Professional Documents
Culture Documents
Appendix B Boyle Et Al. - 2021 - Road-Effect Mitigation Promotes Connectivity and Reduces Mortality at The Population-Level
Appendix B Boyle Et Al. - 2021 - Road-Effect Mitigation Promotes Connectivity and Reduces Mortality at The Population-Level
B.1.2 Planned comparisons, effect size, and interpretation of the After × Impact
effect
Typical BACI designs conceive of the control and impact sites as having an additive relationship
— meaning that they are expected to vary in parallel (on the linear predictor scale) after
accounting for stochastic variability — until the impact site response is potentially deflected by
the effect of the impact in question (Stewart-Oaten and Bence, 2001; Stewart-Oaten et al., 1986).
We specifically examined the significance of the site×period interaction level After × Impact
because this is the key effect in the BACI design (Stewart-Oaten et al., 1986): the change in the
relative response of the sites after the impact occurs (the impact being the implementation of
road mortality mitigation in our case). We also examined the During × Impact level, which
represents the effect of incomplete mitigation. We fitted the models using the (default) treatment
contrasts for fixed effects and specified Before and Control as the reference levels. Therefore, the
Control site during the Before period is the intercept. We treated the interaction levels After ×
Impact and During × Impact as planned comparisons (i.e. without adjusting the p-value as would
be necessary for unplanned, post-hoc comparisons among treatment levels). This further
comparison is necessary because our experimental design included a During period. Without
separating the interaction levels, we could not assess the possibility that any significant overall
period × site interaction might be attributed to the During level effect.
The BACI null hypothesis predicts no substantial impact effect, in which case the impact site
response is simply the additive effect of site and the additive effect of period with no interaction
term. Under the alternate hypothesis, the impact changes the response at the impact site but not at
the control site so that the sites’ responses are no longer parallel and this change is quantified by
the interaction term. Therefore, we quantified the effect size of the After level of the mitigation
as the magnitude of the After × Impact effect relative to the null expectation of no interaction.
Similarly, the magnitude of the During × Impact level effect quantifies the effect of incomplete
mitigation.
We used the default log link for the negative binomial generalized linear models, and therefore
the expected response, μ is an exponential function of the linear predictors:
(β0 +β 1 X i1 +...+ βq X iq )
μi=e
(e.g., Zuur et al., 2009; McDonald et al., 2000), where β values are model coefficients and X iq
are entries of the design matrix. Equivalently:
β0 β1 X i1 βq X iq
μi=e ×e ×... ×e
Therefore, each exponentiated coefficient represents a proportional change in the mean response
due to the corresponding effect.
The BACI treatment coefficients given the treatment contrasts and the factor levels that we
specified are:
β BC ¿ Intercept, given that Control site Before period is the reference level.
βI ¿ Coefficent for the additive Impact site effect in all periods.
βD ¿ Coefficent for the additive effect of the During period at both sites.
βA ¿ Coefficent for the additive effect of the After period at both sites.
βD × I ¿ Coefficent for the interaction effect of the During period at the Impact site.
β A × I ¿ Coefficent for the After period and Impact site interaction (i.e., mitigation effect).
Quantifying partial effects involves examining effects of interest on the response scale (i.e., daily
counts) while holding all other effects constant. (Marginal effects are similar except they are
quantified on the scale of the linear predictors, which is related to the response scale by log link).
We held smooth terms and random effects constant to focus on the BACI partial effects, which
are:
μBC ¿eβ BC
¿ Control site; Before period (i.e., intercept).
βBC βI
μ BI ¿e e ¿ Impact site; Before period.
βBC βD
μ DC ¿e e ¿ Control site; During period.
βBC βD βI β D× I
μDI ¿e e e e ¿ Impact site; During period.
βBC βD βI
μ H 0 DI ¿e e e ¿ Hypothetical Impact site During period under H 0 .
βBC βA
μ AC ¿e e ¿ Control site; After period.
βBC βA βI β A× I
μ AI ¿e e e e ¿ Impact site; After period.
βBC βA βI
μ H 0 AI ¿e e e ¿ Hypothetical Impact site After period under H 0 .
