sampling(NUTS). For each parameter, Eff.Sample
## is a crude measure of effective sample size, and Rhat is the potential
## scale reduction factor on split chains (at convergence, Rhat = 1).
print(koku.res1, digits = 2)
## Warning: The model has not converged (some Rhats are > 1.1). Do not analyse the results!
## We recommend running more iterations and/or setting stronger priors.
## Warning: There were 2751 divergent transitions after warmup. Increasing adapt_delta above 0.8 may help.
## See http://mc-stan.org/misc/warnings.html#divergent-transitions-after-warmup
## Family: gaussian
## Links: mu = identity; sigma = identity
## Formula: koku ~ year + cs.c + year:cs.c + (1 + year | id) + (1 + cs.c | sid)
## Data: koku.nrt.lt (Number of observations: 20718)
## Samples: 4 chains, each with iter = 5000; warmup = 2500; thin = 1;
## total post-warmup samples = 10000
##
## Group-Level Effects:
## ~id (Number of levels: 3453)
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhat
## sd(Intercept) 6.38 1.71 2.82 7.55 2 8.78
## sd(year) 1.15 0.51 0.81 2.04 2 24.09
## cor(Intercept,year) -0.09 0.12 -0.32 0.03 2 4.33
##
## ~sid (Number of levels: 102)
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhat
## sd(Intercept) 1.86 0.43 1.15 2.55 3 1.93
## sd(cs.c) 0.07 0.06 0.00 0.21 15 1.10
## cor(Intercept,cs.c) 0.01 0.52 -0.74 0.93 5 1.33
##
## Population-Level Effects:
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhat
## Intercept 53.72 0.25 53.34 54.24 7 1.18
## year -0.21 0.03 -0.25 -0.16 8 1.16
## cs.c -0.09 0.04 -0.15 -0.00 11 1.12
## year:cs.c -0.01 0.00 -0.02 0.00 6 1.24
##
## Family Specific Parameters:
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhat
## sigma 4.69 0.78 4.20 6.08 2 33.45
##
## Samples were drawn using sampling(NUTS). For each parameter, Eff.Sample
## is a crude measure of effective sample size, and Rhat is the potential
## scale reduction factor on split chains (at convergence, Rhat = 1).
print(koku.res2, digits = 2)
