Hui guys
I have encountered the following challenge: I have two continuous IV's (team commitment TC and organizational commitment OC) and two continuous DV's (i.e. team-directed helping behavior and organizational-directed extra-role behavior), and several covariates (age, gender, tenure).
The aim of my research is to show that there are cross-level additive and interactive effects (i.e. 1 - TC predicts organization-directed extra-role behavior over and above OC; 2 - OC predicts team-directed helping behaviors over and above TC; 3 - the two commitments interact in their positive effects on team- and organization-directed outcomes).
I used two approached to see whether there are additive and interactive effects of the two IV's on the
1. a variable-centered approach (multiple regression)
2. a person-centered apporach (ANCOVA) where I formed different commitemnt patterns based on a median split of the IV's (i.e. low TC - low OC; low TC - high OC; high TC - low OC; and high TC - high OC) and examined the effect of these patterns on the DV (differences in group means).
If I run a multiple regression (variable-centered apporach), I get a significant interaction effect TC*OC on the DV. However, if I compare means based on the ANCOVA, I do not get significant differences in means between the high-high group (dual commitment) and the high-low group (unilateral commitment to the organization but not to the team).
I try to make sense of that. At first glance, the fact that there is a significant additive and interaction term in the regression analysis should also lead to significant mean differences in ANCOVA. However, that was not the case. Can anyone help?
Dear Jörg,
one question: did you also enter the median splitted variables in the regression model? If not, it is not surprising that you get different results. It is not entirely predictable, how the results will change after a median split, but in general you could expect a massive loss of power, i.e. you cant find differences, although they are there. This is due to the loss of information. The information of your continous variable is reuduced to 1/0. Differences between values within each group are ingnored; similar values in the original variable, may now be treated as differently, if they fall in different categories (values around the median for example).
So, you should generally avoid dichotomization (Irwin article, Altman article), there may be only a few reasons to do it (Farrington article).
But please have a look at the articles attached.
I agree with Ranier that simply dichotomizing your continuous variable results in a substantial "loss of information." Essentially, it assume a single point along the continuum where the effect occurs, whereas multiplying the full range of the two independent variables makes no assumption about what is "high" or "low".
- Badges
- Science topic
- Similar topics
- Geoscience
- Asia
- India