Computes the required relative sample size to achieve significance based on power

The relative sample size to achieve significance of the replication study is computed based on the z-value of the original study, the significance level and the power.

sampleSizeSignificance(
  zo,
  power = NA,
  level = 0.025,
  alternative = c("one.sided", "two.sided"),
  designPrior = c("conditional", "predictive", "EB"),
  h = 0,
  shrinkage = 0
)

Arguments

zo: A vector of z-values from original studies.
power: The power to achieve replication success.
level: Significance level. Default is 0.025.
alternative: Either "one.sided" (default) or "two.sided". Specifies if the significance level is one-sided or two-sided. If the significance level is one-sided, then sample size calculations are based on a one-sided assessment of significance in the direction of the original effect estimate.
designPrior: Is only taken into account when power is specified. Either "conditional" (default), "predictive", or "EB". If "EB", the power is computed under a predictive distribution where the contribution of the original study is shrunken towards zero based on the evidence in the original study (with an empirical Bayes shrinkage estimator).
h: Is only taken into account when power is specified and designPrior is "predictive" or "EB". The relative between-study heterogeneity, i.e., the ratio of the heterogeneity variance to the variance of the original effect estimate. Default is 0 (no heterogeneity).
shrinkage: Is only taken into account when power is specified. A number in [0,1) with default 0. Specifies the shrinkage of the original effect towards zero (e.g., shrinkage = 0.25 implies shrinkage by a factor of 25%). Is only taken into account when designPrior is "conditional" or "predictive".

Value

The relative sample size to achieve significance in the specified direction. If impossible to achieve the desired power for specified inputs NaN is returned.

Details

sampleSizeSignificance is the vectorized version of .sampleSizeSignificance_. Vectorize is used to vectorize the function.

References

Held, L. (2020). A new standard for the analysis and design of replication studies (with discussion). Journal of the Royal Statistical Society: Series A (Statistics in Society), 183, 431-448. doi:10.1111/rssa.12493

Pawel, S., Held, L. (2020). Probabilistic forecasting of replication studies. PLoS ONE. 15, e0231416. doi:10.1371/journal.pone.0231416

Held, L., Micheloud, C., Pawel, S. (2022). The assessment of replication success based on relative effect size. The Annals of Applied Statistics. 16:706-720. doi:10.1214/21-AOAS1502

Micheloud, C., Held, L. (2022). Power Calculations for Replication Studies. Statistical Science. 37:369-379. doi:10.1214/21-STS828

Author

Leonhard Held, Samuel Pawel, Charlotte Micheloud, Florian Gerber

Examples

sampleSizeSignificance(zo = p2z(0.005), power = 0.8)
#> [1] 0.9961217
sampleSizeSignificance(zo = p2z(0.005, alternative = "two.sided"), power = 0.8)
#> [1] 0.9961217
sampleSizeSignificance(zo = p2z(0.005), power = 0.8, designPrior = "predictive")
#> [1] 1.337889

sampleSizeSignificance(zo = 3, power = 0.8, designPrior = "predictive",
                       shrinkage = 0.5, h = 0.25)
#> [1] 18.87111
sampleSizeSignificance(zo = 3, power = 0.8, designPrior = "EB",  h = 0.5)
#> [1] 2.412774

# sample size to achieve  0.8 power as function of original p-value
zo <- p2z(seq(0.0001, 0.05, 0.0001))
oldPar <- par(mfrow = c(1,2))
plot(z2p(zo), sampleSizeSignificance(zo = zo, designPrior = "conditional", power = 0.8),
     type = "l", ylim = c(0.5, 10), log = "y", lwd = 1.5, ylab = "Relative sample size",
     xlab = expression(italic(p)[o]), las = 1)
lines(z2p(zo), sampleSizeSignificance(zo = zo, designPrior = "predictive", power = 0.8),
      lwd = 2, lty = 2)
lines(z2p(zo), sampleSizeSignificance(zo = zo, designPrior = "EB", power = 0.8),
      lwd = 1.5, lty = 3)
legend("topleft", legend = c("conditional", "predictive", "EB"),
       title = "Design prior", lty = c(1, 2, 3), lwd = 1.5, bty = "n")

par(oldPar)