## Warning: There were 62 divergent transitions after warmup. Increasing adapt_delta above 0.8 may help.
## See http://mc-stan.org/misc/warnings.html#divergent-transitions-after-warmup
## Family: gaussian
## Links: mu = identity; sigma = identity
## Formula: koku ~ year + cs.c + dai.c + year:cs.c + (1 + year | id) + (1 + cs.c * dai.c | sid)
## Data: koku.nrt.lt (Number of observations: 20718)
## Samples: 4 chains, each with iter = 5000; warmup = 2500; thin = 1;
## total post-warmup samples = 10000
##
## Group-Level Effects:
## ~id (Number of levels: 3453)
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhat
## sd(Intercept) 7.35 0.11 7.14 7.56 4351 1.00
## sd(year) 0.86 0.03 0.80 0.91 3359 1.00
## cor(Intercept,year) -0.02 0.03 -0.08 0.04 7128 1.00
##
## ~sid (Number of levels: 102)
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample
## sd(Intercept) 1.67 0.29 1.08 2.22 1422
## sd(cs.c) 0.09 0.06 0.00 0.24 1212
## sd(dai.c) 7.52 6.32 0.32 23.93 4685
## sd(cs.c:dai.c) 2.51 1.87 0.11 7.02 2032
## cor(Intercept,cs.c) 0.08 0.38 -0.67 0.79 4851
## cor(Intercept,dai.c) -0.04 0.43 -0.81 0.78 11347
## cor(cs.c,dai.c) -0.05 0.45 -0.83 0.79 10731
## cor(Intercept,cs.c:dai.c) -0.23 0.42 -0.86 0.68 5112
## cor(cs.c,cs.c:dai.c) 0.00 0.44 -0.80 0.79 7239
## cor(dai.c,cs.c:dai.c) -0.01 0.45 -0.81 0.80 7737
## Rhat
## sd(Intercept) 1.00
## sd(cs.c) 1.00
## sd(dai.c) 1.00
## sd(cs.c:dai.c) 1.00
## cor(Intercept,cs.c) 1.00
## cor(Intercept,dai.c) 1.00
## cor(cs.c,dai.c) 1.00
## cor(Intercept,cs.c:dai.c) 1.00
## cor(cs.c,cs.c:dai.c) 1.00
## cor(dai.c,cs.c:dai.c) 1.00
##
## Population-Level Effects:
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhat
## Intercept 53.79 0.25 53.30 54.28 5962 1.00
## year -0.21 0.03 -0.26 -0.15 14065 1.00
## cs.c -0.08 0.05 -0.17 0.01 6158 1.00
## dai.c -1.70 9.38 -20.07 16.64 5634 1.00
## year:cs.c -0.01 0.00 -0.02 0.00 14379 1.00
##
## Family Specific Parameters:
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhat
## sigma 4.25 0.03 4.20 4.30 5727 1.00
##
## Samples were drawn using sampling(NUTS). For each parameter, Eff.Sample
## is a crude measure of effective sample size, and Rhat is the potential
## scale reduction factor on split chains (at convergence, Rhat = 1).
print(koku.res3, digits = 2)
## Warning: There were 214 divergent transitions after warmup. Increasing adapt_delta above 0.8 may help.
## See http://mc-stan.org/misc/warnings.html#divergent-transitions-after-warmup
## Family: gaussian
## Links: mu = identity; sigma = identity
## Formula: koku ~ year + cs.c + dai.c + year:cs.c + year:dai.c + (1 + year | id) + (1 + cs.c * dai.c | sid)
## Data: koku.nrt.lt (Number of observations: 20718)
## Samples: 4 chains, each with iter = 5000; warmup = 2500; thin = 1;
## total post-warmup samples = 10000
##
## Group-Level Effects:
## ~id (Number of levels: 3453)
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhat
## sd(Intercept) 7.35 0.11 7.14 7.56 4147 1.00
## sd(year) 0.85 0.03 0.80 0.90 3360 1.00
## cor(Intercept,year) -0.02 0.03 -0.08 0.04 7375 1.00
##
## ~sid (Number of levels: 102)
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample
## sd(Intercept) 1.68 0.28 1.11 2.23 1922
## sd(cs.c) 0.09 0.06 0.00 0.23 1666
## sd(dai.c) 7.64 6.46 0.30 24.08 5085
## sd(cs.c:dai.c) 2.47 1.84 0.10 6.78 2708
## cor(Intercept,cs.c) 0.10 0.38 -0.68 0.80 4579
## cor(Intercept,dai.c) -0.04 0.44 -0.81 0.78 11481
## cor(cs.c,dai.c) -0.04 0.45 -0.82 0.80 10899
## cor(Intercept,cs.c:dai.c) -0.22 0.42 -0.86 0.68 6754
## cor(cs.c,cs.c:dai.c) -0.01 0.45 -0.83 0.81 8273
## cor(dai.c,cs.c:dai.c) -0.01 0.45 -0.82 0.80 8620
## Rhat
## sd(Intercept) 1.00
## sd(cs.c) 1.00
## sd(dai.c) 1.00
## sd(cs.c:dai.c) 1.00
## cor(Intercept,cs.c) 1.00
## cor(Intercept,dai.c) 1.00
## cor(cs.c,dai.c) 1.00
## cor(Intercept,cs.c:dai.c) 1.00
## cor(cs.c,cs.c:dai.c) 1.00
## cor(dai.c,cs.c:dai.c) 1.00
##
## Population-Level Effects:
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhat
## Intercept 53.82 0.24 53.33 54.29 8044 1.00
## year -0.22 0.03 -0.28 -0.17 12300 1.00
## cs.c -0.07 0.05 -0.16 0.02 6970 1.00
## dai.c -5.82 9.63 -24.62 12.76 7877 1.00
## year:cs.c -0.02 0.00 -0.02 -0.01 13730 1.00
## year:dai.c 2.72 0.81 1.13 4.28 13347 1.00
##
## Family Specific Parameters:
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhat
## sigma 4.25 0.03 4.20 4.30 5063 1.00
##
## Samples were drawn using sampling(NUTS). For each parameter, Eff.Sample
## is a crude measure of effective sample size, and Rhat is the potential
## scale reduction factor on split chains (at convergence, Rhat = 1).
print(koku.res4, digits = 2)
