Marine primary producers in a darker future – a meta-analysis of light effects on pelagic and benthic autotrophs
Data files
Dec 16, 2022 version files 2.54 MB
-
irradianceMA.R
50.17 KB
-
latlong.csv
29.50 KB
-
metalight.csv
2.45 MB
-
README.md
4.38 KB
Abstract
The availability of underwater light, as the primary energy source for all aquatic photoautotrophs, is (and will further be) altered by changing precipitation, water turbidity, mixing depth, and terrestrial input of chromophoric dissolved organic matter (CDOM). While experimental manipulations of CDOM input and turbidity are frequent, they often involve multiple interdependent changes (light, nutrients, C-supply). To create a baseline for the expected effects of light reduction alone, we performed a weighted meta-analysis on 240 published experiments (from 108 studies yielding 2,500 effect sizes) that directly reduced light availability and measured marine autotroph responses. Across all organisms, habitats, and response variables, reduced light led to an average 23% reduction in biomass-related performance, whereas the effect sizes on physiological performance did not significantly differ from zero. Especially pigment content increased with reduced light, which indicated strong physiological plasticity in response to diminished light. This acclimation potential was also indicated by light reduction effects minimized if experiments lasted longer. Nevertheless, performance (especially biomass accrual) was reduced the more the less light intensity remained available. Light reduction effects were also more negative at higher temperatures if ambient light conditions were poor. Macrophytes or benthic systems were more negatively affected by light reduction than microalgae or plankton systems, especially in physiological responses where microalgae and plankton showed slightly positive responses. Otherwise, effect magnitudes remained surprisingly consistent across habitats and aspects of experimental design. Therefore, the strong observed log-linear relationship between remaining light and autotrophic performance can be used as a baseline to predict marine primary production in future light climate.
Methods
We followed most recent recommendations for performing research synthesis and reporting meta-analysis in the setup, conduction, and documentation of this meta-analysis (Koricheva and Gurevitch 2014, Nakagawa et al. 2017). We amend the description of the method below with a table of compliance to recommendations (Appendix A) and a detailed assessment of study selection (Appendix B) and publication bias (Appendix C).
Data extraction
Relevant studies for the search were identified on ISI Web of Science (WOS) using the search term: “(light* OR irradianc* OR shadin* OR light treatmen*) AND (experimen* OR manipul* OR field experimen* OR enclosure* mesocosm* OR treatment*) AND (phytoplankt* OR macroalga* OR microphytobenth*”) AND (marin* NOT freshwater)”. The search performed in December 2020 resulted in 1599 published studies. From these, we selected studies manipulating light availability with a clear definition of treatment and control (Appendix B). Consequently, 252 studies were selected for in-depth screening based on the abstracts. From this pre-selection, studies were excluded when i) freshwater systems were addressed, or ii) they did not report mean and a measure of variance (standard deviation, standard error of mean) for an autotroph biomass or a physiological response to light availability. 108 studies remained reporting results of 240 experiments, for which data extraction was done directly from the manuscript or from the incorporated figures using WebPlot Digitizer (Rohatgi 2019). Because single studies reported multiple experiments and single experiments were sampled more than once, data extraction yielded 2500 effect sizes (k). Multiple effect sizes from a single experiment create non-independence in the data, but as detailed below, we accounted for this by using multilevel statistical analyses. In the context of our hypotheses, this approach was needed as we explicitly test for difference in different response variables (hypotheses H1b), at different levels of light reduction (H2a) and with different timing and thus chance for acclimation (H3b).
Each experiment resulted in one or several estimates of the mean and standard deviation of the response variable for both treatment (reduced light) and control (ambient or higher light levels) as well as the respective number of replicates. Additionally, we extracted the following set of categorical and continuous variables (highlighted in bold) that detail the experiment, the habitat and organism group, potential covariates, and the response variables:
- Experiment: Experiments were either done in the lab (k = 1824) or in the field (k = 676) (lab.field), the latter including all outdoor manipulations. More precisely, experimental types distinguished field measurements (k = 228), incubation experiments (k = 177, mostly in situ), mesocosms (k = 363), and microcosms (predominantly in the lab, often with cultures, k = 1,832). The type of light treatment by three categories depended on the used methodology: (i) shading screens (k = 753), (ii) irradiance reduction by e.g. dimming light (k = 1,152), and (iii) gradients (k = 568). A fourth category “other” (k = 27) covered a variety of rarely used approaches for light change. We quantified the strength of the light treatment by the remaining % light as continuous variable, ranging from 91% to almost complete darkness. We also recorded the duration of the experiment in days as continuous, ln-transformed moderator. The size of the experiment and the ambient irradiance were reported using different measures in the studies (area and volume in the former, instantaneous photon flux and daily light doses in the latter). To be able to use both aspects, we calculated instantaneous photon flux from the daily light doses assuming 12 hr daylight and converted area-based sized to volume assuming a third dimension of 1m, thus an area of 1 m² corresponds to 1000 L. The spatial distribution of the studies was surprisingly broad (Appendix D, Figure D1), covering a range from 73.21° N to 77.52° S. We used absolute latitude as a predictor in our analyses. We followed recommendations (Appendix A) in testing whether effect sizes changed with year of publication (publication year) to see whether there were systematic biases in the evolution of this research field.
- Organisms and habitats: We differentiated studies into pelagic (k = 1,844) or benthic (k = 656) systems. Most studies were from coastal habitats (k = 1,467, both pelagic and benthic), which was contrasted to offshore (k = 278) and “culture” for experiments dealing with laboratory cultures (k = 755) categories. With respect to organism groups, we reduced the comparison to macrophytes (seagrass, benthic macroalgae, k = 550) and microalgae (k = 1,950, microphytobenthos and phytoplankton).
- Covariates: The temperature (°C) measured during the experiment was used as a continuous moderator. Only a subset of studies reported nutrient availability in the form of total nitrogen and/or total phosphorus, which were highly correlated (r = 0.90, p < 0.001). We therefore only used ln-transformed concentration of TN (µmol L-1).
- Responses: The variables used for quantifying the treatment effects were divided into two response types, biomass (k = 1,655; standing stock in the form of abundance, biomass, biovolume, or dry mass, primary production per area or volume, growth rates) and physiological responses (k = 845; cellular content of pigments, or primary productivity per unit biomass, or maximum electron transport rate). We used a second variable response category that detailed these types into four categories each: the response type biomass included i) abundance (k = 421), ii) biomass (including biovolume, mass, chlorophyll per volume or area) (k = 920), iii) growth rate (k = 155), and iv) absolute productivity (per unit area or volume) (k = 159). The response type physiology comprised i) specific productivity (per unit biomass) (k = 219), ii) cellular content of pigments (k = 145), iii) cellular content of other molecules (storage molecules, nutrients) (k = 49), and iv) the maximum quantum yield or electron transport (k = 432).
Effect Sizes
We calculated the log response ratio (LRR), which is among the most widely used effect size metrics used to quantify differences between treatment responses (LaJeunesse and Forbes 2003, Koricheva et al. 2013). Specifically, LRR represents relative changes in the response variable, as the treatment mean T is expressed as ln-transformed ratio to the control mean C.
For each effect size, we used the inverse of the sampling variance var.LRR for weighting, which is based on the standard deviation (SD), number of replicates (N) and means of treatment and control, respectively.
Statistical methods
All analyses were performed in R (R Development Core Team 2018) using the package metafor (Viechtbauer 2010). Following recommendations on how to handle non-independent effect sizes (Konstantopoulos 2011, Cheung 2019), we performed a multi-level meta-analysis, with a nested random effect structure with unique experiments nested in unique studies. We chose this approach as studies differed in how many experiments they reported.
