[R-meta] No convergence meta analysis of identical multinomial logistic models.
Viechtbauer, Wolfgang (SP)
wo||g@ng@v|echtb@uer @end|ng |rom m@@@tr|chtun|ver@|ty@n|
Thu Nov 19 12:36:27 CET 2020
Dear Ricardo,
So it seems like you are meta-analyzing 28 different coefficients (based on the length of tau2 you provided). That means that there are 28*29/2 = 406 parameters in the var-cov matrix of the underlying random effects that need to be estimated (28 tau^2 values and 378 correlations). That's *a lot* of parameters compared to how much data you have. Definitely when you 'only' use 50 models/players as input, but even if you use all 2000 it is still a highly complex estimation problem.
The warning about the ratio of the sampling variances is a check that I put into rma.mv() to examine if the sampling variances (the values along the diagonal of the complete V matrix) are very unbalanced. In my experience, there can be cases where this leads to numerical issues. Whether this matters in your case depends on how large/small the sampling variances actually are. If the tau^2 values are sufficiently large (relative to the sampling variances), then the values of the sampling variances tend to matter little and this warning can be safely ignored.
Coming back to the optimization issue, there are several things you can try:
1) Increase the number of iterations. The default for nlminb() is 150 (see help(nlminb)) but that can be increased with control=list(iter.max=10000) (for example). Could also increase the number of function evaluations; e.g., control=list(iter.max=10000, eval.max=10000). In any case, be prepared to wait even longer (days?).
2) Try a different optimizer. rma.mv() allows you to select one of 15+ different ones. See help(rma.mv) and look at the 'Note' section. I would try control=list(optimizer="optim", optmethod="Nelder-Mead") or control=list(optimizer="optim", optmethod="BFGS") as a start. If your computer is beefy and has multiple cores, you can also speed up model fitting with control=list(optimizer="optimParallel", ncpus=<number of cores to use>).
3) Simplify the model.
a) Do you really need those 28 coefficients? Maybe you can do with fewer.
b) And do you even plan on making comparisons between them? If not, you could also run 28 separate models, one for each coefficient. You might lose some efficiency and you cannot test if there are differences between the coefficients, but the results per coefficient are in principle okay.
c) You could also simplify the random effects structure. struct="CS" reduces the estimation down to two parameters, but assumes equal heterogeneity for every coefficient and equal correlation for each pair. struct="HCS" allows for different tau^2 values but still assumes equal corelations. Whether this makes sense for your data I don't know.
d) 'random = ~ 1 | player.id' is not the way to do. If at all, then 'random = ~ 1 | player.id/variable.id' but this is identical to 'random = ~ factor(variable.id) | player.id' with struct="CS" (assuming that the correlation is positive). See: http://www.metafor-project.org/doku.php/analyses:konstantopoulos2011
e) Fit a simplified model (e.g., struct="CS") as a 'working model' and then use cluster-robust inference for 'fixing things up'. You can do this with robust() or, even better, with the clubSandwich package. See help(robust) for an example.
In any case, install the 'devel' version of metafor in case you are using the CRAN (2.4-0) version as I have made some minor improvements here and there that might also be useful:
https://github.com/wviechtb/metafor#installation
Best,
Wolfgang
>-----Original Message-----
>From: R-sig-meta-analysis [mailto:r-sig-meta-analysis-bounces using r-project.org]
>On Behalf Of Ricardo de Boer
>Sent: Thursday, 19 November, 2020 11:14
>To: r-sig-meta-analysis using r-project.org
>Subject: [R-meta] No convergence meta analysis of identical multinomial
>logistic models.
>
>Hi,
>
>I am fairly new to meta analysis and at the moment I ran into an
>convergence error and I was wondering what steps need to be undertaken
>to achieve convergence.
>
>Let me first explain to you my problem. The dataset I am using
>contains 8.5 million observations of pitch data in major league
>baseball. I am trying to investigate whether the pitchers
>I have 2000 multinomial models with the same explanatory variables.
>just to name a few explanatory variables: count (factorvariable with
>amount of pitches and balls), the number of outs beforethe pitch took
>place, if the pitcher is left/right handed. The dependent variable is
>categorical with the values 'Fastball', 'Breakingball' and 'Changeup'.
>
>Every model has data from 1 pitcher, where the dataset varies much in
>the amount of observations per model. The goal is to investigate
>whether these individuals behave differently, as well as what
>characteristics of the pitcher make them behave differently.
>
>I already received feedback from Wolgang on how to conduct the meta
>analysis.
