生態(tài)學(xué)meta分析開源軟件OpenMEE介紹

• 落后的統(tǒng)計方法以及無法升級的確定限制了生態(tài)學(xué)研究對過去meta分析軟件的使用拌屏，作者開發(fā)的開源軟件OpenMEE 擁有交互式平臺，圖形處理界面（GUI）畔规。該軟件雖然基于R，但并不需要使用者擁有R編程的知識。OpenMEE 為連續(xù)的或者分類的數(shù)據(jù)提供了更為先進的meta分析和meta回歸的方法，如多協(xié)變量以及交互變量的meta回歸分析寨躁、進化樹、簡單的缺失數(shù)據(jù)設(shè)算牙勘。軟件還支持?jǐn)?shù)據(jù)的輸入和輸出，數(shù)據(jù)探索分析，圖形化以及表的匯總方面。

OpenMEE 使用簡介

• 載入數(shù)據(jù)集
• 計算效應(yīng)值（連續(xù)或二分類資料）

1. Continuous outcome e?ect sizes: raw mean di?erences, Hedges’
   d, log response ratios, bias-corrected log response ratios (Lajeunesse
   2015).
2. Dichotomous outcome e?ect sizes: log odds ratio, rate di?er
   ence, log relative risk, arcsine transformed risk, raw and log-, logit-
   or Arcsin-transformed proportions
3. Correlation coe?cient transformation to Fisher’s-Z and back
   transformation.
4. User-de?ned e?ect sizes are allowed if users provide
   variances.

• 分析方法

1. Fixed-e?ects modelling (simple inverse variance weighting).
2. Random-e?ects modelling via any of the following estimators:

Maximum Likelihood, Hedges-Olkin, DerSimonian-Laird, Sidik-
Jonkman, Restricted Maximum Likelihood, or Empirical Bayes.

3. Parametric estimation or non-parametric bootstrapping for the vari-

ances and con?dence intervals of pooled e?ect sizes.

4. Non-parametric randomization tests for heterogeneity statistics (Q-

test) in which observed e?ect sizes are resampled to assess the validity
of main e?ect tests.

5. Phylogenetic meta-analysis that account for shared evolutionary his-

tory among species (Lajeunesse 2009).

6. Automatic cumulative (Lau et al. 1992), leave-one-out and sub-group

meta-analyses.

• 具體示例如下：

1. 打開軟件

img

2. 可以采用三種方式載入數(shù)據(jù)

手動輸入

打開早先保存的 (.ome) 文件

導(dǎo)入CSV格式文件

如圖：

img

3. 載入后的數(shù)據(jù)如下圖所示放钦，默認(rèn)各列為cat屬性

img

可以通過右鍵更改數(shù)據(jù)格式和列名

img

文本如下

context

img

更改數(shù)據(jù)類型，如下圖所示數(shù)值列為 'counts' 恭金，而'Country' and 'Year'分類變量, 'study_name'是標(biāo)簽.

img

?

4. 計算效應(yīng)值

• 輸入 (i.e., not entering 'raw data')
• 根據(jù)raw data計算效應(yīng)值
• 也可以通過手動輸入

img

5. 最終的表格如下所示：

img

Here we need to discuss the issue of 'transformed' scale vs. 'raw' scale. To perform a meta-analysis, we need an effect size and a measure of dispersion for each study. For example, if our metric is the odds ratio, we need the effect size on the transformed (log) scale because confidence intervals for odds ratios are generally not symmetric in the raw (untransformed) scale. Fortunately, OpenMEE can transform an effect size and confidence interval (or effect size and variance) from one scale another.

6. 計算效應(yīng)值

img

img

選中Establish linkage......操禀，當(dāng)你更改raw data時，效應(yīng)值會隨之改變

img

計算后的效應(yīng)值如下：

img

*we see the result of transforming the effect size. Three new columns have been added containing the 'raw scale' effect size and upper and lower confidence bounds given the current confidence level shown in the toolbar. Changing the confidence level in the toolbar will cause the confidence interval to be recalculated.

7. 輸出CSV文件

In some cases, one may already have data that one would like to import for analysis. OpenMEE supports importing comma (and otherwise) delimited data. To achieve this, select Import CSV from the File menu. You will be prompted with a screen asking for details about your file format and its location, like so:

img

8. meta相關(guān)的彈性分析横腿，諸如：'bootstrapped' meta regression, meta-regression-based conditional means and boostrapped meta regression-based conditional means

• Standard Meta Analysis

Here we refer to analyses that look to estimate a 'grand mean', i.e., with no covariates. The pages you will see are as follows.

1. Data Type and Metric
   ?

   img

2. ?