## Warning: There were 752 divergent transitions after warmup. Increasing adapt_delta above 0.8 may help.
## See http://mc-stan.org/misc/warnings.html#divergent-transitions-after-warmup
## Family: gaussian
## Links: mu = identity; sigma = identity
## Formula: koku ~ year + cs.c + dai.c + dai.c:cs.c + year:cs.c + year:dai.c + (1 + year | id) + (1 + cs.c * dai.c | sid)
## Data: koku.nrt.lt (Number of observations: 20718)
## Samples: 4 chains, each with iter = 5000; warmup = 2500; thin = 1;
## total post-warmup samples = 10000
##
## Group-Level Effects:
## ~id (Number of levels: 3453)
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhat
## sd(Intercept) 7.35 0.11 7.14 7.56 4292 1.00
## sd(year) 0.85 0.03 0.80 0.90 3033 1.00
## cor(Intercept,year) -0.02 0.03 -0.08 0.04 7028 1.00
##
## ~sid (Number of levels: 102)
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample
## sd(Intercept) 1.68 0.29 1.06 2.24 1628
## sd(cs.c) 0.09 0.06 0.00 0.23 1503
## sd(dai.c) 7.69 6.44 0.27 24.12 5374
## sd(cs.c:dai.c) 2.60 1.87 0.11 6.92 2382
## cor(Intercept,cs.c) 0.10 0.38 -0.66 0.80 4124
## cor(Intercept,dai.c) -0.05 0.43 -0.82 0.78 9510
## cor(cs.c,dai.c) -0.05 0.45 -0.83 0.80 9915
## cor(Intercept,cs.c:dai.c) -0.22 0.42 -0.87 0.67 5305
## cor(cs.c,cs.c:dai.c) -0.01 0.44 -0.80 0.80 7164
## cor(dai.c,cs.c:dai.c) -0.01 0.45 -0.81 0.81 6691
## Rhat
## sd(Intercept) 1.00
## sd(cs.c) 1.00
## sd(dai.c) 1.00
## sd(cs.c:dai.c) 1.00
## cor(Intercept,cs.c) 1.00
## cor(Intercept,dai.c) 1.00
## cor(cs.c,dai.c) 1.00
## cor(Intercept,cs.c:dai.c) 1.00
## cor(cs.c,cs.c:dai.c) 1.00
## cor(dai.c,cs.c:dai.c) 1.00
##
## Population-Level Effects:
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhat
## Intercept 53.79 0.27 53.26 54.33 6839 1.00
## year -0.22 0.03 -0.27 -0.17 10863 1.00
## cs.c -0.07 0.05 -0.16 0.02 6379 1.00
## dai.c -6.63 10.62 -28.02 13.82 6561 1.00
## cs.c:dai.c 0.39 1.71 -2.93 3.82 7047 1.00
## year:cs.c -0.02 0.00 -0.02 -0.01 11354 1.00
## year:dai.c 2.72 0.82 1.12 4.32 10599 1.00
##
## Family Specific Parameters:
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhat
## sigma 4.25 0.03 4.20 4.30 4899 1.00
##
## Samples were drawn using sampling(NUTS). For each parameter, Eff.Sample
## is a crude measure of effective sample size, and Rhat is the potential
## scale reduction factor on split chains (at convergence, Rhat = 1).
print(koku.res5, digits = 2)
