Skip to main content
Cross Validated

Return to Answer

forgot to write "ago"
Source Link
Orielno
Orielno
  • 116
  • 4

Using interaction (product) terms in regressions without using curvilinear (quadratic) terms.

A few years ago I've been thinking about it (after seeing a few papers (in economic/management fields) that were doing it), and I realized that if in the true model the outcome variable depends on the square of some or all the variables in the model, yet those are not included and instead an interaction is included in the examined model, the researcher may find that the interaction has an effect, while in fact it does not.

I then searched to see if there is an academic paper that addressed this, and I did find one (could be more, but that is what I found): https://psycnet.apa.org/fulltext/1998-04950-001.html

You might say that it is a novice mistake, and that a real statistician should know to first try to include all terms and interactions of a certain degree in the regression. But still, this specific mistake seems to be quite common in many fields that apply statistics, and the above linked article demonstrates the misleading results it may lead to.

Using interaction (product) terms in regressions without using curvilinear (quadratic) terms.

A few years I've been thinking about it (after seeing a few papers (in economic/management fields) that were doing it), and I realized that if in the true model the outcome variable depends on the square of some or all the variables in the model, yet those are not included and instead an interaction is included in the examined model, the researcher may find that the interaction has an effect, while in fact it does not.

I then searched to see if there is an academic paper that addressed this, and I did find one (could be more, but that is what I found): https://psycnet.apa.org/fulltext/1998-04950-001.html

You might say that it is a novice mistake, and that a real statistician should know to first try to include all terms and interactions of a certain degree in the regression. But still, this specific mistake seems to be quite common in many fields that apply statistics, and the above linked article demonstrates the misleading results it may lead to.

Using interaction (product) terms in regressions without using curvilinear (quadratic) terms.

A few years ago I've been thinking about it (after seeing a few papers (in economic/management fields) that were doing it), and I realized that if in the true model the outcome variable depends on the square of some or all the variables in the model, yet those are not included and instead an interaction is included in the examined model, the researcher may find that the interaction has an effect, while in fact it does not.

I then searched to see if there is an academic paper that addressed this, and I did find one (could be more, but that is what I found): https://psycnet.apa.org/fulltext/1998-04950-001.html

You might say that it is a novice mistake, and that a real statistician should know to first try to include all terms and interactions of a certain degree in the regression. But still, this specific mistake seems to be quite common in many fields that apply statistics, and the above linked article demonstrates the misleading results it may lead to.

Source Link
Orielno
Orielno
  • 116
  • 4

Using interaction (product) terms in regressions without using curvilinear (quadratic) terms.

A few years I've been thinking about it (after seeing a few papers (in economic/management fields) that were doing it), and I realized that if in the true model the outcome variable depends on the square of some or all the variables in the model, yet those are not included and instead an interaction is included in the examined model, the researcher may find that the interaction has an effect, while in fact it does not.

I then searched to see if there is an academic paper that addressed this, and I did find one (could be more, but that is what I found): https://psycnet.apa.org/fulltext/1998-04950-001.html

You might say that it is a novice mistake, and that a real statistician should know to first try to include all terms and interactions of a certain degree in the regression. But still, this specific mistake seems to be quite common in many fields that apply statistics, and the above linked article demonstrates the misleading results it may lead to.

Post Made Community Wiki by Orielno