3. Data Location
   ?

   img
   
   ?

4. Refine studies/categories/exclude studies with specific missing data

5. Analysis Method and Parameters

6. Summary Page

• Cumulative Meta Analysis

Cumulative meta-analysis is a meta-analytic approach in which studies are added one at a time, and the change in the cumulative effect size is observed. To perform a cumulative meta-analysis in OpenMEE, simply select Cumulative Meta-Analysis from the Analysis menu. We show a sample forest plot generated from a cumulative meta-analysis below. The left-hand side displays the usual study estimates and confidence intervals; the right hand side shows the effect on the overall (summary) estimate as studies are included in the meta-analysis.

img

• Subgroup Meta Analysis

In subgroup meta-analysis, one partitions studies into disjoint groups (e.g., studies conducted in China versus all others) and runs separate meta-analyses over these groups of studies. This is an exploratory exercise that may highlight differences between groups. The options for subgroup meta analysis are the same as for standard meta analysis except for a categorical variable must be selected. For example, if the selected variable is state, then subgroup analyses will be run on studies from each observed state. In the example shown below, this includes Massachusetts (MA) and Rhode Island (RI).

img

Phylogenetic Meta Analysis

phylogenetic meta-analysis

img

Choose effect size and variance and also a species column

img

Choose a phylogenetic tree file

img

Specify phylogenetic

model type

Meta Regression

Meta-regression allows one to explore the relationship between covariates and study outcomes. OpenMEE supports mulitple regression and both continuous and categorical predictors (covariates).

On the 'Select covariates for regression page', select the covariates for regression. On this same page, one can choose whether you want to use a random or fixed effects model and the confidence level to use in the analysis. An example of this is shown below; here we have selected "Avg. Rainfall" to be included in the regression.

img

For any categorical covariates that have been selected to be included in the meta-regression, one needs to specify the value that will serve as its reference value (i.e., the intercept). We show an example below where the covariate is Country and the reference value has been specified as USA.

img

We note that one can also assess the variance of regression coefficient estimates via bootstrapping (see below). To do so, simply select 'bootstrap' rather 'parametric' under 'Type of Analysis' when specifying meta-regression options. We show an example of bootstrap output for the rainfall predictor below.

img

Model Building

OpenMEE can compare two regression-models in terms of model-fit statistics and a likelihood ratio test. This is a strategy to assess the informativeness of a given (set of) predictors.

The process to do this model building comparison is very similiar to that of meta-regression. The only differences are that the covariates you select on the covariate select screen will be the covariates used to construct the Full model. You then select a subset of these covariates/interactions to use for the reduced model as shown in the figure.

img

Specifying the 'reduced' model

Boostrapped Meta Analysis

Bootstrapping is a non-parametric approach for assessing the variance of a parameter estimate. OpenMEE allows for bootstrapped estimates both for meta-regressions and standard meta-analysis.

In general, the options for boostrapped meta-analysis are the same as those for standard meta-analysis, with the addition that one needs to provide bootstrap-specific parameter values. These include the number of bootstrap replicates to perform, for example. We show sample output from a boostrapped meta-analysis below.

img

When running bootstrap operations in OpenMEE, the program tries to generate results using the selected number of bootstrap replicates. We note that depending on the dataset and the meta-anaytic (pooling) model being used, a meta-analysis of a specific sample of the dataset may fail. OpenMEE hands this in a straight-forward way: if one particular replicate attempt fails, another is tried, and so on until 5 times the selected of bootstrap replicates have been attempted, at which point an error message is returned.

Multiple Imputation

Sometimes data elements are missing from studies. OpenMEE supports basic imputation of covariates for meta-regression. To use this, select Multiple Imputations Meta-Regression form the Analysis drop-down menu. You will be prompted with the following screen

img

These parameters specify the imputation method to use and the number of multiple imputations to perform. Providing the methodological details of the imputation is beyond the scope of this document, but we refer the interested reader to the R MICE documentation for more details: we use this package 'under the hood' to perform the imputation.

Publication Bias

OpenMEE includes the standard tools for publication bias diagnostics, including funnel plots. These can be generated by selecting the Funnel Plot option under the Publication Bias sub-menu of Data Exploration. Example output is shown below. We caution users to interpret these with care, and to read this criticism of funnel plots.

img

Under the same Publication Bias sub-menu, one can find an option to calculate fail-safe N

參考資料
OpenMEE: Intuitive, open‐source software for meta‐analysis in ecology and evolutionary biology
https://mp.weixin.qq.com/s/KCLAnQsaLnRBomU1EdZwYg
https://mp.weixin.qq.com/s/4nn1Zof3AKZ29sgpKuki6w