Scottish winter oilseed rape cultivation 2015-2016: impact of the second year of neonicotinoid seed treatment restrictions

Appendix 2 - Survey methodology

Sampling and data collection

On completion of the previous survey, conducted in 2014/15 ⁽⁸⁾ , the 96 participants were asked if they would be willing to take part in a second year of monitoring. Sixty one of these growers initially agreed to participate and 50 were successfully recruited into the 2015/16 survey. The remainder either did not grow WOSR in 2015/16 or were no longer willing to participate.

To supplement these participants, a second sample was drawn from the June 2015 Agricultural Census ⁽¹³⁾ representing winter oilseed rape cultivation in Scotland. The country was divided into 11 land-use regions (Figure 15) and the sample was stratified by these land-use regions and by holding size. The holding size groups were based on the total area of WOSR grown on the holdings. The sampling fractions used within both regions and size groups were based on the areas of relevant crops grown rather than number of holdings, so that smaller holdings would not dominate the sample. This sampling frame took into account the location of the existing participants to ensure a representative geographical spread was obtained. Fifty four growers were recruited from this sample.

These two samples were combined into a single survey sample and WOSR cultivation data were collected from 104 farmers collectively growing 5,553 ha of WOSR, representing 18 per cent of the Scottish crop. The crops surveyed were sown in 2015 and harvested in 2016, representing the second season of WOSR crops sown without insecticidal seed treatments.

Recruited growers were sent an initial explanatory letter outlining the aim of the survey and the data collection process. Data were then collected directly from growers by telephone interview. In some cases growers referred the surveyors to their agronomists for collection of some, or all, of the data.

Growers were contacted twice, once in winter 2015/16 and once in autumn 2016. At the first data collection point, growers were asked for information about their winter oilseed rape cultivation (area, seed rate, drilling date) and about operational changes they had made to mitigate for the lack of insecticidal seed treatments. Growers were also asked for information about their perception of insect pest presence, their use of autumn insecticides, their perception of the efficacy of the insecticides applied and the insect related damage that the crop incurred. Data were also collected relating to the information sources growers used to support decision making about pest and damage assessment and insecticide application. At the second data collection point growers were asked about monitoring and incidence of TuYV, 2016 yields and the about likelihood of their continuing to grow oilseed rape in the future if the restrictions continued. At both data collection points growers were also invited to make any general comments they wished about their experience of growing WOSR.

Statistical Analysis

The survey data collected in 2015/16 was compared to that collected in 2014/15. All statistical analyses were conducted by Elizabeth Duff at BioSS (Biomathematics and Statistics Scotland) using the GLMM and REML routines in GenStat 18 (VSN International Ltd, Hemel Hempstead, Herts., UK).

The traits analysed were: operational changes to crop cultivation, perceived autumn aphid pest pressure, perceived autumn flea beetle pest pressure, number of autumn insecticidal sprays, perceived spray efficacy, perceived autumn insect damage, crop re-drilling, TuYV checking and symptom occurrence and grower likelihood of cultivating WOSR in future.

For these analyses, grower was specified as a random effect to allow for repeated survey responses from those growers surveyed in both years. Fixed effects were specified as region (comparing the 7 regions in which crops were surveyed), year (comparing the 2 survey years), and region.year (testing for interactions between the 7 regions and the 2 years in order to assess whether differences between years are consistent over the 7 regions).

However, since there was no evidence of a significant interaction of region with year for the majority of the response variables, suggesting that increases or decreases with year are constant over all regions, the statistician focussed on analyses where fixed effects were specified as region + year. Where response variables showed evidence of a significant year by region interaction, this is clearly stated in the results.

In relation to reporting of mean values, for analyses using generalized linear mixed models ( GLMMs) means are presented on the transformed scale. For GLMMs with a Poisson distribution and log link function ( e.g. number of sprays) means can be converted from the transformed scale to the response scale by calculating exp(x), where x is the mean on the transformed scale. For GLMMs with a binomial distribution and logit link function ( e.g. re-drilling) means can be converted to the response scale by calculating exp(x)/(1+exp(x)), where x is the mean on the transformed scale.

Detailed analyses and results

1) Number of sprays was analysed using a generalized linear mixed model ( GLMM) with a Poisson distribution and a log link function. The dispersion was estimated from the data. A small number of growers recorded number of sprays as fractions between 0 and 1 (part-sprays). Number of sprays recorded as 0.5 or below were rounded down to 0, whilst numbers between 0.5 and 1 were rounded up to 1, prior to analysis.

There is some evidence that fewer sprays were applied on average in the second year (p=0.009) and also evidence that the number of sprays differed with region (p=0.001). The dispersion parameter was estimated as 0.472.

