Skip to content

sample size in glmer model

4 messages · Ben Bolker, Thierry Onkelinx, Souheyla GHEBGHOUB

#
[please keep r-sig-mixed-models in the Cc: list when replying]

   What is V1, what level does it vary at (among vs within
participants), and what are you interested in comparing/testing with
this model?

   Note that Barr et al 2013's "keep it maximal" protocol would
suggest that you fit (item|Participant) (but ... without
factor-analytic models or some form of regularization this will never
actually be practical, as you'd be trying to estimate a 28x28
covariance matrix (= 378 parameters) from 53 participants ...  you
usually have to choose from some more restricted set of choices
[independence, compound symmetry, etc.]

On Wed, Jan 22, 2020 at 12:36 PM Souheyla GHEBGHOUB
<souheyla.ghebghoub at gmail.com> wrote:
#
Morning Ben,

I am not doing (item|Participant).
V1 is a duration (continuous) , I should have an interaction of Time of two
levels  :  score ~ V1*Time + (Time|Participant) + (Time|item) ,
This means I have 7 parameters including intercept (4 random effects +
intercept + V1 + TimeLevel1).
I have 1484 observations and 53 participants.  The rule is 10 per parameter.

My question: Is it 10 participants per parameter (i.e. I am allowed 5
parameters) or 10 observation per parameter (I am allowed 148 parameters!!!)
I guess it is likely per participant, and if its the case, then should I
still report results even the power is less (53 divided by 7 = I have 7.5
participants per parameter, not 10 as recommended)

Thats all
Thank you,
Souheyla
#
Dear Souheyla,

Your fixed effects need 4 parameters: intercept, V1, TimeLevel2 and
V1:TimeLevel2
(Time|Participant) needs 3 parameters: variance of TimeLevel1, variance of
TimeLevel2 and their covariance.
The same goes for (Time|Item).
So you're using 10 parameters.

(Time|Participant) tries to estimate the difference between two times for
each participant. Such random effect requires much more information than
(1|Participant). In this case I'd recommended that you have 10 or more
observations for the majority of the Time/Participant combinations.

Best regards,

ir. Thierry Onkelinx
Statisticus / Statistician

Vlaamse Overheid / Government of Flanders
INSTITUUT VOOR NATUUR- EN BOSONDERZOEK / RESEARCH INSTITUTE FOR NATURE AND
FOREST
Team Biometrie & Kwaliteitszorg / Team Biometrics & Quality Assurance
thierry.onkelinx at inbo.be
Havenlaan 88 bus 73, 1000 Brussel
www.inbo.be

///////////////////////////////////////////////////////////////////////////////////////////
To call in the statistician after the experiment is done may be no more
than asking him to perform a post-mortem examination: he may be able to say
what the experiment died of. ~ Sir Ronald Aylmer Fisher
The plural of anecdote is not data. ~ Roger Brinner
The combination of some data and an aching desire for an answer does not
ensure that a reasonable answer can be extracted from a given body of data.
~ John Tukey
///////////////////////////////////////////////////////////////////////////////////////////

<https://www.inbo.be>


Op do 23 jan. 2020 om 11:15 schreef Souheyla GHEBGHOUB <
souheyla.ghebghoub at gmail.com>:

  
  
#
Thank you Thierry,

I was mistaken I have 2968 observations (2 Time ? 53 participants ? 28
items).

Is that enough for 7 parameters?
Do I have room for more interactions ( i would later do a backward
step-wise regression).

Thank you,
Souheyla

On Thu, 23 Jan 2020 at 10:39, Thierry Onkelinx <thierry.onkelinx at inbo.be>
wrote: