[R-meta] Issues when using rma.mv

[Viechtbauer, Wolfgang (NP)]

Hi Diego,

Most of the methods used for 'standard' meta-analyses are based on the assumption that the studies are sufficiently large such that some of the 'infidelities' underlying the models (e.g., treating the estimated sampling variances as if they are known constants, assuming normal sampling distributions) can be ignored/accepted. What 'sufficiently large' means depends on the outcome measure used and various others factors, so there is no simple rule here, but n=2 is definitely not going to cut it (whether the SD is zero or not).

Beyond this, I cannot give any more advice here except to note that with sample sizes this small, it may be better to approach this as an 'individual participant data meta-analysis' (or in this case, an 'individual bird data MA' ...). See https://onlinelibrary.wiley.com/doi/book/10.1002/9781119333784 for a thorough introduction to this topic. Depending on the model used, one might also be able to include n=1 studies in such an analysis. Of course, one would then have to get access to the raw data. Not sure how feasible this is.

There are also methods for the case where some studies provide the raw data and others only summary statistics. For example:

Xiong, C., van Belle, G., Zhu, K., Miller, J. P., & Morris, J. C. (2011). A unified approach of meta-analysis: Application to an antecedent biomarker study in Alzheimer's disease. Journal of Applied Statistics, 38(1), 15-27. 

This might also be an option.

Best,
Wolfgang

>-----Original Message-----
>From: R-sig-meta-analysis [mailto:r-sig-meta-analysis-bounces using r-project.org] On
>Behalf Of Diego Gallego García via R-sig-meta-analysis
>Sent: Sunday, 21 May, 2023 20:32
>To: Michael Dewey
>Cc: Diego Gallego García; R Special Interest Group for Meta-Analysis
>Subject: Re: [R-meta] Issues when using rma.mv
>
>Hi Michael,
>
>Thanks for your message. The idea of doing a meta-regression is to take
>into account the different effect sizes of each study (i.e., some studies
>have a bigger N, or a more accurate value of the variable that I am
>studying [the SD is higher or lower depending on the accuracy of the
>measurement). Thus, in order to take into account that the different
>studies are not done with the same accuracy, I need a way to "weigh" those
>issues. I may be wrong, but I think that that is the reason why we use
>these types of analyses while doing meta-analyses and reviews. That is why
>I am not doing a classic GLMM. Am I wrong?
>
>I do not know how I can present the data in a better format. My raw data
>(after the review) has different columns (i.e., Study ID, Species,
>Latitude, Body Weight, Migratory Status, DP (Dependence Period) and SD (of
>each study), as well as N (number of individuals studied in each study).
>That is why I am concerned, in my previous question, about the fact that
>different studies have an SD of 0 and, thus, give some issues in the
>warning while doing a rma.mv
>
>I appreciate the help given!
>*MSc Diego Gallego García*
>
>Proyecto Águila del Chaco - Chaco Eagle Project
>Center for the Study and Conservation of Raptors in Argentina *(CECARA)*
>Institute of Earth and Environmental Sciences of La Pampa *(INCITAP)*
>National Scientific and Technical Research Council *(CONICET)*
>
><https://www.instagram.com/proyectoaguilachaco/>
>
>diegothen using gmail.com
>https://www.researchgate.net/profile/Diego_Gallego_Garcia
>
>El sáb, 20 may 2023 a las 10:23, Michael Dewey (<lists using dewey.myzen.co.uk>)
>escribió:
>
>> Dear Diego
>>
>> If you have all the raw data why do you not just analyse this as a
>> multi-level mode (or mixed effects model)? Perhaps I have misunderstood
>> here.
>>
>> Incidentally your mailer is set to send HTML so your post is mangled
>> since this is a plain text list.
>>
>> Michael
>>
>> On 19/05/2023 17:03, Diego Gallego García via R-sig-meta-analysis wrote:
>> > Good morning,
>> >
>> > First post here, so I apologize if I do not explain my issue in the
>> > best way.
>> >
>> > I am currently conducting a review on the length of the dependence period
>> > (time that the nestling spends with parents) in birds. I want to test
>> which
>> > ecological factors affect this variable.
>> >
>> > My raw dataset has one row per species and per study (i.e., there are
>> some
>> > studies which account for more than one species, so I separate each
>> species
>> > studied in each research), with its corresponding variables of the
>> > dependence period (PD: days), SD, N (number of individuals studied),
>> etc...
>> > My raw dataset should look like this (dummy data):
>> >
>> > Study_ID       Species      Latitude     Body weight      Status
>> >            DP        N        SD
>> > 1                    A                       50                3.4
>> >        Resident              45          3        1.81
>> > 2                    A                       38                3.4
>> >        Migrant                32          1         N/A
>> > 2                    B                        43               2.1
>> >        Migrant                33         11       3.45
>> > 3                    B                        38               2.1
>> >        Resident               38          6        2.45
>> > 3                    C                        12               3.6
>> >       Resident               159        2         0
>> >
>> > Now, as you may see, there are some studies with an n=1 (only one
>> > individual was examined), so there is no valid SD for the study (i.e.,
>> > N/A). Thus, for the escalc() formula, I should delete them. But what
>> about
>> > the last example? There were two individuals studied in that paper, and
>> > both yielded the same result, so the SD is roughly 0.
>> >
>> > Afterwards, when using rma.mv for a mixed-effects model, those studies
>> with
>> > SD=0 (then vi=0) are giving me problems. Specifically, there are some
>> > warnings in my models:
>> >
>> > Warning messages:
>> > 1: There are outcomes with non-positive sampling variances.
>> > 2: 'V' appears to be not positive definite.
>> >
>> > The model is the following:
>> > p1 <- rma.mv(data$yi, data$vi, mods= ~Latitude*Status,
>> >                 random=~1|Species, data=data)
>> >
>> > Should I be worried about these warnings? Any way to fix them?
>> >
>> > I appreciate any help.
>> > Best,
>> > *MSc Diego Gallego García*
>> >
>> > Proyecto Águila del Chaco - Chaco Eagle Project
>> > Center for the Study and Conservation of Raptors in Argentina *(CECARA)*
>> > Institute of Earth and Environmental Sciences of La Pampa *(INCITAP)*
>> > National Scientific and Technical Research Council *(CONICET)*
>> >
>> > <https://www.instagram.com/proyectoaguilachaco/>
>> >
>> > diegothen using gmail.com
>> > https://www.researchgate.net/profile/Diego_Gallego_Garcia