Mean number of sprays applied: GLMM, Poisson distribution, log link function. Arithmetic means are given in brackets.

	Year 1	Year 2	SED	p-value
Random=grower Fixed=Region+year	-0.357 (0.705)	-0.720 (0.529)	0.136	0.009

Summary of tests for fixed effects (sequentially adding terms to fixed models)

Fixed term	Wald statistic	n.d.f.	F statistic	d.d.f.	F pr
Region	23.95	6	3.98	107.2	0.001
year	7.14	1	7.14	128.1	0.009

2) Re-drilling, operational changes and problems controlling pests with sprays were recorded as 'yes' or 'no', and were analysed using a generalised linear mixed model ( GLMM) with a Binomial distribution and a logit link function. For statistical analyses of 'problems controlling pests with foliar sprays', growers responding as ' NA' were removed prior to analysis.

Re-drilling: The proportion of growers that re-drilled was very low (8/96 in year 1; 6/104 in year 2). There was no evidence of significant difference between years in the proportions.

Proportion of growers that re-drilled: GLMM, Binomial distribution, logistic link function. Observed proportions are given in brackets.

	Year 1	Year 2	SED	p-value
Random=grower Fixed=Region+year	-2.286 (0.083)	-2.697 (0.058)	0.574	0.475

Summary of tests for fixed effects (sequentially adding terms to fixed models)

Fixed term	Wald statistic	d.f.	Wald/d.f.	chi pr
Region	5.45	6	0.91	0.488
year	0.51	1	0.51	0.475

Operational changes: There was weak evidence of a decrease in the proportion of growers that made operational changes in the second year (p=0.052) but only when the interaction of region with year was not included in the analysis. When the interaction term was included in the analysis, this effect was no longer evident (results not shown).

Proportion of growers that made operational changes: GLMM, Binomial distribution, logistic link function. Observed proportions are given in brackets.

	Year 1	Year 2	SED	p-value
Random=grower Fixed=Region+year	-1.258 (0.260)	-2.009 (0.154)	0.386	0.052

Summary of tests for fixed effects (sequentially adding terms to fixed models)

Fixed term	Wald statistic	d.f.	Wald/d.f.	chi pr
Region	6.14	6	1.02	0.408
year	3.78	1	3.78	0.052

Problems controlling pests with foliar sprays: Ignoring those who responded NA, there was no evidence of significant differences between years in the proportions reporting problems. Similar results were found when the analysis was repeated excluding respondents who did not spray.

Proportion of growers that had problems controlling pests with foliar sprays: GLMM, Binomial distribution, logistic link function. Observed proportions are given in brackets.

	Year 1	Year 2	SED	p-value
All respondents (n=132)
Random=grower Fixed=Region+year	-3.506 (0.257)	-3.720 (0.210)	0.445	0.631
Only those who sprayed (n=105)
Random=grower Fixed=Region+year	-4.148 (0.155)	-4.151 (0.149)	0.590	0.995

Summary of tests for fixed effects - all respondents (sequentially adding terms to fixed models)

Fixed term	Wald statistic	d.f.	Wald/d.f.	chi pr
Region	4.06	6	0.68	0.669
year	0.23	1	0.23	0.631

Summary of tests for fixed effects - excluding those who did not spray (sequentially adding terms to fixed models)

Fixed term	Wald statistic	d.f.	Wald/d.f.	chi pr
Region	2.17	6	0.36	0.904
year	0.00	1	0.00	0.995

3) Perceived pest pressure and insect damage

Perception of aphid pest pressure, flea beetle pest pressure and insect damage were recorded on a 5 point scale: none; low; moderate; high; unknown. For statistical analyses, growers responding as 'unknown' were removed and the four remaining categories were aggregated as 'none/low' and 'moderate/high' prior to analysis using a generalised linear mixed model ( GLMM) with a Binomial distribution and logit link function.

Aphid pest pressure: The proportions of growers that perceived aphid pest pressure as moderate or high was significantly lower in the second year (p<0.001). The random effect was estimated as zero.

Proportion of growers that perceived aphid pest pressure as moderate or high: GLMM, Binomial distribution, logistic link function. Observed proportions are given in brackets.

	Year 1	Year 2	SED	p-value
Random=grower Fixed=Region+year	-1.010 (0.272)	-2.610 (0.072)	0.464	<0.001

Summary of tests for fixed effects (sequentially adding terms to fixed models)

Fixed term	Wald statistic	n.d.f.	F statistic	d.d.f.	F pr
Region	4.40	6	0.73	179.0	0.624
year	11.91	1	11.91	179.0	<0.001

Flea beetle pest pressure: The proportions of growers that perceived flea beetle pest pressure as moderate or high was significantly lower in the second year (p<0.001).