>For every pitcher, I have a multinomial model where the model
>coefficients are extracted and put in a list (call it: b). This is
>also done for the var-cov of those coefficients(call it: V). Then, I
>unlisted b and V. I created a player.id vector as well as a
>variable.id vector.
>Afterwards it is possible to run a meta analysis on the coefficients
>in the following way:
>(I ran a meta analysis on the first 50 models, not 2000 yet)
>
>rma.mv(b, V, mods = ~ factor(variable.id) - 1, random = ~
>factor(variable.id) | player.id, struct="UN", sparse=TRUE,
>verbose=TRUE)
>
>The above mentioned meta analysis works when struct is set to default,
>"CS" or "HCS", but not with struct = "UN". Although in any meta
>analysis that I have run so far, I receive an warning message that
>says:
>
>Warning message:
>In rma.mv(b, V, mods = ~factor(variable.id) - 1, random =
>~factor(variable.id) | :
> Ratio of largest to smallest sampling variance extremely large. May
>not be able to obtain stable results.
>
>When I set the struct ="UN", I get the following error after many
>hours. I seems that it is stuck in a local maxima after many hours of
>running.
>
>Error in rma.mv(b, V, mods = ~factor(variable.id) - 1, random =
>~factor(variable.id) | :
> Optimizer (nlminb) did not achieve convergence (convergence = 1).
>In addition: Warning message:
>In rma.mv(b, V, mods = ~factor(variable.id) - 1, random =
>~factor(variable.id) | :
>Ratio of largest to smallest sampling variance extremely large. May
>not be able to obtain stable results.
>
>Is there a way to reach convergence? Is there anything I can do about
>these warning messages?
>
>I don't know if it helps, but I will mention the ll, tau2 and rho,
>just before I got an error of no convergence. (at the bottom of this
>message).
>
>I believe that I need to manually set tau2 and rho, but I'm not sure.
>
>Lastly, what is the difference in interpretation if I run the below
>mentioned meta analysis compared to the one Wolgang recommended me.
>The difference lies in the random part where I altered it to random =
>~ 1 | player.id and leaving out the struct = "UN", because that is
>disregarded either way then. What I think is that I ran a mixed
>effects model with only varying intercepts for the pitchers in the
>below mentioned case.
>
>rma.mv(b, V, mods = ~ factor(variable.id) - 1, random = ~ 1 |
>player.id, sparse=TRUE, verbose=TRUE)
>
>To summarize my questions:
>Is there a way to reach convergence?
>Is there anything I can do about these warning messages?
>What is the difference in interpretation?
>
>Any advice on how to proceed is highly appreciated.
>
>ll = -18603.6426
>tau2 = c(0.1508, 0.2016, 0.0593, 0.0729, 0.2317, 0.3542, 0.0452,
>0.0604, 0.0690, 0.0669, 0.1402, 0.3604, 0.2772, 0.2115, 0.1817,
>0.1524, 0.1301, 0.3421, 0.6428, 0.3078, 0.3403, 0.2608, 0.1659,
>0.4243, 0.0119, 0.0363, 0.0862, 0.4163)
>rho = c(0.2289,-0.1471, 0.0524, 0.1389,-0.0911, 0.3515, 0.1262,
>0.2239, 0.4314, 0.1383,-0.0725,-0.0780, 0.0833, 0.5362,-0.2861,
