Visualize meta-regression results with precision-weighted bubble plots, fitted regression lines, and 95% confidence bands to investigate sources of between-study heterogeneity.
Load sample data to see how the tool works, or clear all fields to start fresh.
Drag & drop a file or
CSV, TSV, Excel (.xlsx/.xls) - max 500 rows
| Study | Effect Size | Standard Error | Publication Year | |
|---|---|---|---|---|
Input the effect size, standard error, and moderator value for each study. The moderator can be any continuous study-level covariate such as publication year, mean participant age, or percentage female. Import from CSV or paste from a spreadsheet.
Name your moderator variable to label the x-axis correctly. The tool fits a weighted least squares regression with inverse-variance weights, producing intercept, slope, and R-squared estimates automatically.
Click Generate to render the D3.js bubble plot. Each study appears as a circle sized by its precision. The fitted regression line and 95% confidence band overlay the scatter to visualize the moderator-effect relationship.
Review the slope (b1) value and its p-value to determine whether the moderator significantly predicts variation in effect sizes. The confidence band shows the uncertainty around the predicted regression line at each covariate level.
Examine the R-squared analog to see what proportion of between-study variance the moderator explains. Check residual tau-squared to understand how much unexplained heterogeneity remains after accounting for the moderator.
Download the bubble plot as PNG, SVG, or PDF for your manuscript. Copy the auto-generated methods paragraph for your statistical analysis section, or export the R code to reproduce the analysis with metafor regplot in RStudio.
Need this done professionally? Get a complete systematic review or meta-analysis handled end-to-end.
Get a Free QuoteEach circle on the bubble plot is sized proportionally to the study's precision (1/SE-squared). Larger bubbles represent studies with smaller standard errors that receive greater weight in the meta-regression model, ensuring the visual display matches the statistical influence of each study.
The regression slope (b1) quantifies the predicted change in effect size per one-unit increase in the moderator variable. A statistically significant slope (p < 0.05) suggests that the covariate explains some of the observed between-study heterogeneity in your meta-analysis.
The 95% confidence band around the regression line reflects uncertainty in the true predicted effect at each moderator value. A narrow band indicates a precisely estimated relationship, while a wide band indicates that more studies or less variable data would be needed to confirm the trend.
The R-squared analog in meta-regression represents the proportion of between-study variance (tau-squared) that the moderator explains. An R-squared of 50% means the covariate accounts for half of the heterogeneity. The remaining variance is residual heterogeneity not captured by your model.
Meta-regression operates at the study level. A relationship between mean participant age and effect size does not prove that older individuals benefit more. Study-level associations can be confounded by other correlated characteristics. Only individual patient data meta-analysis can test patient-level moderator effects.
The Cochrane Handbook recommends at least 10 studies per moderator variable tested. With fewer studies, meta-regression has low statistical power and a high risk of producing spurious findings. Pre-specify moderators in your PROSPERO protocol to avoid post hoc data-dredging.
When a meta-analysis reveals substantial between-study heterogeneity, meta-regression models the association between study-level covariates and effect sizes to identify potential explanatory factors. The bubble plot is the standard graphical summary for continuous moderators, formalized by Thompson and Higgins (2002) and recommended by the Cochrane Handbook (Chapter 10). Each study is displayed as a circle sized by its precision (1/SE-squared), the weighted least squares regression line shows the predicted effect at each covariate value, and the 95% confidence band quantifies uncertainty in that prediction.
The slope (b1) indicates the predicted change in effect size per one-unit change in the moderator, with statistical significance assessed via a z-test. The R-squared analog quantifies the proportion of between-study variance explained. A high R-squared paired with a significant slope suggests the moderator is a strong predictor of heterogeneity. However, meta-regression operates at the study level and is subject to the ecological fallacy: a relationship between mean age and effect size does not establish that older patients benefit more individually. Only individual patient data meta-analysis can test individual-level moderator effects (Borenstein et al., 2009). The Cochrane Handbook recommends at least 10 studies per moderator, and pre-specifying moderators in the protocol (e.g., during PROSPERO registration) guards against data-dredging.
Moderator analysis through bubble plots should be approached with caution regarding multiple testing. When researchers test many potential moderators without pre-specification, the probability of finding at least one spurious significant result increases substantially. Thompson and Higgins (2002) recommended limiting the number of moderators tested, reporting all analyses performed (not just significant ones), and interpreting results as exploratory unless the moderator was specified in the review protocol. Permutation tests for meta-regression provide more accurate p-values than standard z-tests when the number of studies is small relative to the number of moderators examined.
The visual interpretation of a bubble plot involves assessing several features simultaneously. The direction and steepness of the regression line indicate the strength and direction of the moderator-effect relationship. The width of the 95% confidence band reflects precision, narrowing where data points are concentrated and widening at the extremes of the moderator range. Explained heterogeneity is quantified by the R-squared analog, but residual heterogeneity (tau-squared residual) shows how much unexplained variability remains. If residual tau-squared is still large after including the moderator, additional unmeasured factors likely contribute to between-study differences.
Ecological bias represents the most important interpretive limitation of bubble plots and meta-regression. Because each data point represents a study (not a patient), the observed relationship between a moderator and effect size may not reflect a causal mechanism at the individual level. For example, studies with higher mean age may also differ in disease severity, treatment adherence, comorbidity burden, or measurement methods. These confounders cannot be separated from the moderator effect in an aggregate-level analysis. Borenstein et al. (2009) emphasized that meta-regression findings should be framed as hypothesis-generating associations that require confirmation through individual patient data analyses or prospective studies designed to test the moderator directly.