Proportion of growers that perceived flea beetle pest pressure as moderate or high: GLMM, Binomial distribution, logistic link function. Observed proportions are given in brackets.

	Year 1	Year 2	SED	p-value
Random=grower Fixed=Region+year	-0.470 (0.366)	-1.699 (0.158)	0.370	<0.001

Summary of tests for fixed effects (sequentially adding terms to fixed models)

Fixed term	Wald statistic	d.f.	Wald/d.f.	chi pr
Region	7.53	6	1.25	0.275
year	11.06	1	11.06	<0.001

Insect damage: The proportions of growers that perceived insect damage as moderate or high was significantly lower in the second year (p=0.001).

Proportion of growers that perceived insect damage as moderate or high: GLMM, Binomial distribution, logistic link function. Observed proportions are given in brackets.

	Year 1	Year 2	SED	p-value
Random=grower Fixed=Region+year	-0.457 (0.355)	-1.631 (0.157)	0.368	0.001

Summary of tests for fixed effects (sequentially adding terms to fixed models)

Fixed term	Wald statistic	d.f.	Wald/d.f.	chi pr
Region	6.04	6	1.01	0.418
year	10.17	1	10.17	0.001

Association between perceived insect damage and pest pressure - from the above, we see that the proportion of growers perceiving insect damage as moderate or high was very similar to the proportion of growers perceiving flea beetle pest pressure as moderate or high. This led us to examine, separately by year, whether it was the same growers that were recording both insect damage perception and flea beetle perception as either none/low or moderate/high. For comparison, we have also shown the association between insect damage and aphid pest pressure. As shown below, Insect damage appears to be more closely related to flea beetle presence than aphid presence.

Observed counts: flea beetle and insect damage

	Year 1		Year 2
	Insect damage		Insect damage
pest pressure	none/low	moderate/high	none/low	moderate/high
none/low	52	6	80	4
moderate/high	6	26	4	12
Total	90		100

Observed counts: aphid and insect damage

	Year 1		Year 2
	Insect damage		Insect damage
pest pressure	none/low	moderate/high	none/low	moderate/high
none/low	46	18	76	13
moderate/high	11	14	5	2
Total	89		96

4) Turnips yellow virus

Visual checking of crops for turnips yellow virus ( TuYV) and observing TuYV symptoms were recorded as 'yes' or 'no', and were analysed using a generalised linear mixed model ( GLMM) with a Binomial distribution and a logit link function.

Turnips yellow virus: visually checking crops: There was no evidence of significant differences between years in the proportions of growers that visually checked crops for TuYV.

Proportion of growers that visually checked crops for TuYV: GLMM, Binomial distribution, logistic link function. Observed proportions are given in brackets.

	Year 1	Year 2	SED	p-value
Random=grower Fixed=Region+year	1.518 (0.793)	1.953 (0.847)	0.4113	0.290

Summary of tests for fixed effects (sequentially adding terms to fixed models)

Fixed term	Wald statistic	d.f.	Wald/d.f.	chi pr
Region	1.79	6	0.30	0.938
year	1.12	1	1.12	0.290

Turnips yellow virus; observing symptoms: Focussing on only those respondents who checked crops for TuYV, the proportion of growers that saw symptoms of TuYV was very low (3/69 in year 1; 5/83 in year 2). There was no evidence of significant differences between years in the proportions of growers that observed symptoms of TuYV. Since models including the effect of region failed to converge, we show results from analysis where fixed effects were specified as year only.

Proportion of growers that saw symptoms of TuYV: GLMM, Binomial distribution, logistic link function. Observed proportions are given in brackets.

	Year 1	Year 2	SED	p-value
Only those who checked for TuYV (n=152)
Random=grower Fixed=year	-3.091 (0.043)	-2.746 (0.060)	0.753	0.646

Summary of tests for fixed effects (sequentially adding terms to fixed models)

Fixed term	Wald statistic	d.f.	Wald/d.f.	chi pr
year	0.21	1	0.21	0.646

Turnips yellow virus: testing crop: The proportion of growers that tested crop for TuYV was extremely low (1/87 in year 1; 1/98 in year 2), so it was not possible to analyse these data using GLMM analysis.

