Pair and brood numbers
Mann-Whitney U -test for independent samples was used to compare
the overall pair and brood densities of all the studied duck species
between Maaninka and Evo. Furthermore, as mallard and teal are
generalist species and common at both study areas, hence providing
sufficient data, their pair and brood densities and brood production
between the two areas were compared separately. G -test for
goodness-of-fit was used to compare species-specific proportions of
brooded and non-brooded pairs in 2017 and 2018 at Evo with those at
Maaninka.
Pair and brood data were zero-inflated, and when exploring the non-zero
part, there was still overdispersion. Thus zero-inflated negative
binomial models were used to explain variation in the number of pairs
and broods using glmmTMB (Brooks et al., 2017). All the analyses
were done in R 3.4.0 (R Core Team, 2017) and the data exploration was
done by following the protocol by Zuur et al. (2010). Water body size
was controlled for by including shoreline length (“SHORE”) as an
explanatory variable in all the models. Field percent (“FIELD”) was
used to indicate the type of the surrounding landscape around each
wetland (1 km buffer around the wetland) in every model. The amount of
food (“FOOD”, continuous) and wetland type (“TYPE”, two levels: lake
or pond) were used as wetland-level explanatory variables. In addition,
pair number (“PAIRS”) was included in the models explaining brood
numbers. However, as data exploration revealed that pair number and
shoreline length were strongly correlated (Pearson correlation r
> 0.6), shoreline length was discarded and the pair number
was kept, because the latter should more directly determine possible
brood production. Due to the nested structure of the data, wetland ID
was entered as random factor. Year effect was excluded because it failed
to improve model fit. All possible model combinations were fitted. We
chose between the best fitting models by using model specific AIC-values
and weights (ΔAIC<2, where Δ = AICi – AICmin) (Burnham &
Anderson, 2002). Due to model selection uncertainty, we calculated the
model-averaged slopes (β-values) of the variables weighted by the Akaike
weights and their unconditional standard errors and 95% unconditional
confidence intervals; all models in the candidate set were used (see
Burnham & Anderson, 2002).