Bubble plots work best within a broader analytic strategy. Begin by visualizing overall results with our forest plot generator and quantifying heterogeneity with I-squared and tau-squared. Assess potential publication bias with our funnel plot and publication bias tool, as funnel asymmetry can sometimes reflect a moderator effect rather than selective publication. Identify influential studies driving heterogeneity with the GOSH plot generator before attributing variability to a moderator. The R code generated by this tool uses the metafor package (Viechtbauer, 2010) with rma() and regplot() for full reproducibility. Format multi-variable models with our meta-regression input formatter.
A bubble plot is the standard graphical method for displaying meta-regression results with a continuous moderator variable. Each included study is represented as a circle (bubble) positioned at its covariate value on the x-axis and its effect size on the y-axis. The size of each bubble is proportional to the study's precision, typically defined as the inverse of the squared standard error (1/SE-squared). Larger bubbles represent more precisely estimated studies, which receive greater weight in the regression. A fitted regression line is overlaid to show the predicted relationship between the moderator and effect size, usually accompanied by a 95% confidence band. This visualization was formalized by Thompson and Higgins (2002) as the primary tool for interpreting continuous moderator effects in meta-regression.
The meta-regression line on a bubble plot is fitted using weighted least squares (WLS), where each study's contribution is weighted by its precision. In a fixed-effect framework, the weight for each study equals 1/SE-squared (the inverse of the sampling variance). The regression minimizes the weighted sum of squared deviations between observed and predicted effect sizes. This produces intercept (b0) and slope (b1) estimates, along with a standard error for the slope that is used to test whether the moderator significantly predicts variation in effect sizes. In a random-effects (mixed-effects) framework, weights are 1/(SE-squared + tau-squared), where tau-squared represents the residual between-study variance not explained by the moderator.
The slope (b1) of the regression line indicates the predicted change in effect size for a one-unit increase in the moderator variable. For example, if the moderator is publication year and b1 = 0.03, the predicted effect size increases by 0.03 units for each additional year. The p-value tests whether this slope differs significantly from zero. The R-squared analog represents the proportion of between-study variance explained by the moderator. An R-squared of 40% means the moderator accounts for 40% of the heterogeneity observed across studies. Both metrics should be interpreted alongside clinical context, as statistical significance does not imply clinical importance, and study-level associations may not reflect individual-level effects.
The widely cited guideline is at least 10 studies per moderator tested in the meta-regression model. With fewer than 10 studies, meta-regression has low statistical power and a high risk of spurious findings. If you want to test two moderators simultaneously, you need at least 20 studies. For a single continuous moderator (the typical bubble plot scenario), 10 studies is the practical minimum. With 3 to 9 studies, you can still generate a bubble plot for visual exploration, but interpret the regression results as hypothesis-generating rather than confirmatory. The Cochrane Handbook (Chapter 10) emphasizes that meta-regression with few studies should be pre-specified in the protocol and findings should be reported as exploratory.
The ecological fallacy is the incorrect assumption that associations observed at the study level apply to individuals within those studies. For example, if studies with a higher mean participant age show larger treatment effects, this does not necessarily mean that older individuals benefit more from the treatment. The observed pattern could reflect confounding with other study characteristics, such as more severe baseline disease in studies recruiting older participants. Meta-regression can only identify between-study associations, not within-study causal mechanisms. Only individual patient data (IPD) meta-analysis can test individual-level moderator effects. When interpreting bubble plots, always acknowledge this limitation and frame findings as exploratory ecological associations rather than causal conclusions.
A bubble plot, funnel plot, and forest plot serve different purposes in meta-analysis. A forest plot displays individual study estimates and the pooled effect, showing what the studies found. A funnel plot assesses publication bias by plotting effect sizes against their precision, looking for asymmetry. A bubble plot visualizes meta-regression results, showing how a moderator variable relates to effect sizes across studies. The x-axis represents the moderator (not precision), and bubble sizes reflect study weight. Unlike funnel plots, bubble plots include a regression line and confidence band. They answer the question of why studies differ, rather than whether they agree (forest plot) or whether results might be missing (funnel plot).
Visualize individual study estimates and pooled effects with our forest plot generator for meta-analysis. Assess publication bias and small-study effects with our funnel plot and publication bias tool. Structure moderator variables for multi-covariate models and export formatted code for R, Stata, or CMA using our meta-regression input formatter.
Reviewed by
Dr. Sarah Mitchell holds a PhD in Biostatistics from Johns Hopkins Bloomberg School of Public Health and has over 15 years of experience in systematic review methodology and meta-analysis. She has authored or co-authored 40+ peer-reviewed publications in journals including the Journal of Clinical Epidemiology, BMC Medical Research Methodology, and Research Synthesis Methods. A former Cochrane Review Group statistician and current editorial board member of Systematic Reviews, Dr. Mitchell has supervised 200+ evidence synthesis projects across clinical medicine, public health, and social sciences. She reviews all Research Gold tools to ensure statistical accuracy and compliance with Cochrane Handbook and PRISMA 2020 standards.
Whether you have data that needs writing up, a thesis deadline approaching, or a full study to run from scratch, we handle it. Average turnaround: 2-4 weeks.