Reduction of variability for the assessment of side effects of toxicants on honeybees and understanding drivers for colony development

The statistical power of studies for the assessment of side effects of toxicants on honeybees conducted according to current guidelines is often limited. A new test design and modified field methods have therefore been developed to decrease uncertainty and variability and to be able to detect small effects. The new test design comprises a monitoring phase (before the tunnel phase) for the selection of honeybee colonies and modified methods, which include assessments of colony strength, an evaluation of the cell content of all cells of hives using photos and digital analysis, and the use of video recordings for the assessment of foraging activity and forager mortality. With the proposed new study design and the modified field methods variability between hives was considerably reduced, which resulted in a marked reduction of the minimum detectable difference (MDD). This makes it possible to address the Specific Protection Goals defined by the European Food Safety Authority and to gain unprecedented insight into the development of hives and driving factors.


Method to measure flight activity and forager mortality using video recordings
Programmable video cameras were placed above the entrance of hives (Figure 1). A white landing board is used to facilitate the distinction of bees from the background. Videos were recorded from dawn to dusk and subsequently analysed in the software VideoCounter 1.1. In the software individual bees are recognized in videos and their movement path is tracked. A transect is used to define the direction into which bees fly (in the hive or outside of the hive).
To exclude bees flying above the hive entrance, the recognition of flying bees was restricted based on the size of bees. The software was previously validated regarding the accuracy of counting (which was 108.7% for bees entering the hive and 82.7% for bees leaving the hive; resolution 320 x 240 px) and correction factors were applied to obtain corrected counts. Counts reflect foraging activity. From these counts forager mortality can be calculated by subtracting the daily number of entering bees from the number of leaving bees. a b c

Detailed results of colony strength
Colony strength was estimated by weight (Table 1), photography of adult bees (Table 2) and by visual estimation (Liebefelder), as recommended by OECD 75 (Table 3).

Impact of the application of the reference substance on forager mortality obtained from dead bee traps, sheets and video analysis
The number of dead counted either in dead bees traps is shown in Figure 2. The underlying data and counts in traps and sheets are provided in Table 4 and 5. Forager mortality obtained by video analysis of bees leaving and entering the hive is presented in Table 6.

Introduction
The aim of this assessment was to understand to what extent brood termination rate (BTR) depends on the number of cells that are evaluated. I.e. the uncertainty related to sample size was analysed.

Methods
Monte Carlo simulations were conducted using the brood development data of a hive from the group treated with the reference substance (hive 17-5). In this hive brood termination rate (BTR) was 83.1% for cells which contained eggs at the time of brood fixing (two days before application, DAA-2). The total number of cells with eggs at brood fixing day 0 (BFD0) was 3418. From these either 100 or 300 brood cells were randomly chosen and brood termination rates were calculated for the resulting subsets of cells. This process was repeated 1000 times to obtain an overview of how much brood termination rate can vary depending which cells are chosen.

Results
When evaluating 100 cells starting with eggs at BFD0, BTR ranged from 65.0% to 97.0%. When choosing 300 cells starting with eggs at BFD0, BTR ranged from 73.3% to 91.3%. In comparison, the true BTR of this colony was 83.1%.

Parameters considered for the evaluation of factors influence brood termination
To understand which factors affect brood termination generalized linear models (GLMs) were used. Since most termination occurred after the egg stage BTR egg was analyzed. The following parameters were taken into account:

Considered parameters for GLMs Adults from weight (AFW) Pollen cells (PC) Nectar cells (NC) Young larvae cells (YL) Capped brood cells (CB) Open brood cells (OB) Pollen cells per young larvae (PCYL) Pollen cells per old larvae (PCOL)
Model parameter estimates are provided in the following for each model (model numbers and parameter abbreviations according to Table 2 in manuscript):

Calculation of power curves for colony size
To illustrate which sample sizes would be required to reach specific MDD values for colony size power curves (Fig. 7 in publication) were calculated using Monte Carlo randomization. Notably, the calculation of these sample sizes is somewhat theoretical, since in practice, sample sizes of more than six to eight hives are not used in OECD 75 trials. However, results illustrate the effort needed to detect effects of a given magnitude.
Calculations were conducted either assuming a random choice of colonies (named 'Conventional trial' in the following 1 ) for the tunnel phase or by considering the data distributions of the actually selected hives (named 'LUV trial' in the following). The analysis was done for the endpoints adults as measured by weight, photography and visual estimations and assuming normal distribution. To estimate distributions resulting from a random choice of colonies, the coefficients of variation (CV weight = 16.12%, CV photo = 20.96%, CV visual = 22.24%) measured for all hives on the last day before the selection of hives were considered (as only the selected hives were further evaluated after the selection). To estimate distributions of selected hives (LUV trial) coefficients of variation measured on the first day in the tunnel were used (control: CV weight = 8.2%, CV photo = 9.2%, CV visual = 7.1%; reference: CV weight = 8.9%, CV photo = 9.9%, CV visual = 15.1%). As a mean of the normal distributions the average number of adults measured on the first measurement day in the tunnels (28.4.) was used to reflect a situation when a pesticide would be applied. The resulting means and standard deviations are shown in Table 9. These were used for conduct Monte Carlo simulations using 1000 iterations to calculate MDD according to Brock et al. (2014).