Therefore, the response of the Impact site in the After period, μ AI , is related to the null
β
expectation as μ AI =μ H 0 AI × e so that e β quantifies the proportional effect of the impact on
A× I A ×I
the response scale (daily counts), which we report as percent change, %=(e β −1)×100 . A× I
However, generalized linear models with a log link, such as Poisson or negative binomial models
of count data, are multiplicative with effects that are proportional on the response scale, so it is
appropriate to compare effects as ratios (McDonald et al., 2000). Therefore, we considered effect
size as a ratio-of-ratios following the approach used by Conner et al. (2016) in the context of
Bayesian BACI analyses. Conner et al. (2016) denote effect size as R BACI and define it as R A /R B,
which is the ratio of mean response values at Impact and Control sites during the After period
divided by the equivalent ratio during the Before period. To illustrate this definition applied our
analysis, we set aside the During period to focus on comparing site proportional responses in the
Before vs. After periods and note that:
μ BI e β e β β
I BC
R B=¿ = β =e I
μBC e BC
μ AI e β e β e β e β
BC A I A ×I
R A =¿ = =e β e β I A× I
μ AC β
e e β BC A
and therefore effect size for this comparison is also equivalent to the exponential of the After ×
Before interaction coefficient:
βI β A× I
e e β A× I
R BACI = β
=e
e I
We generated predicted hypothetical partial effects under H 0 given estimated main effects by
adapting a procedure used with package mgcv to generate estimates and variances for sums of
predictions (in R, see help(predict.gam, package = "mgcv")). Given β as the vector of model
coefficients β q, we created vector x H 0 AI of indicator values ( x q) with the same length as β and set
all x q=0 except the three values corresponding to β BC , β A , and β I , which were set to 1. The
linear predictor for μ H 0 AI was then
T
ln (μ H 0 AI )=x H 0 AI β
and the standard error (on the linear predictor scale) was:
S E2H 0 AI =xTH 0 AI V x H 0 AI
where V is the variance-covariance matrix for the model. We used the same approach to obtain
hypothetical partial effects for μ H 0 DI . We also tested this approach by confirming that it could
recover the same estimates and standard errors for marginal effects available from a direct
application of the predict.gam function.
B.1.2.3Post hoc Before-After comparison
A strength of BACI analyses is that there is no prior assumption about temporal patterns in the
responses, which are instead inferred using observations of the Control site(s) to account for
shared trends that are independent of potential impact effects (Stewart-Oaten et al., 1986).
Therefore, temporal comparisons within-sites are not inherently meaningful for making
inferences about impact effects. Nevertheless, within-period comparisons can provide additional
context to evaluate conservation and management actions (e.g., Baxter-Gilbert et al., 2015).
Therefore, when we found an evidence of a significant After × Impact effect, we also estimated
the proportional change between Before and After periods at the Impact site. We calculated this
proportion as:
μ AI e β e β e β e β
BC A I A×I
= =e β e β A A× I
μBI eβ eβ
BC I
which we estimated with standard error using the same approach that we applied to make
hypothetical predictions for H 0. This comparison was not a focus of our analysis, and we did not
compute p-values, but we did calculate an approximate 95% CI as ± 1.96 ×SE on the linear scale.
After transforming into the response scale, CIs that do not overlap 1 suggest a potentially
significant change, which should be interpreted cautiously because the CIs are approximate and
were not adjusted for additional comparisons.
Table B.1. HGAM candidate model specifications applied to road survey counts of turtles,
snakes, and amphibians. All models contain parametric terms for site and period and vary in
the smoothing model (Smooth) applied to Julian day-of-year and presence of random effects
(RE) of date. Smooth term abbreviations refer to the taxonomy of Pedersen et al. (2019) and
were applied across levels of the effect(s) within parentheses. The Formula column provides
the linear model formula used to construct each model.
