香京julia种子在线播放

We conducted a literature search of fish communities associated with artificial reefs vs. natural reefs in the coastal ocean using Web of Science and Google Scholar. In addition to artificial reefs, other structures, such as shipwrecks (Hoyt et al., 2014) and oil and gas infrastructure (Macreadie et al., 2011), also reside on the seafloor. In this study, we collectively refer to these structures (artificial reefs, shipwrecks, energy infrastructure) as artificial reefs. We conducted the Web of Science search on 5 November 2019 using the advanced search function with Boolean logic with the search query: TS = (artificial reef^* OR artificial habitat^* OR man-made reef^* OR shipwreck^* OR oil rig^* OR oil platform^*) AND TS = (fish^*) AND TS=(rocky reef^* OR coral reef^* OR hardbottom^* OR hard-bottom^* OR hard bottom^* OR livebottom OR live-bottom^* OR live bottom^* OR natural reef^*) AND TS = (abundance^* OR biomass^* OR densit^* OR richness^* OR diversity^*) (note that the ^* acts as a wildcard, so a string such as densit^* would represent density or densities). This search for the dates 1900 to present yielded 524 potentially relevant articles. We imported the title and abstract of each article into Colandr (Cheng et al., 2018) and screened each article according to specified inclusion criteria (Table 1). Briefly, we included field-based studies that reported fish community metrics on artificial habitats and comparable reference reef habitats in the coastal ocean through direct observations, such as diver surveys, video surveys, and net tows at depths shallower than 150 m. This initial screening revealed that 103 of the 524 articles were candidates for inclusion; we then conducted a full-text screening of these 103 articles based on the inclusion criteria. Of these studies, 32 met our criteria and were retained for data extraction.

To ensure that additional relevant papers were not overlooked, we completed a complementary search with a more basic search query (fish assemblage OR fish community AND artificial reef AND natural reef) from 1900 to present in Google Scholar on 6 November 2019. Since this was a complementary and intentionally broader search, it returned 23,600 articles sorted by relevance. We screened the titles and abstracts of the first 200 articles sorted in order of relevance in Google Scholar, conducted a full text-screen on 24 potentially relevant articles, and found 7 that met our inclusion criteria. Thus, 39 papers (32 Web of Science, 7 Google Scholar) in total were targeted for data extraction. Both the Web of Science and Google Scholar searches were for peer-reviewed literature (Table S1).

We built a relational database in Microsoft Access to facilitate data extraction. During data extraction, we recorded identifying information for each study: author, title, publication date, and journal information. We then extracted metadata related to geographic location [continent, country, state (if applicable), ocean basin, local water body, latitude, and longitude], survey approach (method, year), natural reef descriptions (natural reef type, depth, complexity, and description), artificial reef descriptions (artificial reef type, depth, complexity, description, material, date of deployment, and substrate deployed on), and any additional notes that might be important for understanding conditions of the study. If a range of depths were provided, we used the mean of the upper and lower range values as the depth. We categorized complexity as low (<2 m vertical relief), medium (2–4 m vertical relief), or high (>4 m vertical relief). If multiple sites were surveyed with differing complexities, we categorized complexity according to the highest relief reef. If complexity values were not provided, we used typical values (e.g., patch reef = low relief, concrete modules = low relief, shipwreck = high relief). We converted latitude to a factor called “latitude zone,” where 0–23.5° N/S is tropical, 23.5–35° N/S subtropical, 35–50° N/S temperate, and 50–70° N/S subarctic. There were no studies with latitudes >70° N/S. We coded artificial reefs composed of more than one material (e.g., concrete, metal, boulders, and tires) as “mixed.”

Next, we extracted data for each fish community metric (i.e., density, biomass, richness, and diversity) reported in a given study. For each fish community metric, we recorded the units, and if units were not explicitly provided, we determined them based on methodological information, such as the transect dimensions for diver surveys. We then recorded the mean value of the metric reported on artificial reefs and natural reefs, as well as the accompanying precision and noted the type of precision (none, standard error, standard deviation, 95% confidence intervals). We extracted the sample size per reef type and the location within the paper where data were extracted (e.g., table number, figure number, or paragraph location). For several papers, data on fish metrics were only available from figures, so we used the “digitize” package (Poisot, 2011) in R (R Core Team, 2019) to extract the relevant values.

While data extraction was straightforward for most publications, others required us to make decisions about which data were most relevant. In such cases, we made these decisions on the basis of comparability. As an example, in several studies, mean values were reported for one artificial reef and for multiple natural reefs but without information necessary to calculate a pooled group mean for natural reefs. In these cases, we extracted the metric value from a single natural reef that most closely matched the single artificial reef in location (e.g., geography, depth). If studies reported metrics for multiple artificial reef materials separately (e.g., metal vessel and concrete pipes), then these values were recorded as different entries (rows) in the database and associated with the same natural reef data. If a study included only a segment of the fish community (e.g., only resident species or excluded certain fish families), then it was omitted according to our inclusion criteria (Table 1) since it did not report on the entire community.

