Multilevel mixed modelling of ratio values in R

Question

I have a dataset of 10 columns and approx 4,5k rows, each row representing a sampling point. As sampling points share some common features, a categorical variable with 27 levels were created to address some random effects. The columns contain information related to the sex ratio at birth (SRB; number of sons/number of daughters; thus, >1 = male-biased offspring sex ratio, <1 = female-biased ratio, and 1 = equilibrium). Among the columns, one of them contains a general SRB value, while others have SRB values according to the mother's school level and age. I also have one column containing adult sex ratio (ASR).

That said, I'm trying to model the general SRB values as a function of the SRB grouped by mothers' age, school level and by ASR. As I'm working with ratio values, I choose not to log transform my data, even though some ratios are very large, say, 10 to 1, while others are very small, 1 to 10. To work around this issue, however, I started by excluding outliers from my dataset through the IQR method based on the ASR column:

#Data discretization
data$ASR_category <- ifelse(data$ASR > 1.46, "male_biased", ifelse(data$ASR < 0.96, "female_biased", "fisherian_equilibrium"))

library(dplyr)

# Calculate the Interquartile Range (IQR) for ASR
iqr_results <- data %>%
  group_by(ASR_category) %>%
  summarize(out = list(boxplot.stats(ASR)$out))

out_ind <- which(data$ASR %in% unlist(iqr_results$out))

# Exclude outliers from the dataset
outliers_cleaned <- subset(data, !row.names(data) %in% out_ind)

After excluding outliers from the ASR column, in general, all the remaining columns assumed a normal distribution. However, as SRB is positively skewed and consists of ratio values, I modeled them using family = Gamma(link = "log") through glmmTMB pck.

The model I built is as follows:

gc();glmmTMB_model <- glmmTMB(
  SRB ~ ASR + SRB_15to19 + SRB_20to29 +
    SRB_30to39 + SRB_40to49 + SRB_no_schooling_1to3 +
    SRB_4to7 + SRB_8to11 + SRB_12more + (1|state),
    data = data,
    family = Gamma(link = "log"), 
    control=glmmTMBControl(optimizer = optim,
                           optArgs = list(method = "BFGS"))
  )

Here, however, I noticed an issue in the model summary. I have model parameters estimating the effect of my independent variables, though I don't know their effect on each of the discrete states of my dependents (roughly speaking, <1, >1, 1).

Questions:

1. Is it recommended to (log) transform ratio values?
2. Since all estimates were positive, except for ASR, does it mean that all my independent variables (SRB by educational level and age) bias offspring towards sons? (Note: I don't think this is the correct interpretation of the model parameters I have, as there is a "threshold" in y~, though I don't know the correct interpretation in this case.)
3. Could a multinominal regression solve the question above?
4. Is Gamma() the appropriate family to model ratio values? (Note: Although my data assumed a normal distribution after I exclude outliers through IQR, the model with gaussian() family had convergence problems which I wasn't able to workaround even after consulting glmmTMB troubleshooting).

The above model is running smoothly, though you may find some issues when diagnosing it (see code below).

library(DHARMa) x11();plot(simulateResiduals(glmmTMB_model, n = 1000, seed = 123))

glmer() from lme4 also had some warnings(), the model output was:

Model is nearly unidentifiable: very large eigenvalue

And lmer() also didn't fit well:

x11();plot(simulateResiduals(lmer_model, n = 1000, seed = 123))

Here is my session info:

sessionInfo() R version 4.2.3 (2023-03-15 ucrt) Platform: x86_64-w64-mingw32/x64 (64-bit) Running under: Windows 10 x64 (build 22621)

other attached packages: lme4_1.1-33 Matrix_1.5-4.1 DHARMa_0.4.6 glmmTMB_1.1.7

Answer 1

First, you shouldn't exclude outliers just so your data will fit your model.

Second, whether you should log transform ratios depends on what you want to find out and whether a log transform helps you do that. The log of a ratio is, of course, the difference of the logs:

$\log{\frac{a}{b}} = \log(a) - \log(b)$

Third: Multinomial regression models are for categorical dependent variables. You don't have that. If you are thinking of categorizing ASR then a) That is generally a bad idea and b) The result would be ordinal so ordinal logistic would be, at least, a good starting point.

I'm not sure what model you should use, but this thread may help.