Aiming to learn more about the 'soft' determinants of brand post popularity on social networking sites, we conducted a content analysis of Facebook post 'appeals' (e.g. fun, cuteness, adventure, or pride) for a lot of posts. We can show that certain 'post appeals' have a positive influence on likes, comments, and/or shares; and some appeals have a negative influence on these measures.
So far, 24 appeals are included into the regression analyses as dummy variables (0; 1=yes, this appeal is present in this Facebook post). Now, we assume that certain combinations of appeals (e.g. product-performance-related information ['performance' appeal] and an adventurous story ['adventure' appeal]) have a particular positive (or negative) influence on brand post popularity.
We already indentified frequently used 'appeal pairs' and created dummy variables (0; 1=yes, this post contains a combination of [e.g.] 'performance' and 'adventure'). Can we just include that in our regression analyses (dependent variables = likes/comments/shares), in addition to the 24 'single appeals'? Does this make sense or do we perhaps get problems with correlations?
Example:
==> 'performance' appeal => found in 350 posts
==> 'adventure' appeal => found in 250 posts
==> both appeals together => found in 75 posts
In a nutshel, we want to do two things: First, we want to show which appeals have a positve (or negative) influence on post popularity measures (likes, comments, shares). => done. Second, we want to analyze whether there are 'appeal pairs' that are particularly successful in stimulating post popularity (or not). => not sure how...
We are very thankful about your comments and ideas!
The application of correlation coefficients (Pearson; Spearman) may help.
A very interesting study, Timm,
Your use of a new dummy variable to account for joint occurrences of appeals should work but, as you suggest, there is a risk of multicollinearity in your regressors. This can lead to unreliable coefficients. In some cases, multicollinearity leads to apparently significant coefficients when they shouldn't be, or non-significant coefficients when they shouldn't be, or even significant coefficients with the wrong sign. If your result doesn't make sense, then you should suspect something is wrong.
Another approach may be to use ANOVA. All of your are binary variables, so you should get almost identical Partial Eta estimates as the Standardised Betas in your Regression. The advantage of ANOVA is that you can explicitly include different interaction effects in your model.
And, just to be sure that your estimates are robust against possible multicollinearity problems, you could run the models with, say, 1/4 of the observations randomly removed. Do this five or six times to satisfy yourself that you get about the same results.
Hi Timm, there is normally no problem in including interaction terms, on the contrary. Sets of variables linked together via interaction terms will inevitably form a more or less multicollinear set. But if the variables for which you add the interaction term are not the ones you are primarily interested in then multicollinearity is not a problem. The subset may be multicollinear internally, but this will not bias the estimates for your variable of interest.
Even if you use interaction terms for your variables of primary interest, you might still obtain relatively narrow confidence intervals that satisfy your needs. You would of course normally have a good theory why you include the interaction terms. Or your original variables.
By the way, 24 variables of primary interest in a regression seem to me a bit too much. For one, you would expect at least one of them to be a false negative or false positive, if you use 5% confidence levels and powers of around 80%
I agree with Mate. Typically, there is no problem with including interaction terms. Further, if there are significant interactions between specific independent variables, then interpretation of their association with the outcome considered independently is misguided, as the main effects would be less meaningful in the presence of a significant interaction. The interaction terms can be included in the model as the product of the independent variables in the model (i.e., multiply X1 by X2, etc.). But herein lies a potential problem. A full factorial model would include all higher order interaction terms (i.e., if 3 IVs then there are 3 first order interaction terms and 1 second order interaction term). As Mate notes, the 24 IVs you include represents a large set of IVs already, and result in an impossibly large set of interaction terms for a full factorial model. There is a limit to the number of terms you can include since the model cannot include more than n - 1 IVs (where n is the total number of observations). And practically, the problem of collinearity increases with the number of interaction terms included. So, a theory-driven selection of the relevant interaction terms would be very desirable in your situation. HTH
- Badges
- Science topic
- Similar topics
- Social Science
- Sociology