Dataset Information

Bias caused by sampling error in meta-analysis with small sample sizes.

ABSTRACT:

Background

Meta-analyses frequently include studies with small sample sizes. Researchers usually fail to account for sampling error in the reported within-study variances; they model the observed study-specific effect sizes with the within-study variances and treat these sample variances as if they were the true variances. However, this sampling error may be influential when sample sizes are small. This article illustrates that the sampling error may lead to substantial bias in meta-analysis results.

Methods

We conducted extensive simulation studies to assess the bias caused by sampling error. Meta-analyses with continuous and binary outcomes were simulated with various ranges of sample size and extents of heterogeneity. We evaluated the bias and the confidence interval coverage for five commonly-used effect sizes (i.e., the mean difference, standardized mean difference, odds ratio, risk ratio, and risk difference).

Results

Sampling error did not cause noticeable bias when the effect size was the mean difference, but the standardized mean difference, odds ratio, risk ratio, and risk difference suffered from this bias to different extents. The bias in the estimated overall odds ratio and risk ratio was noticeable even when each individual study had more than 50 samples under some settings. Also, Hedges' g, which is a bias-corrected estimate of the standardized mean difference within studies, might lead to larger bias than Cohen's d in meta-analysis results.

Conclusions

Cautions are needed to perform meta-analyses with small sample sizes. The reported within-study variances may not be simply treated as the true variances, and their sampling error should be fully considered in such meta-analyses.

SUBMITTER: Lin L

PROVIDER: S-EPMC6136825 | biostudies-literature |

REPOSITORIES: biostudies-literature

ACCESS DATA

Similar Datasets

Project description:BackgroundEstimates of under-5 mortality at the national level for countries without high-quality vital registration systems are routinely derived from birth history data in censuses and surveys. Subnational or stratified analyses of under-5 mortality could also be valuable, but the usefulness of under-5 mortality estimates derived from birth histories from relatively small samples of women is not known. We aim to assess the magnitude and direction of error that can be expected for estimates derived from birth histories with small samples of women using various analysis methods.MethodsWe perform a data-based simulation study using Demographic and Health Surveys. Surveys are treated as populations with known under-5 mortality, and samples of women are drawn from each population to mimic surveys with small sample sizes. A variety of methods for analyzing complete birth histories and one method for analyzing summary birth histories are used on these samples, and the results are compared to corresponding true under-5 mortality. We quantify the expected magnitude and direction of error by calculating the mean error, mean relative error, mean absolute error, and mean absolute relative error.ResultsAll methods are prone to high levels of error at the smallest sample size with no method performing better than 73% error on average when the sample contains 10 women. There is a high degree of variation in performance between the methods at each sample size, with methods that contain considerable pooling of information generally performing better overall. Additional stratified analyses suggest that performance varies for most methods according to the true level of mortality and the time prior to survey. This is particularly true of the summary birth history method as well as complete birth history methods that contain considerable pooling of information across time.ConclusionsPerformance of all birth history analysis methods is extremely poor when used on very small samples of women, both in terms of magnitude of expected error and bias in the estimates. Even with larger samples there is no clear best method to choose for analyzing birth history data. The methods that perform best overall are the same methods where performance is noticeably different at different levels of mortality and lengths of time prior to survey. At the same time, methods that perform more uniformly across levels of mortality and lengths of time prior to survey also tend to be among the worst performing overall.