To calculate effect sizes (e.g., outcome representing relationship between fish community metrics for artificial vs. natural reefs), we required standard deviation (SD), sample sizes (n), and mean values (X) for each community metric and each study on both artificial and natural reefs. Computations or conversions were necessary to obtain the required values for some studies. When standard error (SE) was presented but not SD, we calculated SD as:

(Higgins et al., 2019). Similarly, when 95% confidence intervals (CI) were provided, we calculated the SD as:

where CI_lower and CI_upper represent the lower and upper bounds of the 95% confidence interval (Higgins et al., 2019). When multiple groups were presented for a reef type (e.g., three artificial reefs of the same material with metrics reported separately), we pooled the means and weighted by the sample size following:

where X_i is the value of the ith group, n_i is the sample size of the ith group, and X¯ is the weighted mean (Borenstein et al., 2009; Higgins et al., 2019). We calculated pooled SD as:

where N is the number of groups and SD_i is the SD of the ith group (Borenstein et al., 2009).

We calculated effect size as a standardized mean difference. Specifically, we used Hedges' g because it contains a correction factor that reduces bias. Using the sample means, standard deviation, and sample size for artificial (AR) and natural reef (NR) groups, we then calculated the Hedges' g effect size as:

(Borenstein et al., 2009), where J is a bias correction factor calculated as:

(Koricheva et al., 2013), and where s is the standard deviation pooled across reef type as:

(Borenstein et al., 2009). We calculated the variance of Hedges' g as:

(Borenstein et al., 2009). For one study, fish community metric means on artificial and natural reefs were unavailable, but ANOVA results examining the effect of artificial vs. natural reefs on metrics were available. We used the reported F value to calculate Hedges' g as:

and ensured that g was positive if the metric was higher on artificial than natural reefs and negative if the opposite (Koricheva et al., 2013). Similarly, for another study we used the reported t-value to calculate Hedges' g as:

(Koricheva et al., 2013).

For studies that did not report measures of precision (n = 10 density, n = 4 biomass, n = 17 richness, n = 5 diversity), we imputed SD by fitting a linear model between the known mean and SD, both log transformed, from the other studies (Marinho et al., 2003; Higgins et al., 2019). It is generally preferable to impute missing values than to exclude those studies with missing values from the meta-analysis (Batson and Burton, 2016; Weir et al., 2018). We then used the fitted model to impute the unknown SD values for the studies that reported means but not SD. For the one study that provided a t-value and the one study that provided an F-value that permitted calculation of Hedges' g but not the variance in Hedges' g, we used the mean variance in Hedges' g from the other studies for the same community metric and applied it (Batson and Burton, 2016). We later verified that our model outcomes were robust to imputation, as described below.

We used the extracted data to explicitly test whether fish community metrics differ between artificial reefs and natural reefs using random effects models. We fit these meta-analytic models with the “metaphor” package (Viechtbauer, 2010), providing the Hedges' g effect size and variance in Hedges' g for each study. We used random effects meta-analysis models because of their assumption that there may be different effect sizes underlying different studies (Borenstein et al., 2009). For each fish community metric, we first fit a base model without moderators (i.e., independent predicator variables). We then fit a series of mixed models using categorical moderators: ocean basin (Atlantic, Pacific, Mediterranean), latitude zone (tropical, subtropical, temperate, subarctic), and artificial reef material (metal, concrete, boulders, mixed, tires). We chose not to include depth as a moderator since our inclusion criteria specified that all studies were shallower than 150 m. We also did not include reef complexity as a moderator because reef complexity closely relates to reef material. For example, metal ships are usually high complexity, whereas concrete structures are usually low complexity. We fit models for each combination of one, two, or three moderators (ocean basin, latitude zone, artificial material), resulting in eight total models per fish community metric.

We selected the final model from each of the eight candidate models per fish community metric using Akaike Information Criteria (Burnham and Anderson, 2004), with the lowest AIC value indicating the best performing model. We also examined the heterogeneity metrics associated with each model to verify our model selection (Borenstein et al., 2009). These included several metrics: T² (tau²), the estimated total amount of heterogeneity (or residual heterogeneity if moderator included in model); I², the percent of total variability due to heterogeneity; H², the total variability divided by the within-study variance, and R², the proportion of true variance explained by the moderators relative to the base model without moderators. Model selection aimed to pick low T², I², H², values and high R² values (Borenstein et al., 2009). We additionally examined effect sizes for the models. When models contained moderators, we obtained the effect sizes of each moderator level by subtracting the model intercept. To then test for effects of moderators, we conducted omnibus tests of moderator significance (yielding QMp values, which provide a p-value for the moderator associated with the residual heterogeneity test). We also conducted likelihood ratio tests (LRTs) between the full and reduced versions of the model to examine the effect of each moderator. We visualized model results by plotting mean effect sizes and associated 95% confidence intervals, as well as by creating forest plots to visualize the effect sizes and confidence intervals of individual studies.

