main effect insignificant interaction term significant

[rai rai]

Hello Everyone,

I am running a logistic regression and I have around 15 variables (including control variables). I have an interaction term which is significant in the interaction model but the variables that constitute the interaction term are insignificant in the main effects model. Is it unusual to include interaction terms for variables whose main effects in insignificant. My variables are continuous in nature and the Dependent Variable is binary.

Thanks

[Nick Cox]

This member has already rejected the request to use a full real name

http://www.statalist.org/forums/foru...uals-by-groups

so the issue need not be raised again.

My answer is that I can have no idea what is usual or unusual or not in your field, as you do not state it and almost certainly it is not mine too. But I see nothing inherently objectionable in not being a slave to 5% or 1%. You should test the hypotheses that arise in your research. Naturally, reviewers in your field might disagree about conventional levels.

[Carlo Lazzaro]

As an aside, after 10 messages the original poster should be aware of and act on FAQ #12 about how to post what (s)he typed and report what Stata gave her(him) back.

[Nick Cox]

I should add that a more important rule than any mentioned is that the number of data points should be **much** greater than the number of parameters being estimated. Trouble with significance levels may mean at least partly that you just don't have a large enough sample to do well what you want to do.

[Richard Williams]

Coefficients for main effects and their P values have a different meaning once interaction terms are added. Usually you don't care if the main effects are significant or not. For more, see

http://www3.nd.edu/~rwilliam/stats2/l53.pdf

especially the discussion on p. 4.

Since your variables are continuous, you may also wish to look at

http://www3.nd.edu/~rwilliam/stats2/l55.pdf

[rai rai]

Thank you everyone for your insights. My sample has 5671 data points.

[rai rai]

Nick Cox Could you please let me know how to change my username. I have really benefited from this group and its discussions and I do not want other dedicated members to consider me as recalcitrant. My only objective is to learn and share my views, and the pressure to use my real name is becoming too much and hence would want to change it.

[Nick Cox]

Thanks for your willingness to change, but we have already done this! The references in the thread linked in #2 explain how to do it.

[Clyde Schechter]

To change your user name, click on "CONTACT US" (in the lower right hand corner of the window) and, on the page that appears, send a message to the system administrators requesting the name change.

Thank you.

[Carlo Lazzaro]

rai rai:
thanks for this.
Echoing in part the title of one of Nick's most famous article on Stata Journal, below you will find (as per FAQ) how to re-register with your real names -step by step- (bold is mine).

6. Why are real names preferred?



You are asked to post on Statalist using your full real name, including given name(s) and a family name, such as "Ronald Fisher" or "Gertrude M. Cox". Giving full names is one of the ways in which we show respect for others and is a long tradition on Statalist. It is also much easier to get to know people when real names are used.

If you overlook this on first registration, it is easy to fix. Click on “Contact us” located at the bottom right-hand corner of every page.

PS: Crossed in the cyberspace with Clyde's helpful reply.

[Alex Mai]

Coefficients for main effects and their P values have a different meaning once interaction terms are added. Usually you don't care if the main effects are significant or not. For more, see

http://www3.nd.edu/~rwilliam/stats2/l53.pdf

especially the discussion on p. 4.

Since your variables are continuous, you may also wish to look at

http://www3.nd.edu/~rwilliam/stats2/l55.pdf

Dear Richard,

I have read the document about the interpretation of interaction. However, I have a question about the case with more than one interaction terms. For a regression with single interaction term (depvar=c+ax+by+cx*y, where y is a dummy variable and x is continuous), the coefficient c is the difference in x for those y=1 and y=0. So for a regression with two interaction terms (depvar=c+ax+by+cz+dx*y+ez*y, where y is a dummy variable and x and z are continuous), how should I explain the coefficient of the dummy variable (b)?

Than you very much!

[Clyde Schechter]

So for a regression with two interaction terms (depvar=c+ax+by+cz+dx*y+ez*y, where y is a dummy variable and x and z are continuous), how should I explain the coefficient of the dummy variable (b)?

In this model b is the expected difference in depvar associated with a unit change in y conditional on x and z both being 0.