Camera Trap Predator Monitoring: 2021 - 2025
Josh Carrell, GISP | Statistical Ecologist / Geospatial Analyst @ Headwaters Corporation
2025-10-03
Camera Trap Information
|
Five years of camera monitoring data (2021–2025) from six off-channel sand and water (OCSW) sites have recently become available with the close of the 2025 Plover and Tern monitoring season. Camera trap monitoring was implemented to identify predator presence and document predation events at nests. This document provides both exploratory and statistical analyses to:
The results are intended to inform ongoing management strategies and improve understanding of how monitoring practices intersect with predator–prey dynamics in these systems. A visualization of nests with cameras and without cameras shows varying counts over the 5-year period of monitoring at each site (Figure 1.). |
|
Fates
Cameras are used specifically for monitoring nest fate. This analysis is subsetted to only 3:
Failed: Predation
Failed: Unknown
Successful Nest
Nest fates are shown proportionally for all years broken down by site and camera presence in figure 2.
Question
Our primary objective for camera monitoring at plover and tern nesting sites was to quantify nest fate. Prior to monitoring, fate could only be classified by in-field observation and some nest detective work. However, these habitats are fairly flat and lack any vegetation, making cameras very noticeable to predators. This analysis aims to not only describe what fates have been observed but explore,
Does camera presence increase the odds of predation?
Additionally, unknown fates still occur despite having cameras present, which leaves uncertainty if the unknown was predation. Both scenarios (excluding unknowns vs. including unknowns as predation) are examined.
To evaluate the effect of camera presence on nest fate outcomes, I applied a chi-squared test to assess differences in proportions between camera treatments, conducted a power and effect size analysis to evaluate sample adequacy, and fit a generalized logistic model.
Analysis Results
Predation vs. Success
When examining only known outcomes (predation and sucessful nesting), proportions are quite extreme. A contingency table describes proportional outcomes as:
| Failed-Predation | Successful Nest | |
|---|---|---|
| 0 | 24 | 152 |
| 1 | 123 | 245 |
where the proportion of the data are calculated as:
\[ \frac{Predated}{Predated + Not Predated} = Proportion Predated \]
Thus, without camera:
\[ \frac{24}{24 + 152} = {\sim.14}\] and with camera:
\[ \frac{123}{123 + 245} = {\sim.33}\]
These proportions are shown visually below:
Power, Effect Size, and Chi-Squared
The effect size for the difference in proportions was medium-to-large (\(h = 0.48\)), yielding very high statistical power (\(\approx\) 99%). The observed association between camera presence and nest outcomes was statistically significant (\(\chi^2 = 23.64\), df = 1, \(p < 0.001\)) with a small-to-moderate effect (Cramér’s \(V = 0.20\)).
| Statistic | Value | Notes |
|---|---|---|
| Effect size (h) | 0.476 | Medium-to-large effect (Cohen’s) |
| Sample size (n₁) | 368 | Nests with Cameras |
| Sample size (n₂) | 176 | Nests without Cameras |
| Significance level (α) | 0.05 | Two-sided test |
| Power | 0.999 | Very high (≈ 99%) |
| Chi-squared (χ²) | 23.64 | df = 1 |
| p-value | 1.16 × 10⁻⁶ | Highly significant |
| Cramér’s V | 0.204 | Small-to-moderate association |
GLM Interpretation
The generalized linear model becomes a logistic regression when the response is fit with a logit link function. The full GLM is shown as:
\[ y_i \sim \text{Bernoulli}(p_i) \] \[ \text{logit}(p_i) = \beta_0 + \beta_1Camera_i \]
where \(y_i\) = 1 if the nest failed and 0 if it survived. \(Camera_i\) is the binary explantory variable where 0 = no camera and 1 = camera present. \(p_i\) is the probability of nest failure which is fit with the logit link function (\(\text{ln}(\frac{p_i}{1-p_i})\)), with log-odds response transformed to probability (\(p_i = \frac{e^{\beta_0+\beta_1Camera_i}}{1 + e^{beta_0+\beta_1Camera_i}}\))
| Predictor | Log-Odds (β) | SE | Odds (e^β) | 95% CI (OR) | Probability (p) | p-value | Significance | Interpretation |
|---|---|---|---|---|---|---|---|---|
| Intercept (No Camera) | −1.846 | 0.220 | 0.16 | 0.10 – 0.24 | 0.14 | < 2 × 10⁻¹⁶ | *** | Baseline odds and probability of predation when no camera is present. |
| Camera (1 = Present) | +1.157 | 0.246 | 3.18 | 1.99 – 5.25 | 0.34 | 2.54 × 10⁻⁶ | *** | Predation odds increase ≈ 3.2× (≈ +20 pp in failure probability). |
Without cameras, nests had an estimated 86% probability of survival (odds \(\approx\) 6.3:1).
