When Does Adding a New Variable Turn a Replication Into an Extension?

If You Add a Variable, Are You Still Replicating the Original Study?

You want to replicate a previous study, but you also want to collect one additional variable. Perhaps you want to measure age, prior experience, socioeconomic status, motivation, self-efficacy, technology acceptance, or another construct that the original researchers did not include.

Does that addition mean your project is no longer a replication?

Not necessarily. Simply collecting another variable does not determine what kind of study you are conducting. What matters is the role that variable plays in the research. It might have no effect on the replication question, it might help you conduct a more defensible test of the original claim, or it might introduce an additional hypothesis that extends the original research.

The Variable Matters When It Changes What You Are Trying to Learn

Classify the Variable by Its Inferential Role, Not by Its Mere Presence

Start With the Claim You Are Replicating

Before thinking about the additional variable, identify exactly what the original study claimed.

Suppose an earlier study reported that students' academic self-efficacy was positively associated with academic persistence. Your replication question might be:

Does the association between academic self-efficacy and academic persistence appear again in new data under the conditions being tested?

That is your inferential anchor. Nosek and Errington argue that replication is better understood through its relationship to a prior claim than through mechanical duplication of every feature of the original study. Under their formulation, a replication produces evidence for which outcomes consistent with the prior claim would increase confidence in it and inconsistent outcomes would decrease confidence in it.

Once the original claim is clear, you can ask whether the additional variable changes that inferential target.

Simply Measuring an Additional Variable Does Not Automatically Create an Extension

Imagine that the original self-efficacy study did not report participants' employment status. In your replication, you collect employment status because you want a more complete description of the sample.

You have added a variable, but you have not necessarily added a research question.

If employment status appears only in the participant description and plays no substantive role in your hypotheses or interpretation, the core study may remain a replication. The same principle can apply to variables collected for administrative purposes, eligibility checks, manipulation checks, data-quality assessments, or other methodological reasons.

The number of columns in your dataset therefore tells you surprisingly little about whether the study is a replication or an extension.

A New Predictor Can Create an Extension When It Introduces a New Relationship

Now suppose the original study examined only the relationship between self-efficacy and persistence. You add perceived instructor support and hypothesize that both self-efficacy and instructor support independently predict persistence.

The original self-efficacy-persistence relationship can still be replicated if you preserve an appropriate test of it. But the proposed relationship between instructor support and persistence was not part of the original claim.

That additional relationship is an extension.

Your project can therefore contain two inferential components:

The study has not ceased to contain replication evidence simply because an extension has been added. It is more accurate to describe the project as combining replication and extension, provided the replication component remains interpretable.

Adding a Moderator Usually Introduces an Additional Claim

A particularly clear example involves moderation.

Suppose the original study reported that X is associated with Y. You add Z and hypothesize that the strength or direction of the X-Y relationship depends on Z.

You are now asking something the original study did not answer:

Under what conditions does the original relationship change?

That is an extension question.

For example, if an earlier study reported that formative feedback improves writing performance, you might replicate that comparison while testing whether students' feedback literacy moderates the effect. The replication component concerns whether the original feedback effect appears again. The extension concerns whether that effect differs according to feedback literacy.

This distinction matters analytically. Testing whether the original effect exists and testing an interaction are not the same statistical question. A nonsignificant interaction does not mean the original effect failed to replicate, and evidence for an interaction does not by itself establish that the original result replicated.

Adding a Mediator Also Moves Beyond Simply Retesting the Original Effect

Mediation creates a similar but distinct situation.

Suppose previous research reports that an instructional intervention improves academic performance. Your new study measures cognitive engagement and proposes that the intervention improves performance because it increases cognitive engagement.

The original claim concerns whether the intervention affects performance. The mediation hypothesis concerns a possible mechanism explaining that effect.

These questions should not be collapsed into one. A study can obtain evidence consistent with the original effect but little support for the proposed mediation model. It can also produce a more complicated pattern in which evidence for the original effect differs from evidence concerning the proposed mechanism.

In either case, the replication and extension claims should be evaluated separately.

An Additional Outcome Can Also Create an Extension

Variables are not limited to predictors, moderators, or mediators. Adding another outcome may also extend the study.

Imagine that the original experiment measured learning performance immediately after an intervention. You reproduce that outcome but also measure retention four weeks later.

The immediate performance measure can contribute to replication of the original finding. The delayed outcome addresses something additional: whether the effect persists over time.

That second outcome is not merely "more data." It permits a new substantive inference.

The same logic applies if an original educational technology study measured achievement and your project additionally examines cognitive load, satisfaction, transfer, persistence, or another outcome. If you interpret that outcome as evidence for an additional substantive claim, the relevant component is an extension.

A Covariate Does Not Automatically Make the Study an Extension

The classification becomes more complicated when an additional variable is used as a covariate or adjustment variable.

Suppose you measure baseline achievement because the design requires adjustment for pre-existing differences between participants. If the purpose is to improve estimation of the original effect rather than investigate a new substantive hypothesis about baseline achievement, the variable does not necessarily create an extension.