Model Smooth RE Formula
GS(year:site count ~ period + site + period:site + s(jday, k=20,
1
) m=2) + s(jday, year:site, bs="fs", k=20, m=2)
count ~ period + site + period:site + s(jday,
2 GI(year:site) bs="tp", k=20, m=2) + s(jday, by=year:site,
bs="tp", k=20, m=1) + s(year, site, bs="re")
count ~ period + site + period:site + s(jday,
3 S(year:site)
year:site, bs="fs", k=20, m=2)
count ~ period + site + period:site + s(jday,
4 I(year:site)
by=year:site, bs="tp", k=20, m=2)
count ~ period + site + period:site + s(jday, k=20,
5 GS(year)
m=2) + s(jday, year, bs="fs", k=20, m=2)
count ~ period + site + period:site + s(jday,
6 GI(year) bs="tp", k=20, m=2) + s(jday, by=year, bs="tp",
k=20, m=1) + s(year, bs="re", k=6)
count ~ period + site + period:site + s(jday, year,
7 S(year)
bs="fs", k=20, m=2)
count ~ period + site + period:site + s(jday,
8 I(year)
by=year, bs="tp", k=20, m=2)
count ~ period + site + period:site + s(jday, site,
9 S(site) date
bs="fs", k=20, m=2) + s(date, bs="re")
count ~ period + site + period:site + s(jday, site,
10 S(site)
bs="fs", k=20, m=2)
count ~ period + site + period:site + s(jday,
11 I(site) date
by=site, bs="tp", k=20, m=2) + s(date, bs="re")
count ~ period + site + period:site + s(jday,
12 I(site)
by=site, bs="tp", k=20, m=2)
count ~ period + site + period:site + s(jday,
13 G date
bs="tp", k=20, m=2) + s(date, bs="re")
count ~ period + site + period:site + s(jday,
14 G
bs="tp", k=20, m=2)
count ~ period + site + period:site + s(date,
15 date
bs="re")
16 count ~ period + site + period:site
B.2 Results
B.2.1 Data summary
Figure B.1. Histograms of road survey counts. Each observation is a daily total at one of two
sites site (Control and Impact) for A) turtles (range 0–7), B) snakes (range 0–6), and C)
amphibians (range 0–324). Plotted data include both sites and all years (2013–2018)
surveyed. Bars represent bins with widths = 1. The highest 1% amphibian counts are
indicated with circles for visibility.
Table B.2. Summary of daily road survey counts of turtles, snakes, and amphibians at
Impact and Control sites at Presqu’ile Provincial Park. Daily surveys in 2013–2017 were
simultaneous except during initial training at the start of each field season. In 2018, surveys
alternated between sites (one site-survey/day).
Mean daily count
Survey days
Turtles Snakes Amphibians
Year Period
Impac Contro Impac Contr Ratio Impac Contr Ratio Impac Contr Ratio
* * *
t l t ol t ol t ol
201 Befor
96 94 0.35 0.40 0.87 0.46 0.180 2.6 21.0 5.7 3.70
3 e
201 Befor
94 91 0.45 0.37 1.20 0.57 0.120 4.8 19.0 6.8 2.80
4 e
201 Befor
97 94 0.30 0.33 0.91 0.21 0.074 2.8 11.0 1.9 5.80
5 e
201 Durin
100 99 0.28 0.41 0.68 0.34 0.110 3.1 4.7 1.4 3.40
6 g
201
After 92 98 0.12 0.37 0.32 0.68 0.210 3.2 12.0 11.0 1.10
7
Mean daily count
Survey days
Turtles Snakes Amphibians
Year Period
Impac Contro Impac Contr Ratio Impac Contr Ratio Impac Contr Ratio
* * *
t l t ol t ol t ol
201
After 57 55 0.07 0.36 0.19 0.07 0.036 1.9 4.1 5.9 0.69
8
*
Impact mean daily count / Control mean daily count
B.2.2 AIC model selection tables
B.2.2.1 Turtles
Table B.3. HGAMs of turtle counts ranked by AIC. Smooth and random effects terms varied
among models but the fixed effects components were period + site + period × site for all
models. Column RE indicates whether a random effect of date was included.
Model Smooth RE AIC AIC
11 I(site) date 1409.8 0.0
9 S(site) date 1410.0 0.2
6 GI(year) 1410.4 0.6
13 G date 1419.1 9.3
8 I(year) 1424.8 15.0
2 GI(year:site) 1431.1 21.2
1 GS(year:site) 1431.3 21.5
5 GS(year) 1435.8 26.0
4 I(year:site) 1436.1 26.3
7 S(year) 1448.0 38.2
10 S(site) 1448.4 38.6
12 I(site) 1448.4 38.6
14 G 1451.1 41.3
3 S(year:site) 1456.0 46.2
15 date 1477.8 68.0
16 1544.2 134.4
B.2.2.2Snakes
Table B.4. HGAMs of snake counts ranked by AIC. Smooth and random effects terms
varied among models but the fixed effects components were period + site + period × site
for all models. Column RE indicates whether a random effect of date was included.