With cameras, survival was approximately 67% (odds \(\approx\) 2:1), implying roughly a two-thirds reduction in survival odds.
GLM with Interaction Effect of Year
A fixed interaction effect within a GLM pulls on the entire trend of the regression. It is logical to think that the year of the study influences the effect of camera presence on nest fate. When including the effect, the structure of the GLM is expressed as:
\[ \text{logit}(p_i) = \beta_0 + \beta_1Camera_i + \beta_2Year_i + \beta_3(Camera_i \times Year_i) \]
| Predictor | Estimate (log-odds) | SE | p-value | Significance |
|---|---|---|---|---|
| Intercept (No Camera, baseline year) | -613.17 | 331.98 | 0.065 | · |
| Camera | -86.38 | 377.78 | 0.819 | n.s. |
| Year | 0.30 | 0.16 | 0.066 | · |
| Camera × Year | 0.04 | 0.19 | 0.817 | n.s. |
Model fit: Residual deviance = 589.5 (df = 540), AIC = 597.5.
The interaction between Camera × Year was not statistically significant (p = 0.817), indicating that the effect of camera presence on nest failure did not vary meaningfully across years.
Although the main effect of Year was marginally positive (β = 0.30, p = 0.066), suggesting a slight temporal increase in nest failure probability, the overall model indicates no evidence that camera effects changed through time.
The negative but nonsignificant main effect of Camera (β = –86.38) suggests no consistent relationship between camera presence and failure probability after accounting for annual variation.
GLM with Interaction Effect of Site
A fixed interaction effect within a GLM separates the regression trend by group, allowing the effect of camera presence to vary among sites. It is logical to assume that the influence of camera monitoring on nest fate may differ depending on local site conditions, predator community, or management context. When including this effect, the structure of the GLM is expressed as:
\[ \text{logit}(p_i) = \beta_0 + \beta_1Camera_i + \beta_2Site_i + \beta_3(Camera_i \times Site_i) \]
| Predictor | Estimate (log-odds) | SE | p-value |
|---|---|---|---|
| Intercept (No Camera, reference site) | -0.29 | 0.76 | 0.706 |
| Camera | -0.88 | 0.85 | 0.301 |
| Site (Dyer) | -0.52 | 0.84 | 0.533 |
| Site (Kearney Broadfoot South) | -1.91 | 1.07 | 0.074 · |
| Site (Leaman) | -15.28 | 594.16 | 0.979 |
| Site (Newark East) | -4.08 | 1.26 | 0.001 ** |
| Site (Newark West) | -0.81 | 0.90 | 0.366 |
| Camera × Site (Dyer) | 1.33 | 0.95 | 0.161 |
| Camera × Site (Kearney Broadfoot South) | 2.77 | 1.16 | 0.017 * |
| Camera × Site (Leaman) | 15.01 | 594.16 | 0.980 |
| Camera × Site (Newark East) | 3.40 | 1.36 | 0.012 * |
| Camera × Site (Newark West) | 1.98 | 1.01 | 0.049 * |
Model fit: Residual deviance = 544.4 (df = 532), AIC = 568.4.
The inclusion of the Camera × Site interaction
allows the model to test whether the relationship between camera
presence and nest failure probability differs across sites.
Here, several interaction terms were statistically significant
(p < 0.05), indicating that the camera effect varied
meaningfully among sites.
In particular: - Kearney Broadfoot South (β
= 2.77, p = 0.017),
- Newark East (β = 3.40, p = 0.012),
and
- Newark West (β = 1.98, p =
0.049)
all showed strong positive camera interactions, suggesting that
camera presence at these sites was associated with higher odds of nest
failure.
The main effect of Camera (β = –0.88,
p = 0.30) was not significant, reflecting that the camera
effect cannot be generalized across all sites—its direction and strength
depend on local context.
