Building a suitable mixed effects logistic regression model taking into account gender (R) I'm new to regression analysis but I'm currently building a mixed effects logistic regression model having Pitch as the dependent variable and DurationZ, PitchMax, IntenMax, Gender as fixed effects and Subject (person) as a random effect.
Pitch is a binary yes/no and other variables are either numeric or factors.
I have a few versions of the model and I don't know which one of them will be the most accurate one. And how should I treat Gender if I know that it has an effect on other independent parameters (PitchMax*Gender or (1+PitchMax|Gender)?:
library(lme4)
testModel1 <- glmer(Pitch~DurationZ+PitchMax+IntenMax+Gender+
                   (1+DurationZ|Subject)+(1+PitchMax|Subject)+
                   (1+IntenMax|Subject),
                   data = TotalOutlScaled5, family = binomial, 
                   control = glmerControl(optimizer = "bobyqa"))

Or
testModel2 <- glmer(Pitch~DurationZ+PitchMax+IntenMax+Gender+ 
                    PitchMax*Gender + IntenMax*Gender + DurationZ*Gender+
                   (1+DurationZ|Subject)+(1+PitchMax|Subject)+
                   (1+IntenMax|Subject),
                   data = TotalOutlScaled5, family = binomial, control = 
                   glmerControl(optimizer = "bobyqa"))

Or
testModel3 <- glmer(Pitch~DurationZ+PitchMax+IntenMax+
                    (1+DurationZ|Subject)+(1+PitchMax|Subject)+ 
                    (1+IntenMax|Subject)+ 
                    (1+DurationZ|Gender)+ (1+PitchMax|Gender)+ 
                    (1+IntenMax|Gender),
                    data = TotalOutlScaled5, family = binomial, 
                    control = glmerControl(optimizer = "bobyqa"))

Or
 testModel4 <- glmer(Pitch~DurationZ+PitchMax+IntenMax+
                        (1|Gender)+(1|Subject),
                        data = TotalOutlScaled5, family = binomial, 
                        control = glmerControl(optimizer = "bobyqa"))




> str(TotalOutlScaled5)
'data.frame':   66917 obs. of 6 variables:
 $ Pitch     : Factor w/ 2 levels "no","yes": 1 2 1 1 1 1 1 2 1 2 ...
 $ DurationZ : num  0.516 0.356 0.351 0.268 0.26 ...
 $ PitchMax  : num  0.2638 0.3348 0.2603 0.1732 0.0703 ...
 $ IntenMax  : num  0.74 0.734 0.718 0.626 0.647 ...
 $ Gender    : Factor w/ 2 levels "F","M": 2 2 2 2 2 2 2 2 2 2 ...
 $ Subject   : Factor w/ 8 levels "A","B","C","D",..: 1 1 1 1 1 1 1 1 1 1 ...

 A: Gender as random effect 
Your first and second model seem most plausible. The third and fourth model with random effects for gender are a bit strange. This would imply as if you had been picking gender randomly (creating the random error) out of a pool/population with many genders.
Model 1 vs 2
Here you did not change the mixed effects but only the fixed effects (where the extra cross term with gender could be written by a:b, and a*b, equaling a + b + a:b, is all fixed + cross terms).
So the difference between these two models is an issue of model selection.
Correlation
One more point that you would have to find out is whether the glmer function allows all slopes to correlate with the intercept. The 
(1+DurationZ|Subject) + (1+PitchMax|Subject) + (1+IntenMax|Subject)

differs from 
(1|Subject) + (DurationZ - 1|Subject) + (PitchMax - 1|Subject) + (IntenMax - 1|Subject) 

and 
(1+DurationZ+PitchMax+IntenMax|Subject)

In how the random errors in the slope and intercept correlate and if the random errors in the slopes are considered to correlate with each other. Possibly the first one (slopes correlating with the intercept but not, directly, with each other) might not be possible in every software implication.