## Warning: There were 415 divergent transitions after warmup. Increasing adapt_delta above 0.8 may help.
## See http://mc-stan.org/misc/warnings.html#divergent-transitions-after-warmup
## Family: gaussian
## Links: mu = identity; sigma = identity
## Formula: koku ~ year + cs.c + dai.c + year:cs.c + year:dai.c + year:cs.c:dai.c + (1 + year | id) + (1 + cs.c * dai.c | sid)
## Data: koku.nrt.lt (Number of observations: 20718)
## Samples: 4 chains, each with iter = 5000; warmup = 2500; thin = 1;
## total post-warmup samples = 10000
##
## Group-Level Effects:
## ~id (Number of levels: 3453)
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhat
## sd(Intercept) 7.35 0.11 7.15 7.56 4110 1.00
## sd(year) 0.85 0.03 0.80 0.90 3455 1.00
## cor(Intercept,year) -0.02 0.03 -0.08 0.04 7591 1.00
##
## ~sid (Number of levels: 102)
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample
## sd(Intercept) 1.67 0.28 1.12 2.22 1879
## sd(cs.c) 0.09 0.06 0.00 0.23 1544
## sd(dai.c) 7.74 6.49 0.28 24.36 5567
## sd(cs.c:dai.c) 2.51 1.90 0.09 7.05 1658
## cor(Intercept,cs.c) 0.10 0.38 -0.66 0.81 4609
## cor(Intercept,dai.c) -0.04 0.43 -0.81 0.78 8512
## cor(cs.c,dai.c) -0.05 0.45 -0.84 0.80 10392
## cor(Intercept,cs.c:dai.c) -0.22 0.41 -0.86 0.67 5893
## cor(cs.c,cs.c:dai.c) -0.02 0.44 -0.81 0.80 7526
## cor(dai.c,cs.c:dai.c) -0.01 0.45 -0.83 0.81 6714
## Rhat
## sd(Intercept) 1.00
## sd(cs.c) 1.00
## sd(dai.c) 1.00
## sd(cs.c:dai.c) 1.00
## cor(Intercept,cs.c) 1.00
## cor(Intercept,dai.c) 1.00
## cor(cs.c,dai.c) 1.00
## cor(Intercept,cs.c:dai.c) 1.00
## cor(cs.c,cs.c:dai.c) 1.00
## cor(dai.c,cs.c:dai.c) 1.00
##
## Population-Level Effects:
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhat
## Intercept 53.81 0.25 53.33 54.30 6513 1.00
## year -0.23 0.03 -0.28 -0.17 8745 1.00
## cs.c -0.07 0.05 -0.16 0.02 7051 1.00
## dai.c -5.69 9.59 -24.74 12.78 7624 1.00
## year:cs.c -0.02 0.00 -0.02 -0.01 11718 1.00
## year:dai.c 2.30 1.05 0.22 4.35 10211 1.00
## year:cs.c:dai.c 0.10 0.16 -0.22 0.41 10034 1.00
##
## Family Specific Parameters:
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhat
## sigma 4.25 0.03 4.20 4.30 4579 1.00
##
## Samples were drawn using sampling(NUTS). For each parameter, Eff.Sample
## is a crude measure of effective sample size, and Rhat is the potential
## scale reduction factor on split chains (at convergence, Rhat = 1).
print(koku.res6, digits = 2)
## Warning: There were 187 divergent transitions after warmup. Increasing adapt_delta above 0.8 may help.
## See http://mc-stan.org/misc/warnings.html#divergent-transitions-after-warmup
## Family: gaussian
## Links: mu = identity; sigma = identity
## Formula: koku ~ year + dai.c + cs.c + dai.c:cs.c + year:cs.c + year:dai.c + year:cs.c:dai.c + (1 + year | id) + (1 + cs.c * dai.c | sid)
