Hypothetical:
I conduct a survey and ask the respondents to answer a yes/no question "do you consider yourself to be an alcoholic". Their responses correlates with high/low averages on another measure "Qualities of Alcoholism" (QoA, e.g. yes=high, no=low). Most of the responses are in the 'no' camp. My questions are:
- Should I include everyone in the sample (yes and no's) if I'm comparing QoA to another measure (e.g. 'Effects of Drinking' scale) or should I just section out the 'yes' cases?
- What would the effects be on regression analyses if I took either approach?
Kellie Vella Focusing on the 'yes' group allows for a more targeted analysis of those who identify as alcoholics, potentially yielding clearer and more relevant insights into the relationship between QoA and the "Effects of Drinking" scale. But analyzing the full sample could reveal subtle relationships that might be missed if you only focus on one group.
So I would suggest you to include both.
Kellie Vella Before making a decision, you should consider your sample size (and the number of participants in each subgroup) and run regression diagnostics separately with each subgroup. You mention majority of participants are in the "no" categories, which means the few in the "yes" categories may not be sufficient to obtain a reliable estimate or satisfy the assumptions of the regression.
If you run a regression with largely unbalanced groups (i.e. too few individuals in some categories), the coefficients may underestimate (or overestimate) the difference that truly exists between the respective populations simply because there are not enough people in some subgroups. In short, the results will not be particularly accurate.
- **Include Everyone**: Provides a broader, generalizable analysis but may weaken the relationship in regression due to the inclusion of the 'no' responses.
- **Section Out 'Yes' Cases**: Focuses on a specific subgroup, likely strengthening the relationship in regression but limiting generalizability.
To add to other answers stated, the specifics of the survey design may also play a role, i.e. the research questions. From your description, it seems you have a low frequency of yeses on QoA. If my research question is about the dependence of QoA on other measures, as suggested in other answers, I too would include the full set of responses in a logistic regression model ( yes and no responses) with other measures as explanatory variables. Here, I assume that the QoA is my response variable. If I ignore the low frequency answers, I cannot use the above dichotomous variable in a regression setting as a response variable and the analysis has to focus on other variables.
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Graphs