Incongruity between t.test result and sample size estimation?

Question

Not sure if I'm misunderstanding the pwr::pwr.t.test() function and result.

set.seed(1299448)
g1 <- rnorm(35, 12, 3)
g2 <- rnorm(35, 14, 3)
t.test(g1, g2)
pwr.t.test(d = ((mean(as.numeric(g1), na.rm = T) -
                 mean(as.numeric(g2), na.rm = T)) /
                 sqrt(((sd(g1)^2)+ (sd(g2)^2))/2)),
           power = .80,
           sig.level = .05,
           alternative = "two.sided")

Results

Welch Two Sample t-test

data:  g1 and g2
t = -2.1738, df = 63.822, p-value = 0.03343


Two-sample t test power calculation 

          n = 59.10848
          d = 0.519647
  sig.level = 0.05
      power = 0.8
alternative = two.sided

NOTE: n is number in *each* group

So the effect is significant with a sample size of 70, but with such an effect size, a sample size of 120 would be required for 80% power.

Can someone ELI5? Sorry if this is more of a cross-validated question. I started originally thinking this was a programming/misunderstanding of the pwr package, but now I'm starting to think I just don't understand the sample size estimation/power well enough.

Answer 1

The issue is that the effect size for that particular realization of the simulation is different than the effect size of the model that generated it, and just because the effect is strong enough to produce a p-value less than 0.05 one time, doesn't mean that it's strong enough to create a p-value less than 0.05 80% of the time:

set.seed(1299448)
g1 <- rnorm(35, 12, 3)
g2 <- rnorm(35, 14, 3)
(eff <- ((mean(as.numeric(g1), na.rm = T) -
           mean(as.numeric(g2), na.rm = T)) /
          sqrt(((sd(g1)^2)+ (sd(g2)^2))/2)))
# [1] -0.519647

n.samp <- 10000
out70 <- rep(NA,n.samp)
out120 <- rep(NA,n.samp)

for(i in 1:n.samp) {
  g1 <- rnorm(35, 0)
  g2 <- rnorm(35, eff)
  out70[i] <- t.test(g1, g2)$p.value<0.05
  g1 <- rnorm(60, 0)
  g2 <- rnorm(60, eff)
  out120[i] <- t.test(g1, g2)$p.value<0.05
}

mean(out70)
# [1] 0.5682
mean(out120)
# [1] 0.8109