These findings support the interpretation that camera impacts on nest fate are site-specific, potentially driven by differences in predator communities, habitat openness, or management interventions at each location. Intercept variability is shown below:
GLMM: Mixed Effect Model with Random Year Intercepts
A mixed-effect model behaves slightly differently than the previous fixed effect models with interactions. Instead, a random effect is included, which essentially treats the model as several distinct models that are grouped, in this case by year or site. This allows the overall groups to borrow strength from each other, and each model grouped is pulled towards the overall global mean. The math looks slightly different:
\[ \text{logit}(p_{ij}) = \beta_0 + \beta_1Camera_{ij} + \mu_{ij}\]
where \(i\) represent the individual observation (nest) and \(j\) represents the group (year). \(\beta_0\) still represents the overall fixed intercept of the regression and \(\beta_1\) is the fixed effect of camera per observation (nest) in each group (year). This is where it gets more interesting as \(\mu_{ij}\) is the random intercept for year, which allows each year to have its own baseline odds of failure. \(\mu_{ij}\) is distributed normally, \(\text{Normal}(0, \sigma^{2}_\text{Site})\) by a mean of zero and the variance of each year (e.g., group-level variability).
| Predictor | Estimate (log-odds) | SE | p-value |
|---|---|---|---|
| Intercept (No Camera) | -2.03 | 0.39 | < 0.001 *** |
| Camera | 1.15 | 0.25 | < 0.001 *** |
| Random Effect | Group | Variance | Std. Dev. |
|---|---|---|---|
| Intercept | Year | 0.4789 | 0.6921 |
Model fit: AIC = 586.9, BIC = 599.7, logLik = -290.4, n = 544 (5 years).
The positive fixed effect of Camera (β = 1.15, p < 0.001) indicates that camera presence significantly increased the probability of nest failure after accounting for annual variation.
The random intercept variance for Year (σ² = 0.48) suggests moderate variation among years in baseline nest failure risk, with some years experiencing higher or lower overall failure probabilities.
These results imply that while the camera effect is consistently positive across years, the underlying baseline failure rates differ annually—potentially reflecting interannual variability in predator pressure, environmental conditions, or management actions, which is shown below.
GLMM: Mixed Effect Model with Random Site Intercepts
| Predictor | Estimate (log-odds) | SE | p-value |
|---|---|---|---|
| Intercept (No Camera) | -1.84 | 0.38 | < 0.001 *** |
| Camera | 1.01 | 0.26 | < 0.001 *** |
| Random Effect | Group | Variance | Std. Dev. |
|---|---|---|---|
| Intercept | Site | 0.499 | 0.706 |
Model fit: AIC = 580.3, BIC = 593.2, logLik = -287.2, n = 544 (6 sites).
The positive fixed effect of Camera (β = 1.01, p < 0.001) indicates that camera presence significantly increased the probability of nest failure after accounting for variation among sites.
The random intercept variance for Site (σ² = 0.50, SD = 0.71) suggests moderate variation in baseline failure risk across sites, implying that some locations experienced higher or lower overall failure probabilities even after controlling for camera presence.
These results indicate that while the camera effect is consistently positive across sites, the baseline risk of nest failure varies among locations—potentially reflecting differences in predator communities, habitat openness, or management intensity, as visualized below.
Predation + Unknown vs. Success
When examining all failed outcomes jointly (predation + Failed-Unknown) , proportions are quite similar. A contingency table describes proportional outcomes as:
| Failed | Successful Nest | |
|---|---|---|
| 0 | 91 | 152 |
| 1 | 152 | 245 |
where the proportion of the data are calculated as:
\[ \frac{Predated}{Predated + Not Predated} = Proportion Predated \]
Thus, without camera:
\[ \frac{91}{91 + 152} = {\sim.37}\] and with camera:
\[ \frac{152}{152 + 245} = {\sim.38}\]
These proportions are shown visually below:
Power, Effect Size, and Chi-Squared
The observed difference in predation rates between nests with and
without cameras was negligible (\(h =
0.02\)), with very low statistical power (\(\approx\) 5%), indicating insufficient
sample size to detect such a small effect.
The chi-squared test showed no significant association between camera
presence and nest outcome (\(\chi^2 =
0.05\), df = 1, \(p = 0.83\)),
and Cramér’s \(V = 0.01\) suggests
virtually no relationship between the two variables.
| Statistic | Value | Notes |
|---|---|---|
| Effect size (h) | 0.017 | Negligible difference in proportions (Cohen’s) |
| Sample size (n₁) | 397 | Nests with Cameras |
| Sample size (n₂) | 243 | Nests without Cameras |
| Significance level (α) | 0.05 | Two-sided test |
| Power | 0.055 | Very low (≈ 5%) |
| Chi-squared (χ²) | 0.05 | df = 1 |
| p-value | 0.832 | Not significant |
| Cramér’s V | 0.005 | Very small (no association) |
GLM Interpretation
| Predictor | Log-Odds (β) | SE | Odds (e^β) | 95% CI (OR) | Probability (p) | p-value | Significance | Interpretation |
|---|---|---|---|---|---|---|---|---|
| Intercept (No Camera) | −0.513 | 0.133 | 0.60 | 0.46 – 0.78 | 0.38 | 0.0001 | *** | Baseline odds of failure for nests without cameras (≈38% probability of failure). |
| Camera (1 = Present) | +0.036 | 0.168 | 1.04 | 0.75 – 1.44 | 0.39 | 0.832 | n.s. | No significant difference in nest failure between camera and non-camera nests. |
AIC = 853.8, Residual deviance = 849.8 (df = 638).
