What to extract from each study
The analysis is only as good as the numbers you pull out of each paper, so decide the format before you start reading. For a binary outcome, record the events and the total in each arm (for example 12 of 84 versus 20 of 80). For a continuous outcome, record the mean, the standard deviation, and the sample size per arm. When a study reports neither, but gives an effect with a confidence interval (an odds ratio of 1.8 with a 95% CI of 1.1 to 2.9, or an adjusted hazard ratio), record the effect and both interval bounds. That is enough for a generic inverse-variance analysis.
The common snag is a study that reports only the effect and its CI, not a standard error. You can recover the standard error from the interval. For a ratio measure, take the natural log of the effect and of both bounds, then SE equals (ln upper minus ln lower) divided by 3.92. For a difference measure, drop the logs: SE equals (upper minus lower) divided by 3.92. The effect size calculator does this for you and also converts between OR, RR, RD, and NNT.
Choosing the effect measure
Match the measure to the outcome. Use an odds ratio (OR) or risk ratio (RR) for binary outcomes; RR is easier to explain to clinicians, OR is what logistic regression and case-control designs report. Use a hazard ratio (HR) for time-to-event outcomes such as survival. Use a mean difference (MD) when every study measures the outcome on the same scale, and a standardized mean difference (SMD) when they use different scales for the same construct. Pick one measure for the whole analysis and keep every study on it.
Fixed effect or random effects
A fixed-effect model assumes every study estimates one shared true effect and differences are chance alone. A random-effects model assumes the true effect varies across studies and estimates the spread as well as the average. When your studies differ in population, dose, or follow-up, which is almost always, expect heterogeneity and use random effects. For the between-study variance you will see two estimators: DerSimonian-Laird (DL) is the classic default, REML is often preferred with few studies. Both are available; if you are unsure, random effects with REML is a defensible choice.
Running it and reading the forest plot
Point-and-click means you paste or upload your table, choose the measure and the model, and read the output. The forest plot is the main result: one row per study with its effect and CI, a box sized by the study's weight, and a diamond at the bottom for the pooled estimate. The diamond's center is the summary effect and its width is the summary CI. Read left to right against the null line (1 for ratios, 0 for differences): rows crossing the line are not individually significant, and a diamond clear of the line is a significant pooled effect.
Heterogeneity without hand-waving
Three numbers describe how much the studies disagree. Cochran's Q tests whether the variation is more than chance; a small p-value says the effects differ. I-squared is the percentage of total variation due to real differences rather than sampling error, so 0 percent is none and 75 percent is substantial. Tau-squared is the actual estimated variance of the true effects, in the units of the effect measure, and it is what the random-effects model uses to widen the interval. Report all three rather than I-squared alone, because I-squared can look small when studies are small and large when they are precise.
Small-study effects and publication bias
Small studies with null results often go unpublished, which biases the pooled effect. A funnel plot puts each study's effect against its precision; symmetry is reassuring, and a gap in one bottom corner suggests missing studies. Egger's test puts a number on that asymmetry. The standing rule of thumb is to skip the funnel plot and Egger's test when you have fewer than 10 studies, because with few points the test has almost no power and asymmetry is hard to judge.
Subgroups, meta-regression, and sensitivity
When heterogeneity is high, ask where it comes from. Subgroup analysis pools studies within groups (for example by risk of bias or by dose) and tests whether the groups differ. Meta-regression relates the effect to a study-level variable, such as mean age or year, and is better than subgrouping when the moderator is continuous. Leave-one-out analysis re-runs the pooled estimate dropping each study in turn, so you can see whether one study is driving the result. Treat these as exploratory unless you pre-specified them.
Reporting it so a reviewer accepts it
Three things carry the writeup. A PRISMA flow diagram shows how you went from records found to studies included. A methods sentence names the measure, the model, the heterogeneity statistics, and the software, for example: "We pooled risk ratios with a random-effects model (REML), reporting Cochran's Q, I-squared, and tau-squared, and assessed small-study effects with a funnel plot and Egger's test." Reviewers increasingly want reproducibility, so download the R replication code for your analysis and include it as a supplement; the numbers were validated against R's metafor, so the code reproduces them.
When you genuinely still need R
Point-and-click covers standard pairwise meta-analysis and the extensions above. Some designs need real code. Multilevel and multivariate models (multiple correlated outcomes per study, or effects nested within studies) need packages like metafor's rma.mv. Dose-response meta-analysis with study-specific trends is a coding job. So are exotic data structures and custom likelihoods. If that is your case, write it in R. If your review is a standard comparison of a treatment against a control, you do not need to. You can run the whole meta-analysis in the browser and hand the R code to your reviewers.