>0.1613,-0.1386,-0.3373, 0.0174,-0.2871, 0.1560, 0.2784, 0.3446,
>0.1405, 0.4257, 0.5349, 0.0060,-0.1630, 0.1319, 0.2658, 0.7631,
>0.5231, 0.5883, 0.0441,-0.0861, 0.5200, 0.1121,-0.0348, 0.2969,
>0.8384, 0.1838, 0.5219,-0.0372, 0.6533, 0.5988,-0.0022, 0.1343,
>0.4633, 0.1505, 0.8475,-0.3028,-0.0324, 0.4166, 0.6198, 0.1990,
>0.0920, 0.0044, 0.0830, 0.5364,-0.1806, 0.8285, 0.1668, 0.2280,
>0.1424, 0.5914,-0.2417, 0.3360,-0.0623,-0.4578,
>0.0022,-0.0712,-0.0746, 0.5757, 0.1784,-0.1073, 0.0367,-0.1341,
>0.0013, 0.0931, 0.3820, 0.1053, 0.0035, 0.1518, 0.1013,-0.0153,
>0.4991, 0.1533, 0.2000, 0.1243, 0.2846, 0.2767,
>0.3803,-0.0541,-0.1176, 0.3366, 0.1470, 0.0853, 0.6318, 0.3464,
>0.2588, 0.3308, 0.1981, 0.1800, 0.7962, 0.2815, 0.1303, 0.4582,
>0.1903, 0.2537, 0.3107, 0.2717,-0.1092, 0.7615, 0.3814, 0.5234,
>0.1995, 0.4969, 0.1606, 0.8527, 0.2849, 0.1483, 0.1083, 0.2908,
>0.2007, 0.7324,-0.1987, 0.2675, 0.4506, 0.5507, 0.2096,
>0.8548,-0.1065, 0.2166, 0.2160, 0.5937, 0.2572, 0.2003, 0.0453,
>0.0985, 0.5423, 0.0211, 0.6948, 0.0843, 0.5098, 0.2643,
>0.7222,-0.1023, 0.8114, 0.2588, 0.5064, 0.3885, 0.7213, 0.0550,
>0.5369,-0.0114,-0.3025,-0.1885,-0.0535,-0.3893,-0.0170,-0.0777,-0.1449,-
>0.2225,-0.1658,-0.2592,
>0.3576,-0.0914, 0.0188,-0.2216,-0.2069,-0.2428, 0.2988, 0.1031,
>0.1030, 0.1722, 0.0957, 0.3964,-0.1560, 0.1722, 0.0907, 0.2357,
>0.0104, 0.2482,-0.0033, 0.5061, 0.0755, 0.4075,-0.0022, 0.3639,
>0.0933, 0.5728, 0.2217,-0.1232, 0.1043, 0.1923, 0.1117, 0.2038,
>0.3263, 0.0959, 0.1589, 0.0522, 0.2504, 0.6467, 0.5219, 0.5852,
>0.4965, 0.3067, 0.5289, 0.2853, 0.2642,-0.0712, 0.3174,
>0.0415,-0.0727, 0.0858, 0.1702,-0.1773, 0.3730,-0.0070, 0.0960,
>0.0545, 0.3180,-0.1435, 0.7252,-0.0970, 0.6729,-0.0139,
>0.3325,-0.2098, 0.5342, 0.0148, 0.3279, 0.1504,-0.0100, 0.1517,
>0.4155, 0.1276, 0.3376, 0.3259, 0.1824, 0.1891, 0.3813, 0.1976,
>0.5315, 0.3752, 0.6874, 0.4051, 0.6222, 0.3945, 0.0516, 0.1515,
>0.7049,-0.0276,-0.0703, 0.1794, 0.0877, 0.1596, 0.1051,
>0.3875,-0.2280, 0.4786, 0.1099, 0.3915,-0.0411, 0.5162,-0.0749,
>0.5694,-0.0772, 0.7401,-0.1371, 0.6767,-0.2185, 0.3935, 0.1458,
>0.7784, 0.0420,-0.2098,-0.4941, 0.1300, 0.2519, 0.0660,-0.0708,
>0.2505, 0.0839, 0.1530, 0.2263, 0.1785,-0.1884, 0.0688,-0.4143,
>0.2752,-0.2915, 0.2570,-0.0885,-0.3592,-0.4473,-0.3541,-0.4311,
>0.0270,-0.2614,-0.1877,-0.2318, 0.0422, 0.0571,
>0.0623,-0.0238,-0.0887, 0.1438,-0.0286, 0.0653,
>0.0626,-0.0020,-0.0627,-0.0744,-0.0528,-0.0138, 0.0122,
>0.0434,-0.2359,-0.2191,-0.2303, 0.2720,-0.1875, 0.3434,
>0.3613,-0.0998, 0.0538, 0.3671, 0.4282, 0.4665, 0.2392,-0.0677,
>0.5099, 0.4577, 0.4424, 0.3033, 0.1845, 0.2091, 0.1229, 0.3054,
>0.2979, 0.3781, 0.3892,-0.3077,-0.0617, 0.1617,-0.0626, 0.3297,
>0.1311, 0.3860, 0.2505,-0.4281,-0.4427, 0.0005,-0.0658,-0.4593,
>0.0899,-0.2325,-0.2140,-0.2812, 0.0446,-0.3755, 0.0915,-0.2307,
>0.1002,-0.3179, 0.0127,-0.5230,-0.0075,-0.1111, 0.1253,-0.3566,
>0.2783,-0.4834, 0.2381,-0.0100,0.1206,-0.3672)
More information about the R-sig-meta-analysis
mailing list