*Replication seems a sensible way to assess whether a scientific result is right. The intuition is clear: if a result is right, you should get a significant result when repeating the work; if it it’s wrong, the result should be non-significant. I test this intuition across a range of conditions using simulation. For exact replications, the intuition is dead on, but when replicas diverge from the original studies, error rates increase rapidly. Even for the exact case, false negative rates are high for small effects unless the samples are large. These results bode ill for large, systematic replication efforts, which typically prioritize uniformity over fidelity and limit sample sizes to run lots of studies at reasonable cost.*

The basic replication rationale goes something like this: (1) many published papers are wrong; (2) this is a serious problem the community must fix; and (3) systematic replication is an effective solution. (In recent months, I’ve seen an uptick in pre-registration as another solution. That’s a topic for another day.) In this post, I focus on the third point and ask: viewed as a statistical test, how well does systematic replication work; how well does it tell the difference between valid and invalid results?

By “systematic replication” I mean projects like Many Lab, Reproducibility Project: Psychology (RPP), Experimental Economics Replication Project (EERP), and Social Sciences Replication Project (SSRP) that systematically select studies in a particular field and repeat them in a uniform fashion. The main publications for these projects are Many Lab, RPP, EERP, SSRP.

I consider a basic replication scheme in which each original study is repeated once. This is like RPP and EERP, but unlike Many Lab as published which repeated each study 36 times and SSRP which used a two-stage replication strategy. I imagine that the replicators are trying to closely match the original study (*direct replication*) while doing the replications in a uniform fashion for cost and logistical reasons.

My test for replication success is the same as SSRP (what they call the *statistical significance criterion*): a replication succeeds if the replica has a significant effect in the same direction as the original.

A replication is *exact* if the two studies are sampling the same population. This is an obvious replication scenario. You have a study you think may be wrong; to check it out, you repeat the study, taking care to ensure that the replica closely matches the original. Think cold fusion. A replication is *near-exact* if the populations differ slightly. This is probably what systematic replication achieves, since the need for uniformity reduces precision.

Significance testing of the replica (more precisely, the statistical significance criterion) works as expected for exact replications, but error rates increase rapidly as the populations diverge. This isn’t surprising when you think about it: we’re using the replica to draw inferences about the original study; it stands to reason this will only work if the two studies are very similar.

Under conditions that may be typical in systematic replication projects, the rate of false positive mistakes calculated in this post ranges from 1-71% and false negative mistakes from 0-85%. This enormous range results from the cumulative effect of multiple unknown, hard-to-estimate parameters.

My results suggest that we should adjust our expectations for systematic replication projects. These projects may make a lot of mistakes; we should take their replication failure rates with a grain of salt.

The software supporting this post is open source and freely available in GitHub.

The software simulates *studies* across a range of conditions, combines pairs of studies into *pairwise replications*, calculates which replications pass the test, and finally computes false positive and false negative rates for conditions of interest.

The studies are simple two group comparisons parameterized by sample size \(n\) and population effect size \(d_{pop}\) (\(d_{pop}\ge0\)). For each study, I generate two groups of \(n\) random numbers. One group comes from a standard normal distribution with \(mean=0\); the other is standard normal with \(mean=d_{pop}\). I then calculate the p-value from a t-test. When I need to be pedantic, I use the term *study set* for the ensemble of studies for a given combination of \(n\) and \(d_{pop}\).

The program varies \(n\) from 20 to 500 and \(d_{pop}\) from 0 to 1 with 11 discrete values each (a total of \(11^2=121\) combinations). It simulates \(10^4\) studies for each combination yielding about 1.2 million simulated studies. An important limitation is that all population effect sizes are equally likely within the range studied. I don’t consider *publication bias* which may make smaller effect sizes more likely, or any prior knowledge of expected effect sizes.

To generate pairwise replications, I consider all (ordered) pairs of study sets. For each pair, the software permutes the studies of each set, then combines the studies row-by-row. This multiplies out to \(121^2=14,641\) pairs of study sets and almost 150 million simulated replications. The first study of the pair is the *original* and the second the *replica*. I consistently use the suffixes \(1\) and \(2\) to denote the original and replica respectively.

Four variables parameterize each pairwise replication: \(n1\), \(n2\), \(d1_{pop}\), and \(d2_{pop}\). These are the sample and population effect sizes for the two studies.

After forming the pairwise replications, the program discards replications for which the original study isn’t significant. This reflects the standard practice that non-significant findings aren’t published and thus aren’t candidates for systematic replication.

Next the program determines which replications should pass the replication test and which do pass the test. The ones that *should pass* are ones where the original study is a true positive, i.e., \(d1_{pop}\ne0\). The ones that *do pass* are ones where the replica has a significant p-value and effect size in the same direction as the original.

A *false positive replication* is one where the original study is a false positive (\(d1_{pop}=0\)) yet the replication passes the test. A *false negative replication* is one where the original study is a true positive (\(d1_{pop}\ne0\)), yet the replication fails the test. The program calculates *false positive* and *false negative rates* (abbr. *FPR* and *FNR*) relative to the number of replications in which the original study is significant.

My definition of which replications *should pass* depends only on the original study. A replication in which the original study is a false positive and the replica study a true positive counts as a false positive replication. This makes sense if the overarching goal is to validate the original *study*. If the goal were to test the *result* of the original study rather than the study itself, it would make sense to count this case as correct.

To get “mistake rates” I need one more parameter: \(prop.true\), the proportion of replications that are true. This is the issue raised in Ioannidis’s famous paper, “Why most published research findings are false” and many other papers and blog posts including one by me. The terminology for “mistake rates” varies by author. I use terminology adapted from Jager and Leek. The *replication-wise false positive rate* (*RWFPR*) is the fraction of positive results that are false positives; the *replication-wise false negative rate* (*RWFNR*) is the fraction of negative results that are false negatives.

A replication is *exact* if the two studies are sampling the same population; this means \(d1_{pop}=d2_{pop}\).

Figure 1 shows FPR for \(n1=20\) and \(n2\) varying from 50 to 500. The x-axis shows all four parameters using \(d1\), \(d2\) as shorthand for \(d1_{pop}\), \(d2_{pop}\). \(d1_{pop}=d2_{pop}=0\) throughout because this is the only way to get false positives with exact replications. Figure 2 shows FNR for the same values of \(n1\) and \(n2\) but with \(d1_{pop}=d2_{pop}\) ranging from 0.1 to 1.

I mark the conventionally accepted thresholds for false positive and negative error rates (0.05 and 0.2, resp.) as known landmarks to help interpret the results. I do **not** claim these are the right thresholds for replications.