Model Smooth RE AIC AIC
6 GI(year) 1292.6 0.0
5 GS(year) 1292.9 0.3
2 GI(year:site) 1295.8 3.2
7 S(year) 1295.9 3.3
3 S(year:site) 1297.9 5.3
1 GS(year:site) 1298.0 5.4
10 S(site) 1324.4 31.8
12 I(site) 1324.5 31.9
8 I(year) 1328.7 36.1
14 G 1330.3 37.7
4 I(year:site) 1332.3 39.7
11 I(site) date 1341.1 48.5
9 S(site) date 1341.3 48.7
16 1344.7 52.1
13 G date 1349.3 56.8
15 date 1360.6 68.0
B.2.2.3Amphibians
Table B.5. HGAMs of amphibian counts ranked by AIC. Smooth and random effects terms
varied among models but the fixed effects components were period + site + period × site
for all models. Column RE indicates whether a random effect of date was included.
Model Smooth RE AIC AIC
6 GI(year) 5595.8 0.0
13 G date 5622.1 26.3
11 I(site) date 5631.4 35.6
8 I(year) 5634.0 38.2
9 S(site) date 5639.1 43.3
7 S(year) 5643.3 47.5
15 date 5646.1 50.3
2 GI(year:site) 5658.3 62.6
1 GS(year:site) 5675.6 79.9
5 GS(year) 5685.6 89.8
4 I(year:site) 5688.0 92.2
3 S(year:site) 5704.9 109.1
12 I(site) 5725.8 130.0
10 S(site) 5727.4 131.6
14 G 5729.9 134.1
16 5907.9 312.2
Figure B.2. Comparison of BACI partial effects on the response scale (mean daily counts; A,
B, C) with BACI marginal effects on the linear predictor scale (log mean daily counts; D, E,
F) for turtles (A,D), snakes (B,E), and amphibians (C,F) observed during road surveys of
Control and Impact sites. ‘ H 0 Impact’ indicates hypothetical main-effects-only predictions
of the null hypothesis of no BACI interaction effects. H 0 is the additive model with no
interaction. In our negative binomial model using a log link function an additive model
implies a consistent ratio between treatments, which are then parallel on a logarithmic scale.
Results are shown for the top AIC-ranked model for each taxon while holding random and
smoothing effects constant. Error bars indicate 95% CIs.
Figure B.3. Response-scale fitted values from negative binomial hierarchical generalized
additive models (HGAMs) of daily road survey counts fitted with a smoothing function on
Julian day with approximate 95% CI bands. The fitted values were calculated while holding
the random effect of date constant if present in the model. The top model chosen using AIC
model selection was plotted for each taxon. Amphibian CI bands were cropped to visualize
means at an appropriate scale.
Turtle model 11 planned comparisons examining both period × site interaction levels.
BA × CI level Est.a SE z Pb
During:Impact -0.254 0.304 -0.837 0.4
After:Impact -1.18 0.342 -3.47 0.00053 ***
a
coefficient value on the linear predictor (log link) scale
b
Significance. codes: 0 *** 0.001 ** 0.01 * 0.05
Turtle model 9 planned comparisons examining both period × site interaction levels.
BA × CI level Est.a SE z Pb
During:Impact -0.259 0.303 -0.854 0.39
After:Impact -1.2 0.341 -3.53 0.00042 ***
a
coefficient value on the linear predictor (log link) scale
b
Significance. codes: 0 *** 0.001 ** 0.01 * 0.05
Turtle model 6 planned comparisons examining both period × site interaction levels.
BA × CI level Est.a SE z Pb
During:Impact -0.381 0.335 -1.14 0.26
After:Impact -1.27 0.351 -3.61 0.00031 ***
a
coefficient value on the linear predictor (log link) scale
b
Significance. codes: 0 *** 0.001 ** 0.01 * 0.05
Snake model 6 planned comparisons examining both period × site interaction levels.
BA × CI level Est.a SE z Pb
During:Impact -0.0977 0.435 -0.224 0.82
After:Impact -0.0872 0.357 -0.244 0.81
a
coefficient value on the linear predictor (log link) scale
b
Significance. codes: 0 *** 0.001 ** 0.01 * 0.05
Snake model 5 planned comparisons examining both period × site interaction levels.
BA × CI level Est.a SE z Pb
During:Impact -0.0934 0.435 -0.215 0.83
After:Impact -0.0858 0.357 -0.24 0.81
a
coefficient value on the linear predictor (log link) scale
b
Significance. codes: 0 *** 0.001 ** 0.01 * 0.05
Amphibian model 6 planned comparisons examining both period × site interaction levels.
BA × CI level Est.a SE z Pb
During:Impact 0.0673 0.27 0.249 0.8
After:Impact -1.2 0.214 -5.6 2.2e-08 ***
a
coefficient value on the linear predictor (log link) scale
b
Significance. codes: 0 *** 0.001 ** 0.01 * 0.05