Table 1.
Different methods for meta-analysis in the genome-wide association setting
| Issues and caveats | P-value meta-analysis | Effect size meta-analysis | |
|---|---|---|---|
| Fixed effects | Random effects | ||
| Direction of effect is considered | In some methods | Yes | Yes |
| Effect size is considered | No | Yes | Yes |
| Summary p-value is obtained | Yes | Yes | Yes |
| Summary effect is obtained | No | Yes | Yes |
| Summary result can be converted to credibility based on priors for the anticipated effect sizes | No | Yes | Yes |
| Between-study heterogeneity can be taken into account | No | No | Yes |
| Between-study heterogeneity can be estimated/tested | No | Yes | Yes |
| Consensus on if/how datasets should be weighted | No | Yes | Yes |
| Commonly used weights | None, SQRT(N), N | Inverse variance | Inverse variance |
| Prior assumptions on the effect size can be used | No | In Bayesian meta-analysis | In Bayesian meta-analysis |
| Prior uncertainty on heterogeneity can be accommodated | No | No | In Bayesian meta-analysis |
| Prior uncertainty on the genetic model can be accommodated | No | In Bayesian M-A | In Bayesian meta-analysis |
| Normality assumptions typically made within each study | Yes | Yes | Yes |
| Normality assumptions within each study easily testable | Yes, rarely done | Yes, rarely done | |
| Normality assumptions for distribution of effects across studies easily testable | No effects assumed | Single common effect assumed(assumption may be visibly wrong) | Not easily testable |
| Heavy-tail alternative methods exist | No | Yes, rarely used | Yes, rarely used |
| Use with uncommon alleles (small genotype groups, or even zero allele counts in 2 ×2 tables) | Need to use exact methods | Quite robust | Between-study variance estimation unstable |
| Power for discovery | Good | Good | Less than others |
| False-positives from single biased dataset | Susceptible | Susceptible | Less susceptible |
| False-positives when evidence from small studies is most biased | Susceptible | Susceptible | More susceptible |
| False-positives when evidence from large studies is most biased | Susceptible | Susceptible | Less susceptible |
| Can predict range of effect sizes in future similar populations | No | Too narrow confidence intervals | Appropriate with predictive intervals |
| Can convey uncertainty for practical applications (e.g. to be used in clinical prediction test) | Useless | Inappropriate | Most appropriate with prediction intervals |