However, adding covariates can affect comparability with the original analysis. If the original study estimated an unadjusted relationship and your primary replication result comes from a substantially different adjusted model, differences between the results may partly reflect analytical specification rather than new data alone.

A sensible solution is often to preserve an appropriately comparable analysis of the original claim and then report justified alternative or adjusted analyses separately.

Adding a Variable Can Alter the Original Replication Test If You Change the Model

Consider an original study that estimated a simple relationship between X and Y. You collect X and Y again but add Z, W, and V to the primary model. You then compare your adjusted coefficient for X with the original unadjusted estimate.

Are those estimates testing exactly the same quantity?

Not necessarily.

Statistical coefficients can change when additional variables enter a model, and the interpretation of an adjusted coefficient can differ from that of an unadjusted association. Whether the models are meaningfully comparable depends on the design and substantive context.

Consequently, researchers conducting replication-extension studies should avoid allowing the extension analysis to silently replace the replication analysis.

This separation can make the contribution much easier to understand.

The Variable Can Extend Generalizability Rather Than Introduce a Completely Different Topic

Extensions do not have to abandon the original phenomenon.

Suppose an earlier finding appears consistently in the overall population, but theory suggests that the effect may be weaker among participants with a particular characteristic. Measuring that characteristic and testing the predicted difference can identify a possible boundary condition.

In this sense, the extension deepens understanding of the original claim by asking where or for whom it applies.

Nosek and Errington emphasize that differences between studies can provide information about the conditions under which findings generalize. This also illustrates why the boundary among conceptual replication, generalizability testing, and extension is not universally fixed. Different research traditions may apply these terms differently.

For practical purposes, transparency is more important than winning a terminological argument. State exactly what the new variable is intended to test.

The Same Variable Can Play Different Roles in Different Studies

Consider age. Merely recording participants' ages for sample description does not ordinarily create an extension. Including age as a prespecified adjustment variable may alter the analytical model but still serve the original inferential purpose. Testing whether age moderates the original effect introduces an additional substantive claim.

The classification therefore follows the inferential function of the variable, not its name or measurement scale.

Replication-Extension Designs Can Be Deliberate Rather Than Accidental

There is nothing methodologically suspect about designing a project to do both. Bonett describes replication-extension studies as studies specifically designed to combine new results with prior evidence while both replicating and extending earlier findings.

The advantage of an explicit replication-extension design is conceptual clarity. You can identify in advance which hypothesis confronts the previous claim and which hypothesis goes beyond it.

That is preferable to beginning with a replication, adding several variables because they happen to be available, and later describing whichever relationships are statistically interesting as the study's "extension." The latter strategy risks turning a theoretically motivated extension into a post hoc search through the dataset.

If your additional variable introduces a genuinely new inferential target, the broader project fits naturally within the distinction between replication and extension research.

Adding More Variables Does Not Necessarily Make the Study Better

It is easy to assume that a replication becomes more valuable if you collect additional constructs. More variables appear to offer more possible findings.

They also create costs.

Additional measures can increase participant burden, lengthen surveys or experiments, complicate data management, create additional hypotheses and analytical decisions, and raise multiplicity concerns when many relationships are tested. Moderation and mediation questions may also require design and sample-size considerations beyond those needed for the original replication target.

A new variable should therefore earn its place in the design.

Ask what uncertainty it addresses and whether the study is capable of answering that additional question well. A focused replication plus one well-justified extension can be considerably more informative than a replication surrounded by a small ecosystem of variables collected "just in case."

One Added Variable Can Play Three Very Different Roles

The dataset can therefore be almost identical across these scenarios while the inferential purpose changes substantially. What matters is not that the variable called "AI literacy" appears in the spreadsheet. What matters is what claim you intend to make with it.

Common Mistakes When Adding Variables to a Replication

Give Every Added Variable a Defined Job Before You Collect It

If you are planning to add a variable to a replication, write down why it is there before deciding what label to give the study.

Do not begin with "I need something new." Begin with the uncertainty the variable is intended to address.

For a replication-extension design, state the replication objective and extension objective separately. The hypotheses should make the distinction visible, and the analysis plan should preserve it.

For example, you might first test whether the original X-Y relationship recurs and then test whether Z moderates that relationship. If you preregister the study, the distinction can also be documented before the outcomes are known.

Finally, resist the urge to add a variable simply because it is available in an instrument or dataset. Each additional substantive test creates another claim that needs theoretical justification and appropriate evidence. The fact that a questionnaire already contains twelve subscales is not, regrettably, a research framework.

If your added variable is motivated mainly by a different population or context, consider whether the central issue is actually what the population change means for the status of the study rather than the variable itself.

Before Adding a Variable to Your Replication

Questions About Adding Variables to Replication Studies

A New Variable Becomes an Extension When It Creates a New Claim

How to Cite This Guide

This guide is intended to be read, shared, and used in research, teaching, and academic work. If you draw on its ideas, explanations, or other content, please acknowledge the source by citing the guide. Doing so gives appropriate credit and helps your readers locate the original resource.