5) Yield

Yield was analysed using residual maximum likelihood ( REML) analysis. Plots of the residuals from the analysis indicated that the data did not require any transformation prior to analysis. There is evidence that yields were higher on average in the first year (p<0.001). There is also evidence that yield differed significantly on average with region (p=0.001), and evidence of a significant interaction between year and region (p=0.002), suggesting that the magnitude of changes between years was not constant over all regions but was dependent on region (see table and figure below). Since there was evidence of a significant interaction of year and region, we show results from the analyses where fixed effects were specified as region*year.

Mean yield: REML.

Region	Year 1	Year 2	Mean
Aberdeen	3.583	3.490	3.536
Angus	4.703	3.449	4.076
Central Lowlands	4.254	3.489	3.871
East Fife	3.960	3.547	3.754
Lothian	4.312	3.111	3.712
Moray Firth	4.500	2.942	3.721
Tweed Valley	4.664	3.849	4.257
Mean	4.282	3.411

SED=0.1182 for overall year comparisons; average SED=0.2586 for overall region comparisons; average SED=0.3349 for comparisons between year by region means.

Summary of tests for fixed effects (sequentially adding terms to fixed models)

Fixed term	Wald statistic	n.d.f.	F statistic	d.d.f.	F pr
Region	23.09	6	3.85	130.9	0.001
year	66.73	1	66.73	106.5	<0.001
Region.year	22.11	6	3.68	99.7	0.002

Similar results were found when the analysis was repeated including yields for 2014 (Year 0), with significant overall effects of year (p<0.001), region (p<0.001) and the interaction of year by region (p=0.002); the significant interaction suggesting that changes between years was dependent on region (see table and figure below). Since there was evidence of a significant interaction of year and region, we again show results from the analyses where fixed effects were specified as region*year. However, care must be taken when interpreting comparisons of yield between years. Year-to-year variation in yield is generally high so significant differences between years does not imply that this is due to something else that is also changing with year.

Mean yield: REML

Region	Year 0	Year 1	Year 2	Mean
Aberdeen	4.005	3.584	3.470	3.686
Angus	4.625	4.695	3.454	4.258
Central Lowlands	4.046	4.263	3.466	3.925
East Fife	3.822	3.960	3.547	3.776
Lothian	3.905	4.314	3.108	3.776
Moray Firth	4.619	4.506	2.916	4.014
Tweed Valley	4.146	4.649	3.863	4.219
Mean	4.167	4.282	3.403

Average SED=0.1170 for overall year comparisons; average SED=0.2192 for overall region comparisons; average SED=0.3281 for comparisons between year by region means.

6) Growing WOSR in future

The likelihood of growing WOSR in future if restrictions remain was recorded on a 4 point scale: less; same; more; don't know. For statistical analysis, growers responding as 'don't know' were removed and the three remaining categories were aggregated into 2 categories, 'less' and 'same/more', prior to analysis using a generalised linear mixed model ( GLMM) with a Binomial distribution and logit link function. There was no evidence of significant differences between years in the proportions that would be equally likely or more likely to grow WOSR in future.

Proportion of growers that would be equally likely or more likely to grow WOSR in future: GLMM, Binomial distribution, logistic link function. Observed proportions are given in brackets.

	Year 1	Year 2	SED	p-value
Random=grower Fixed=Region+year	1.860 (0.866)	1.616 (0.837)	0.4749	0.607

Summary of tests for fixed effects (sequentially adding terms to fixed models)

Fixed term	Wald statistic	d.f.	Wald/d.f.	chi pr
Region	7.60	6	1.27	0.269
year	0.26	1	0.26	0.607

Statistical Conclusions

Some evidence of differences between years was detected in six of the response variables analysed. However, analyses such as these describing changes over time need to be interpreted with caution since there are limitations of the inference that can be drawn. When a difference in response is detected between years, this does not imply that it is driven by something else that also changes between years.

Data quality assurance

The dataset underwent several validation processes as follows; (i) checking for any obvious errors upon data receipt (ii) checking and identifying inconsistencies and omissions once entered into spreadsheets (iii) 100 per cent checking of data held in the spreadsheets against the raw data. Where inconsistencies or errors were found these were checked against the records and with the farmer where necessary. Additional quality assurance is provided by sending reports for independent review before publication.

Main sources of bias

The data presented in this report were produced by surveying a representative sample of holdings rather than conducting a census of all the holdings in Scotland. The data, therefore, represents that sample of crop only and not all Scottish winter oilseed rape cultivation.

This survey may be subject to measurement bias as it is reliant on respondents recording and reporting data accurately. As this survey was not compulsory it may also be subject to non-response bias, as some farmers may be more likely to agree to participate than others. However, experience indicates that stratified random sampling coupled with collection of data by personal interview, delivers the highest quality data and minimises non-response bias.