SNAPP, BFD and likelihood

[Susan Miller]

Hi,

I'm new to SNAPP and BFD analyses and I'm trying to figure out how to tell when I've run the program for "long enough".  I see in the BFD manual that "Stable marginal likelihood estimates usually require at least 48 steps (sometimes 100), chainLength = 100,000 (sometimes 1,000,000), and preBurnin=10,000 (sometimes 100,000)."

My question is, the likelihood values that I get from my run (48 steps, 1,000,000 chainLength, 100,000 preBurnin) are not stable. They decrease very slightly, then dramatically and then slowly again:

Step        theta         likelihood   contribution ESS

0            1            -2196.333    -151.9752    433.5722     

1            0.9308       -2197.6168   -144.5497    383.7362     

2            0.8651       -2199.0227   -137.3365    482.009      

3            0.8026       -2200.4268   -130.327     548.8232     

4            0.7434       -2201.9686   -123.5378    667.4383     

5            0.6873       -2203.156    -116.9257    689.2903     

6            0.6343       -2205.0958   -110.5577    535.8091     

7            0.5842       -2206.8963   -104.3819    538.7384     

8            0.5369       -2208.5406   -98.392      755.0615     

9            0.4924       -2211.0301   -92.6445     561.7323     

10           0.4505       -2213.518    -87.0827     692.2465     

11           0.4112       -2215.561    -81.6923     823.7485     

12           0.3743       -2218.4427   -76.5375     632.9455     

13           0.3398       -2220.0189   -71.505      898.9604     

14           0.3076       -2223.869    -66.7444     880.52       

15           0.2777       -2227.0617   -62.1391     800.9236     

16           0.2498       -2230.7504   -57.733      798.2301     

17           0.2239       -2234.0295   -53.4999     822.611      

18           0.2          -2237.45     -49.4417     901          

19           0.1779       -2241.9479   -45.589      795.2054     

20           0.1576       -2245.9114   -41.8897     901          

21           0.139        -2252.6261   -38.418      901          

22           0.1219       -2257.5462   -35.0744     649.4863     

23           0.1064       -2265.4384   -31.9438     834.5882     

24           0.0923       -2274.0714   -28.9856     901          

25           0.0796       -2285.5995   -26.2109     691.8972     

26           0.0682       -2295.1304   -23.5533     665.7409     

27           0.058        -2304.1951   -21.0446     787.8366     

28           0.0488       -2320.7937   -18.7582     727.0695     

29           0.0408       -2336.8732   -16.6159     713.6427     

30           0.0337       -2355.3282   -14.6016     682.8985     

31           0.0275       -2391.7971   -12.8545     671.5166     

32           0.0222       -2417.1843   -11.1256     705.0736     

33           0.0177       -2462.8888   -9.642       760.5135     

34           0.0138       -2572.8082   -8.6252      354.6241     

35           0.0106       -2811.9237   -7.8549      226.1951     

36           0.0079       -3101.1509   -6.8424      240.0633     

37           0.0058       -3256.1827   -5.6068      495.2796     

38           0.004        -3325.2399   -4.3962      636.9773     

39           0.0027       -3354.13     -3.3051      632.3215     

40           0.0018       -3372.5755   -2.3798      820.3404     

41           0.001        -3407.6556   -1.6291      729.2771     

42           0.0006       -3414.4595   -1.0233      863.7367     

43           0.0003       -3428.446    -0.574       901          

44           0.0001       -3434.4217   -0.2646      805.5019     

45           0            -3426.8201   -0.083       757.9917     

46           0            -3440.6993   -0.0092      626.1314     

47           0            -3425.6581   0            759.2131     

marginal L estimate = -2235.90459950448

Is this a normal pattern for the likelihood values over the steps? Have I done something wrong? If it is normal, do I just need more steps to allow it to stabilise? Am I misinterpreting the manual and/or my results? 

Full output files from the cluster, .xml and log and likelihood files from the first and last steps attached.

Any insights would be greatly appreciated.

thanks,

Susan

[Remco Bouckaert]

Hi Susan,

Judging from the ESSs, the estimates for each individual step can be considered to be very stable (that is, has a low variance in each estimate) so adding more steps will not change the pattern in estimates that you see.

Cheers,

Remco

--
You received this message because you are subscribed to the Google Groups "beast-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to beast-users...@googlegroups.com.
To post to this group, send email to beast...@googlegroups.com.
Visit this group at https://groups.google.com/group/beast-users.
For more options, visit https://groups.google.com/d/optout.
<likelihood_step47.log><runa3_step47.log><likelihood_step0.log><runa3_step0.log><SNAPP_klcruna3.o2006986><runa3.xml><SNAPP_klcruna3.e2006986>

[Susan Miller]

Hi Remco,

Okay, thanks! I guess I misinterpreted the tutorial - I thought the likelihood estimates for each step (not just within steps) were supposed to stabilise eventually.

So just to make sure I understand, if I want to compare different species grouping, I must just use the same number of steps (and other parameters) in each run, make sure the likelihood estimates within the steps have stabilised for each species grouping, and then I can calculate the BF as outlined in the BDF tutorial?

cheers,

Susan