exploring inference for 'separated' trunc distribution

trunc_compois_phenom.Rmd

```{r}
dd <- read_excel("Truncated.compois.problem.xlsx")
dd <- transform(dd, pop = factor(pop), morph = factor(morph))
m0 <- glmmTMB(mates ~ pop*morph,
              family='truncated_compois', data=dd)
m1 <- update(m0, contrasts = list(pop = my_sum(dd$pop), morph = my_sum(dd$morph)))
tt <- purrr::map_dfr(list(treatment=m0,sum=m1),
                     tidy, effects = "fixed", .id = "contrasts", conf.int = TRUE) %>%
    filter(term != "(Intercept)") %>%
    mutate(across(term, ~gsub("(pop\\.?|sum_)","",.)))
```

```{r dwplot1}
ggplot(tt, aes(y = term, x = estimate, colour = contrasts))  +
    scale_x_continuous(limits=c(-25,20), oob = scales::squish) +
    geom_pointrange(aes(xmin=conf.low, xmax = conf.high), position = position_dodge(width = 0.5)) +
    facet_wrap(~contrasts,scales = "free_y")
```


`drop1` gives identical results (overall significance of the interaction):

```{r drop1}
drop1(m0, test="Chisq")    
drop1(m1, test="Chisq")
```

These are different even though they're using type II estimates ... ? (Could check and see if they're the same if we drop `ELS`)
```{r anova}
car::Anova(m0)
car::Anova(m1)
```

```{r}
m0B <- update(m0, . ~ morph + pop)
m1B <- update(m1, . ~ morph + pop)
drop1(m0B, test="Chisq")
drop1(m1B, test="Chisq")
```

Again identical (except warning  from `m1B`, maybe because of weird custom contrasts?

```{r}
noELS <- droplevels(filter(dd, pop != "ELS"))
m2 <- update(m1, data = noELS,
             contrasts = list(pop = my_sum(noELS$pop)))
dwplot(m2, effects = "fixed") + geom_vline(xintercept = 0, lty=2)
```