To verify that our final model results for each of the four fish community metrics were not dependent upon the imputation of the missing SD values, we bootstrapped the imputation procedure 1,000 times. In each bootstrap iteration, we generated a new log-log linear model, with model coefficients drawn from a 95% truncated multivariate normal distribution using the “tmvtnorm” package (Wilhelm and Manjunath, 2015). We tested how many times the significance of the moderators in each of the final models changed relative to the original fits. We also tested how many times the Hedges' g value was positive vs. negative for the base models without predictors to provide a supplementary test of the robustness of our overall results. To examine potential effects of publication bias in the final models, we created funnel plots, which help visually diagnose publication bias, and quantitatively tested for funnel plot asymmetry using regression tests (Viechtbauer, 2010). Specifically, we tested whether the observed outcomes for models without moderators or the residuals for models with moderators were related to their standard error values.

Artificial reefs hosted similar fish densities to natural reefs (Figure 2). When we examined how fish density by reef type differed as a function of moderator variables, we discovered nuances to this pattern (Figures 3, 4). Although ocean basin was not significantly related to fish density, artificial reefs in the Atlantic Ocean but not the Pacific Ocean or the Mediterranean Sea tended to host higher density than natural reefs (Figure 3A; LRT χ² = 4.13, p = 0.13). Fish density on artificial and natural reefs was associated with the latitude zone that the reefs were located in Figure 3B (LRT χ² = 10.76, p = 0.01). Artificial reefs in subarctic (50–70° N/S) zones supported higher fish density than natural reefs, whereas artificial reefs in tropical regions tended to support lower density, although this latter pattern was not statistically significant (Figures 3B, 4). Additionally, reefs composed of concrete tended to have higher density than natural reefs (Figure 3C), but material was not significant overall (LRT χ² = 6.34, p = 0.18). These patterns were represented in the final model selection that included latitude zone as a moderator (QMp < 0.001), which explained 45.97% more variance than the base model without the moderator (Table S2). These results were not influenced by the imputation of standard deviation, as all 1,000 bootstraps of the final model imputation yielded p-values for the moderator that remained significant. Additionally, the regression test for publication bias indicated no asymmetry in the funnel plot, suggesting there was not significant publication bias (Figure S4A; p =0.79).

Fish biomass did not differ on artificial and natural reefs (Figure 2, Figure S1). Biomass remained similar on artificial and natural reefs regardless of the ocean basin (Figure 5A; LRT χ² = 0.33, p = 0.85) and latitude zone (Figure 5B; LRT χ² = 2.36, p = 0.31). However, biomass was higher on artificial materials of mixed composition relative to natural reefs (Figure 5C, Table S3; LRT χ² = 15.30, p = 0.004). The final model including artificial reef material as the moderator (QMp < 0.0001) explained 100% more variance than the base model without the moderator (Table S3). The funnel plot (Figure S4B) and bootstrapping procedure both verified that these results were robust to imputation, and the regression test suggested no asymmetry (publication bias) in the funnel plot (Figure S4B; p = 0.91).

Species richness was similar on artificial vs. natural reefs (Figure 2, Figure S2). Neither ocean basin (Figure 6A; LRT χ² = 2.86, p = 0.24), latitude zone (Figure 6B; LRT χ² = 4.53, p = 0.10), artificial material (Figure 6C; LRT χ² = 0.41, p = 0.98), nor imputation affected this outcome. Although not statistically significant, artificial reefs located in the Mediterranean tended to have lower species richness than natural reefs (Figure 6A), and artificial reefs located in subtropical latitudes tended to have higher species richness than natural reefs (Figure 6B). The final model, therefore, did not include moderators (Table S4). The regression test indicated asymmetry in the funnel plot (Figure S4C; p < 0.01), suggesting the possibility that studies with small or non-significant findings were not published and thus not included in this analysis. If so, such a publication bias could influence our results.

Artificial and natural reefs exhibited similar fish diversity (Figure 2, Figure S3). Nuances emerged with both ocean basin (Figure 7A; LRT χ² = 6.61, p = 0.04) and latitude zone (Figure 7B; LRT χ² = 12.46, p = 0.006) influencing diversity (QMp < 0.0001). In general, artificial reefs in the Mediterranean or in tropical or temperate locations seemed to have higher diversity than natural reefs, although the difference was not statistically significant. Artificial material did not influence the similarity in diversity on artificial and natural reefs (Figure 7C; LRT χ² = 0.08, p = 0.78). The final model with both ocean and latitude included as moderators explained 78.25% more variance than the base model without these moderators (Table S5). The imputation procedures did not affect the model outcome, and funnel plot asymmetry was not significant, per the regression test (Figure S4D; p = 0.72).