We performed this multi-level weighted analysis without moderators (k = 2,500) to test H1a (evaluating general light reduction effects on autotrophic performance) and as a multivariate weighted metaregression using moderators to test all other hypotheses. For the latter, we chose an additive model without interactions, equivalent to a main effect analysis. A fully interactive model was not possible as in our dataset, like in most meta-analyses, the distribution of studies across categories was highly unequal and non-orthogonal. This reflects that certain level combinations are either not possible or not used in experimental designs.
To produce unbiased average effect sizes and their confidence intervals, we use the multi-level weighted analysis without moderators for a) all data, b) the biomass and physiology responses separately, and c) for all categorical groups in each predictor variable separately for physiology and biomass. Here, confidence intervals not including zero indicated significant positive or negative effects overall. For the complete model, the same random effect structure (experiments within studies) was used and amended by fixed effect moderators, for which we included response type, response category, lab.field, experiment unit type, habitat, system, organism group, and type of light treatment as categorical variables and absolute temperature, duration, publication year and remaining % light as continuous variables. Significance of predictors rejected the null hypotheses corresponding to H2 and H3).
As not all studies reported all moderator values, the additive model comprised 2078 effect sizes. In order to evaluate whether this change in the data set and the presence of other moderators mattered, we also performed univariate multi-level meta-analyses and compared the outcome to the effect of the same moderator in the additive model (Appendix E). This comparison served three additional purposes. First, it allowed investigating whether collinearity between predictors changed the sign and significance of single moderators. As detailed there, the effects were highly congruent between univariate and multivariate explanatory models. From 22 estimates in the complete model, only 2 estimates were significant in one analysis and changed sign in the other coinciding with becoming non-significant. Both these cases are detailed in the Results but overall the complete model was not strongly affected by collinearity or the reduction in k. Additionally, the univariate models provided predictor-specific intercepts in contrast to the complete model, which gives a single intercept for a certain combination of predictor groups at 0 (or 1 if log-transformed) values for continuous predictors.
For four variables, the amount of missing information was so high that their inclusion in the additive model would have caused massive reduction in the database by >500 effect sizes. Therefore, we tested the moderating effects of incoming irradiance, experiment size, latitude and TN concentrations only in the univariate analyses (Appendix E).
Appendix B: Systematic Literature Review
We used the following keywords for literature search for ISI Web of Science (WOS), with access in December 2020: “(light* OR irradianc* OR shadin* OR light treatmen*) AND (experimen* OR manipul* OR field experimen* OR enclosure* mesocosm* OR treatment*) AND (phytoplankt* OR macroalga* OR microphytobenth*”) AND (marin* NOT freshwater)”. The search resulted in 1,599 studies. 25 (default, from newest to oldest, not sorted by relevance) of the first abstracts of these studies were looked at by 3 researchers (LK, JW, MS) independently, to see if they would pick the same articles to be included in the meta-analysis. As the results were consistent, the remaining studies were divided between three of the authors. The initial exclusion criteria were
1) freshwater studies
2) no presence of light manipulation as a treatment
3) no reporting of primary production / biomass of marine primary producers, including phytoplankton, microphytobenthos, macroalgae and seagrass (but not heterotrophic bacteria)
4) have at least 1-day duration of the experiment to allow response to light treatment.
Based on the abstracts, 252 studies passed these criteria and were treated as the primary database for in-depth inspection. From these, further exclusion had to be performed as studies lacked control data for irradiance manipulation (i.e. a light reduction treatment had been performed but no control was conducted), data were not extractable (neither present in text, figures or appendix), or replication was missing.
For this secondary database, data were extracted using the free online software Webplot Digitizer. However, some studies were flagged if data were presented in a way not allowing extraction (e.g., cluttered scatterplots) or standard deviations were missing. These reasons are stated in the database and the R-script provides all steps taken. These steps reduced the database to 108 studies, which contained 240 unique experiments.