The intercept indicates that, without cameras, nests had an estimated probability of failure of about 37.7% (odds ≈ 0.60:1).
The camera coefficient (β = 0.036, p = 0.832) was
not statistically significant, indicating no detectable
difference in nest failure between camera and non-camera
nests.
This suggests that, when all failure types are combined (not only
predation), camera presence does not meaningfully alter overall
nest success rates.
GLM with Interaction Effect of Year
A fixed interaction effect within a GLM pulls on the entire trend of the regression. It is logical to think that the year of the study influences the effect of camera presence on nest fate. When including the effect, the structure of the GLM is expressed as:
\[ \text{logit}(p_i) = \beta_0 + \beta_1Camera_i + \beta_2Year_i + \beta_3(Camera_i \times Year_i) \]
| Predictor | Estimate (log-odds) | SE | p-value | Significance |
|---|---|---|---|---|
| Intercept (No Camera, baseline year) | −256.38 | 183.77 | 0.163 | n.s. |
| Camera | −496.53 | 248.14 | 0.045 | * |
| Year | +0.13 | 0.09 | 0.164 | n.s. |
| Camera × Year | +0.25 | 0.12 | 0.045 | * |
Model fit: Residual deviance = 825.9 (df = 636), AIC = 833.9.
The negative coefficient for Camera (β = −496.53, p = 0.045) indicates that, at the baseline year, nests with cameras were less likely to fail, a statistically significant reduction in predation risk. The positive interaction (β = +0.25, p = 0.045) suggests that this beneficial camera effect diminished across years, with camera-equipped nests showing increasing odds of failure over time.
The main effect of Year (β = +0.13, p = 0.164) was not significant, indicating no overall temporal trend in failure risk independent of camera presence.
Although the model fit (AIC = 833.9) represents modest improvement over the null, the extremely large intercept and slope magnitudes hint at sparse or unbalanced data across certain camera–year combinations, so results should be interpreted with caution.
GLM with Interaction Effect of Site
A fixed interaction effect within a GLM separates the regression trend by group, allowing the effect of camera presence to vary among sites. It is logical to assume that the influence of camera monitoring on nest fate may differ depending on local site conditions, predator community, or management context. When including this effect, the structure of the GLM is expressed as:
\[ \text{logit}(p_i) = \beta_0 + \beta_1Camera_i + \beta_2Site_i + \beta_3(Camera_i \times Site_i) \]
| Predictor | Estimate (log-odds) | SE | p-value |
|---|---|---|---|
| Intercept (No Camera, reference site) | 0.00 | 0.71 | 1.000 |
| Camera | -1.17 | 0.80 | 0.145 |
| Site (Dyer) | -0.04 | 0.76 | 0.960 |
| Site (Kearney Broadfoot South) | 0.11 | 0.78 | 0.892 |
| Site (Leaman) | -0.69 | 1.00 | 0.488 |
| Site (Newark East) | -1.23 | 0.75 | 0.098 · |
| Site (Newark West) | -0.18 | 0.79 | 0.817 |
| Camera × Site (Dyer) | 0.95 | 0.88 | 0.276 |
| Camera × Site (Kearney Broadfoot South) | 1.02 | 0.89 | 0.253 |
| Camera × Site (Leaman) | 1.46 | 1.13 | 0.195 |
| Camera × Site (Newark East) | 0.86 | 0.88 | 0.329 |
| Camera × Site (Newark West) | 1.42 | 0.91 | 0.120 |
Model fit: Residual deviance = 800.7 (df = 628), AIC = 824.7.
The inclusion of the Camera × Site interaction allows the model to test whether the relationship between camera presence and nest failure probability differs across sites.
Although none of the interaction terms were statistically significant (p > 0.05), several sites (notably Kearney Broadfoot South, Leaman, and Newark West) exhibited relatively large positive coefficients, suggesting potential site-level differences in the direction and magnitude of the camera effect.
The negative but non-significant main effect of Camera (β = –1.17, p = 0.145) indicates that, on average, nests with cameras tended to have slightly lower odds of failure, though this pattern was not consistent across sites.