## Data: koku.nrt.lt (Number of observations: 20718)
## Samples: 4 chains, each with iter = 5000; warmup = 2500; thin = 1;
## total post-warmup samples = 10000
##
## Group-Level Effects:
## ~id (Number of levels: 3453)
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhat
## sd(Intercept) 7.35 0.11 7.15 7.56 4287 1.00
## sd(year) 0.85 0.03 0.80 0.90 3361 1.00
## cor(Intercept,year) -0.02 0.03 -0.08 0.04 6580 1.00
##
## ~sid (Number of levels: 102)
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample
## sd(Intercept) 1.68 0.29 1.11 2.24 1395
## sd(cs.c) 0.09 0.06 0.00 0.23 1206
## sd(dai.c) 7.65 6.32 0.30 23.57 4676
## sd(cs.c:dai.c) 2.64 1.94 0.11 7.13 1796
## cor(Intercept,cs.c) 0.10 0.38 -0.67 0.79 4363
## cor(Intercept,dai.c) -0.05 0.43 -0.81 0.77 9417
## cor(cs.c,dai.c) -0.05 0.44 -0.82 0.78 10805
## cor(Intercept,cs.c:dai.c) -0.22 0.41 -0.86 0.68 5354
## cor(cs.c,cs.c:dai.c) -0.01 0.44 -0.80 0.80 7119
## cor(dai.c,cs.c:dai.c) -0.00 0.44 -0.80 0.80 6988
## Rhat
## sd(Intercept) 1.00
## sd(cs.c) 1.00
## sd(dai.c) 1.00
## sd(cs.c:dai.c) 1.00
## cor(Intercept,cs.c) 1.00
## cor(Intercept,dai.c) 1.00
## cor(cs.c,dai.c) 1.00
## cor(Intercept,cs.c:dai.c) 1.00
## cor(cs.c,cs.c:dai.c) 1.00
## cor(dai.c,cs.c:dai.c) 1.00
##
## Population-Level Effects:
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhat
## Intercept 53.80 0.27 53.28 54.34 6355 1.00
## year -0.23 0.03 -0.28 -0.17 11823 1.00
## dai.c -6.48 10.75 -27.71 14.98 5493 1.00
## cs.c -0.07 0.05 -0.16 0.02 5410 1.00
## dai.c:cs.c 0.24 1.79 -3.30 3.72 4992 1.00
## year:cs.c -0.01 0.00 -0.02 -0.01 12588 1.00
## year:dai.c 2.33 1.09 0.21 4.45 12636 1.00
## year:dai.c:cs.c 0.10 0.16 -0.22 0.42 11296 1.00
##
## Family Specific Parameters:
## Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhat
## sigma 4.25 0.03 4.20 4.30 5082 1.00
##
## Samples were drawn using sampling(NUTS). For each parameter, Eff.Sample
## is a crude measure of effective sample size, and Rhat is the potential
## scale reduction factor on split chains (at convergence, Rhat = 1).
brms::waic(koku.res1)
##
## Computed from 10000 by 20718 log-likelihood matrix
##
## Estimate SE
## elpd_waic -64276.3 140.6
## p_waic 5393.3 81.1
## waic 128552.6 281.2
## Warning: 3116 (15.0%) p_waic estimates greater than 0.4. We recommend
## trying loo instead.
brms::waic(koku.res2)
##
## Computed from 10000 by 20718 log-likelihood matrix
##
## Estimate SE
## elpd_waic -61778.2 126.9
## p_waic 4093.6 45.9
## waic 123556.4 253.8
## Warning: 2594 (12.5%) p_waic estimates greater than 0.4. We recommend
## trying loo instead.
brms::waic(koku.res3)
##
## Computed from 10000 by 20718 log-likelihood matrix
##
## Estimate SE
## elpd_waic -61775.7 126.9
## p_waic 4085.7 45.8
## waic 123551.4 253.7
## Warning: 2591 (12.5%) p_waic estimates greater than 0.4. We recommend
## trying loo instead.
brms::waic(koku.res4)
##
## Computed from 10000 by 20718 log-likelihood matrix
##
## Estimate SE
## elpd_waic -61776.1 126.8
## p_waic 4088.7 45.8
## waic 123552.3 253.6
## Warning: 2599 (12.5%) p_waic estimates greater than 0.4. We recommend
## trying loo instead.
brms::waic(koku.res5)
##
## Computed from 10000 by 20718 log-likelihood matrix
##
## Estimate SE
## elpd_waic -61771.3 126.9
## p_waic 4085.9 45.8
## waic 123542.6 253.7
## Warning: 2579 (12.4%) p_waic estimates greater than 0.4. We recommend
## trying loo instead.
brms::waic(koku.res6)
##
## Computed from 10000 by 20718 log-likelihood matrix
##
## Estimate SE
## elpd_waic -61776.0 126.9
## p_waic 4088.4 45.9
## waic 123552.1 253.8
## Warning: 2594 (12.5%) p_waic estimates greater than 0.4. We recommend
## trying loo instead.
brms::loo(koku.res1)
## Warning: Found 1861 observations with a pareto_k > 0.7 in model
## 'koku.res1'. With this many problematic observations, it may be more
## appropriate to use 'kfold' with argument 'K = 10' to perform 10-fold cross-
## validation rather than LOO.
##
## Computed from 10000 by 20718 log-likelihood matrix
##
## Estimate SE
## elpd_loo -64244.9 130.1
## p_loo 5361.9 65.7
## looic 128489.8 260.2
## ------
## Monte Carlo SE of elpd_loo is NA.
##
## Pareto k diagnostic values:
## Count Pct. Min. n_eff
## (-Inf, 0.5] (good) 15607 75.3% 0
## (0.5, 0.7] (ok) 3250 15.7% 0
## (0.7, 1] (bad) 1326 6.4% 0
## (1, Inf) (very bad) 535 2.6% 0
## See help('pareto-k-diagnostic') for details.
brms::loo(koku.res2)