Project description:BackgroundMediation analysis can be used to evaluate the effect of an exposure on an outcome acting through an intermediate variable or mediator. For studies with small sample sizes, permutation testing may be useful in evaluating the indirect effect (i.e., the effect of exposure on the outcome through the mediator) while maintaining the appropriate type I error rate. For mediation analysis in studies with small sample sizes, existing permutation testing methods permute the residuals under the full or alternative model, but have not been evaluated under situations where covariates are included. In this article, we consider and evaluate two additional permutation approaches for testing the indirect effect in mediation analysis based on permutating the residuals under the reduced or null model which allows for the inclusion of covariates.MethodsSimulation studies were used to empirically evaluate the behavior of these two additional approaches: (1) the permutation test of the Indirect Effect under Reduced Models (IERM) and (2) the Permutation Supremum test under Reduced Models (PSRM). The performance of these methods was compared to the standard permutation approach for mediation analysis, the permutation test of the Indirect Effect under Full Models (IEFM). We evaluated the type 1 error rates and power of these methods in the presence of covariates since mediation analysis assumes no unmeasured confounders of the exposure-mediator-outcome relationships.ResultsThe proposed PSRM approach maintained type I error rates below nominal levels under all conditions, while the proposed IERM approach exhibited grossly inflated type I rates in many conditions and the standard IEFM exhibited inflated type I error rates under a small number of conditions. Power did not differ substantially between the proposed PSRM approach and the standard IEFM approach.ConclusionsThe proposed PSRM approach is recommended over the existing IEFM approach for mediation analysis in studies with small sample sizes.

Project description:BACKGROUND: Supervised learning for classification of cancer employs a set of design examples to learn how to discriminate between tumors. In practice it is crucial to confirm that the classifier is robust with good generalization performance to new examples, or at least that it performs better than random guessing. A suggested alternative is to obtain a confidence interval of the error rate using repeated design and test sets selected from available examples. However, it is known that even in the ideal situation of repeated designs and tests with completely novel samples in each cycle, a small test set size leads to a large bias in the estimate of the true variance between design sets. Therefore different methods for small sample performance estimation such as a recently proposed procedure called Repeated Random Sampling (RSS) is also expected to result in heavily biased estimates, which in turn translates into biased confidence intervals. Here we explore such biases and develop a refined algorithm called Repeated Independent Design and Test (RIDT). RESULTS: Our simulations reveal that repeated designs and tests based on resampling in a fixed bag of samples yield a biased variance estimate. We also demonstrate that it is possible to obtain an improved variance estimate by means of a procedure that explicitly models how this bias depends on the number of samples used for testing. For the special case of repeated designs and tests using new samples for each design and test, we present an exact analytical expression for how the expected value of the bias decreases with the size of the test set. CONCLUSION: We show that via modeling and subsequent reduction of the small sample bias, it is possible to obtain an improved estimate of the variance of classifier performance between design sets. However, the uncertainty of the variance estimate is large in the simulations performed indicating that the method in its present form cannot be directly applied to small data sets.

Project description:Limited sample sizes can lead to spurious modeling findings in biomedical research. The objective of this work is to present a new method to generate synthetic populations (SPs) from limited samples using matched case-control data (n = 180 pairs), considered as two separate limited samples. SPs were generated with multivariate kernel density estimations (KDEs) with unconstrained bandwidth matrices. We included four continuous variables and one categorical variable for each individual. Bandwidth matrices were determined with Differential Evolution (DE) optimization by covariance comparisons. Four synthetic samples (n = 180) were derived from their respective SPs. Similarity between observed samples with synthetic samples was compared assuming their empirical probability density functions (EPDFs) were similar. EPDFs were compared with the maximum mean discrepancy (MMD) test statistic based on the Kernel Two-Sample Test. To evaluate similarity within a modeling context, EPDFs derived from the Principal Component Analysis (PCA) scores and residuals were summarized with the distance to the model in X-space (DModX) as additional comparisons. Four SPs were generated from each sample. The probability of selecting a replicate when randomly constructing synthetic samples (n = 180) was infinitesimally small. MMD tests indicated that the observed sample EPDFs were similar to the respective synthetic EPDFs. For the samples, PCA scores and residuals did not deviate significantly when compared with their respective synthetic samples. The feasibility of this approach was demonstrated by producing synthetic data at the individual level, statistically similar to the observed samples. The methodology coupled KDE with DE optimization and deployed novel similarity metrics derived from PCA. This approach could be used to generate larger-sized synthetic samples. To develop this approach into a research tool for data exploration purposes, additional evaluation with increased dimensionality is required. Moreover, given a fully specified population, the degree to which individuals can be discarded while synthesizing the respective population accurately will be investigated. When these objectives are addressed, comparisons with other techniques such as bootstrapping will be required for a complete evaluation.