Overall, these results suggest that camera impacts on nest fate vary by site, but that the magnitude of this variation is uncertain with current sample sizes. Site-specific predator activity, habitat openness, or management intensity may explain subtle differences in camera effects across locations.
GLMM: Mixed Effect Model with Random Year Intercepts
| Predictor | Estimate (log-odds) | SE | p-value |
|---|---|---|---|
| Intercept (No Camera) | -0.61 | 0.29 | 0.035 * |
| Camera | -0.00 | 0.18 | 0.991 |
| Random Effect | Group | Variance | Std. Dev. |
|---|---|---|---|
| Intercept | Year | 0.327 | 0.572 |
Model fit: AIC = 824.2, BIC = 837.6, logLik = -409.1, n = 640 (5 years).
The fixed effect of Camera (β = –0.00, p = 0.991) indicates no detectable relationship between camera presence and overall nest failure probability after accounting for annual variation.
The intercept (β₀ = –0.61, p = 0.035) corresponds to a baseline nest failure probability of approximately 35% (odds ≈ 0.54:1) in the absence of cameras.
The random intercept variance for Year (σ² = 0.33, SD = 0.57) suggests modest interannual variation in baseline failure risk, implying that while years differ slightly in overall failure probability, the effect of cameras remained essentially constant across time.
These findings suggest that camera presence does not significantly influence nest outcomes when all failure types (not only predation) are considered, and that most of the variation in nest failure is due to year-to-year differences rather than camera treatment.
GLMM: Mixed Effect Model with Random Site Intercepts
| Predictor | Estimate (log-odds) | SE | p-value |
|---|---|---|---|
| Intercept (No Camera) | -0.39 | 0.25 | 0.119 |
| Camera | -0.14 | 0.18 | 0.424 |
| Random Effect | Group | Variance | Std. Dev. |
|---|---|---|---|
| Intercept | Site | 0.237 | 0.487 |
Model fit: AIC = 826.2, BIC = 839.6, logLik = -410.1, n = 640 (6 sites).
The fixed effect of Camera (β = –0.14, p = 0.424) indicates no significant relationship between camera presence and overall nest failure probability after accounting for site-level variation.
The intercept (β₀ = –0.39, p = 0.119) corresponds to a baseline nest failure probability of approximately 40% (odds ≈ 0.66:1) in the absence of cameras.
The random intercept variance for Site (σ² = 0.24, SD = 0.49) indicates relatively small variability in baseline failure probabilities among sites, suggesting that differences between sites were modest once camera effects were included.
Overall, these results show that camera presence did not significantly affect nest fate across sites, and that between-site differences in baseline failure risk were minor, consistent with a generally stable relationship between nest outcomes and camera presence across locations.
Interpretation
Unknown nest fates still occur even with cameras, a reminder that monitoring never captures all outcomes in ecology. Focusing only on known outcomes (predation vs. success), camera presence was associated with a ~3.2:1 increase in the odds of predation. This translates to an estimated 33% probability of predation with cameras compared to 16% without, a difference supported by adequate statistical power, effect size, and 95% confidence intervals.
Yes, Cameras do increase the odds of predation.
However, monitoring remains important. It documents predator activity and can guide adaptive management. Yet these results highlight a trade-off: the very tools we use to understand predation may elevate risk for the species we are trying to protect. A more holistic interpretation of productivity, incorporating both the benefits and costs of monitoring, is therefore essential for informing management decisions.
Uncertainty & Decision Making
In this analysis, there are two components of nest failure: Failed–Predation and Failed–Unknown. The inclusion of unknowns introduces an unavoidable layer of uncertainty and these failures could reflect unobserved predation, abandonment, weather, or other causes.
The true effect of camera presence likely lies somewhere between the two modeled extremes:
The observed odds using only confirmed predation events, and
The inclusive odds assuming all unknown failures were caused by predation.
To explore this range, a sensitivity analysis was conducted. The dashed line in the figure below represents how the odds ratio (OR) for the camera effect on nest fate would change if different proportions of unknown failures were treated as predation (x-axis).
If half of the unknowns were actually predation (a plausible assumption) the camera effect would decline to roughly 1.7 : 1, meaning cameras increase the odds of failure by ~70%. If all unknowns were predation, the relationship would approach 1 : 1, implying no overall effect.
In reality, the answer lies between these two boundaries. This uncertainty underscores the importance of interpreting model results not as absolutes, but as decision-support tools, guiding whether to continue, modify, or discontinue current camera deployments based on what we can, and cannot, directly measure.
