I don't think you have provided enough information about your research question, variables, and methods of analysis to answer specifically. In general, however, if there is in reality a relationship that involves interaction, omitting the interaction from the model can suppress the true relationship and yield a non-significant result. In that circumstance, it is certainly better to report a model with interaction than without.

First of all, it is not necessarily to have a significant difference, secondly, the nonsignificant difference has a good meaning too, thirdly, if you think that there should be a significant difference then raise the probability to 10% , and that will mean that there is a significant difference at least 90 times out of each 100 observations. Regards.

Dear Professor Burke D. Grandjean

Thanks for your answer. let me make it little elaborate. I am looking the effect of gender role attitude on Labor force participation. I have three gedner role attitudes( sub group 3 in each). Directly they donot produce significant rtesult. I divided them into two groups( say High and low). Now I am using an interaction term of all high attitudes. The interaction of three high attitudes produces negative impact on labour force participation, but individually they don't. My question is: Do I follow the right approach to generate an interactive variable?.

Dear Medhat Elsahookie,

What agrugent I will give if I raise the significance level to 10%, when the standard is 5%?

Thanks for answer.

Here is my understanding of what you did: Your dependent variable is labor force participation (LFP). If LFP is a 2-category variable (yes, in the labor force vs. no, not in the labor force), I assume you are running binary logistic regression models. If it is a continuous variable (such as number of hours employed for pay per week or per year), then I assume you are running ordinary least squares multiple regression. You have 3 predictor variables, each measuring a different aspect of gender role attitudes (and each one being a scale made up of 3 specific items from a questionnaire, probably Likert-type items). I will call these Att1, Att2, and Att3. You have also created 3 other variables by splitting each of the attitude scales into a 2 categories, hi/lo. I will call these Hi1, Hi2, and Hi3. And you created an interaction term, a 2-category variable scored 1 if Hi1, Hi2 and Hi3 are all hi and scored zero if at least one of those is lo. I will call it AllHi. Do I have all that approximately correct?

You ran a model predicting LFP from Att1, Att2, and Att3 and found no significant effects. Some possible reasons include: the theory predicting effects does not work in the population from which your sample was drawn; your sample is too small to show significance when effects exist but are weak; your 3 attitude scales have low measurement reliability and/or low validity (have you examined their scale properties?), the 3 attitudes are highly correlated with each other so that when all 3 are in the model together, none of them uniquely explains enough to be significant (have you checked for multicollinearity? is a bivariate model of LFP with any one of the attitudes, alone, significant? is the 3-variable model as a whole significant, even though none of the individual predictors is?)

By the way, if the direction (positive or negative) of the non-significant effects in the 3-variable model is consistent with your theoretical expectation, then you could justify using a one-tailed significance test. That would allow you to divide by 2 the p-values given by your software; in other words, significance at the 10% level in the default 2-tailed test can be interpreted as significance at the 5% level if a 1-tailed test is appropriate.

It sounds like what you are calling your interaction model is just a bivariate model predicting LFP from AllHi. That is an unconventional approach, and I would be reluctant to use it. Usually, an "interaction" model would include at least 2 different predictors plus an additional term formed by multiplying the predictors together. A 2-way interaction would include the product of 2 predictors. A 3-way interaction would include the product of 3 predictors, and usually a 3-way interaction model would also include all possible 2-way interactions between pairs of predictors. It sounds like you have included only AllHi (which is the product of Hi1 x Hi2 x Hi3), without also including Hi1, Hi2, and Hi3 as individual predictors and without any of the 2-way terms.

Your result with AllHI may suggest that the 3 attitude scales are not measuring 3 different things, but are better viewed as all measuring the same underlying attitude toward gender roles. That would produce multicollinearity when you try to include them all as separate predictors. You might consider examining the scale reliability and validity of a single, composite scale combining all (or most) of the items comprising Att1, Att2, and Att3. I'll call that big composite scale AllAtts. Because you got a significant effect of AllHi on LFP, I think you might well get a significant effect of AllAtts on LFP -- assuming all of the other possible reasons for non-significance that I mentioned above can be dealt with. Good luck with your work.

Dear Prof. Burke D. Grandjean,

I am so pleased to get such a detailed answer. I will give details of what exactly I did in the model.

Thank you so much