## Warning: Found 124 observations with a pareto_k > 0.7 in model 'koku.res2'.
## With this many problematic observations, it may be more appropriate to use
## 'kfold' with argument 'K = 10' to perform 10-fold cross-validation rather
## than LOO.
##
## Computed from 10000 by 20718 log-likelihood matrix
##
## Estimate SE
## elpd_loo -62000.8 129.1
## p_loo 4316.2 48.5
## looic 124001.5 258.3
## ------
## Monte Carlo SE of elpd_loo is NA.
##
## Pareto k diagnostic values:
## Count Pct. Min. n_eff
## (-Inf, 0.5] (good) 19488 94.1% 534
## (0.5, 0.7] (ok) 1106 5.3% 93
## (0.7, 1] (bad) 120 0.6% 15
## (1, Inf) (very bad) 4 0.0% 8
## See help('pareto-k-diagnostic') for details.
brms::loo(koku.res3)
## Warning: Found 121 observations with a pareto_k > 0.7 in model 'koku.res3'.
## With this many problematic observations, it may be more appropriate to use
## 'kfold' with argument 'K = 10' to perform 10-fold cross-validation rather
## than LOO.
##
## Computed from 10000 by 20718 log-likelihood matrix
##
## Estimate SE
## elpd_loo -61995.9 129.0
## p_loo 4305.9 48.3
## looic 123991.8 258.0
## ------
## Monte Carlo SE of elpd_loo is NA.
##
## Pareto k diagnostic values:
## Count Pct. Min. n_eff
## (-Inf, 0.5] (good) 19521 94.2% 329
## (0.5, 0.7] (ok) 1076 5.2% 116
## (0.7, 1] (bad) 118 0.6% 19
## (1, Inf) (very bad) 3 0.0% 9
## See help('pareto-k-diagnostic') for details.
brms::loo(koku.res4)
## Warning: Found 129 observations with a pareto_k > 0.7 in model 'koku.res4'.
## With this many problematic observations, it may be more appropriate to use
## 'kfold' with argument 'K = 10' to perform 10-fold cross-validation rather
## than LOO.
##
## Computed from 10000 by 20718 log-likelihood matrix
##
## Estimate SE
## elpd_loo -61998.1 129.0
## p_loo 4310.6 48.3
## looic 123996.1 258.0
## ------
## Monte Carlo SE of elpd_loo is NA.
##
## Pareto k diagnostic values:
## Count Pct. Min. n_eff
## (-Inf, 0.5] (good) 19524 94.2% 284
## (0.5, 0.7] (ok) 1065 5.1% 112
## (0.7, 1] (bad) 126 0.6% 11
## (1, Inf) (very bad) 3 0.0% 27
## See help('pareto-k-diagnostic') for details.
brms::loo(koku.res5)
## Warning: Found 131 observations with a pareto_k > 0.7 in model 'koku.res5'.
## With this many problematic observations, it may be more appropriate to use
## 'kfold' with argument 'K = 10' to perform 10-fold cross-validation rather
## than LOO.
##
## Computed from 10000 by 20718 log-likelihood matrix
##
## Estimate SE
## elpd_loo -61994.2 129.2
## p_loo 4308.8 48.5
## looic 123988.4 258.4
## ------
## Monte Carlo SE of elpd_loo is NA.
##
## Pareto k diagnostic values:
## Count Pct. Min. n_eff
## (-Inf, 0.5] (good) 19551 94.4% 384
## (0.5, 0.7] (ok) 1036 5.0% 79
## (0.7, 1] (bad) 123 0.6% 11
## (1, Inf) (very bad) 8 0.0% 5
## See help('pareto-k-diagnostic') for details.
brms::loo(koku.res6)
## Warning: Found 139 observations with a pareto_k > 0.7 in model 'koku.res6'.
## With this many problematic observations, it may be more appropriate to use
## 'kfold' with argument 'K = 10' to perform 10-fold cross-validation rather
## than LOO.
##
## Computed from 10000 by 20718 log-likelihood matrix
##
## Estimate SE
## elpd_loo -61999.5 129.2
## p_loo 4311.8 48.6
## looic 123999.0 258.5
## ------
## Monte Carlo SE of elpd_loo is NA.
##
## Pareto k diagnostic values:
## Count Pct. Min. n_eff
## (-Inf, 0.5] (good) 19494 94.1% 408
## (0.5, 0.7] (ok) 1085 5.2% 90
## (0.7, 1] (bad) 134 0.6% 13
## (1, Inf) (very bad) 5 0.0% 2
## See help('pareto-k-diagnostic') for details.