I have a binary DV (0/1) and I am running a univariate then a multivariate regression, initially in the univariate test, I end up with SPSS flagging cases as outliers (I use the casewise list Zresid criteria) the issue is that when I remove the cases considered outliers (I used my criteria with Zresid over 3 to be an outlier) the cases I'm left with are all dependant variable = 0 (so I have none of my DV=1 left) which prompts an error message. My question is, do I just use the results from the initial regression and just highlight that I have outliers or are there other tests I should use?
Also, do I have to ensure I meet the assumptions for my univariate and be as stringent as if it was a multivariate model?
Thank you
some info:
my main independent is also categorical, 3 levels (and I have both binary and multinomial that will be included in the multivariate model 1 + and a continuous for my multiV model 2 )
my DV is binary
I have had many difficulties with the SPSS approach. I have switched to R with no problem. Suggest you download the attached reference from the attached place and look at logistic regression programs which are completely specified. I have a recently accepted paper to give you you an idea of the power and ease of doing this attached too. Questions gladly accepted. Good luck, David Booth
Oops I forgot to mention that all of the above downloaded materials and software are free. Best wishes, David Booth
Why are you removing outliers? Do you believe these points are not from your population of interest? They are often the most interesting in a data set!
...especially since all the data that deviates from the general mean seems to yield DV=1. You should consider plotting DV against zresid to investigate
Sven Greving I know about plotting a continuous DV against Z residuals in the context of linear models but I don't quite see what it will be telling me in this context of a logistic model. Thank you
Hi Meryem (hey Sven ;)),
I am not sure what outliers or residuals in the logistic case should be anyway. And I also know nobody who could give an answer to this. Your DV has two values 0 and 1, so logistic regression is the way to go. There are no outliers, and also no "residuals" in the standard sense (i.e. in Case 1).
Case 1:
If you have one measurement per individual (e.g, participants in an experiment each having a specific attribute with 'present=1' and 'absent=0'), then the logistic regression estimates effects (condition-wise average estimates) on the log-scale, which then are logit-transformed to probabilties, and these probabilities are the proportion of individuals having this attribute 'present'. (i.e., no outliers possible, not important) depending on some IV (predictor/condition). As mentioned in other replies: What you call outliers here are actually required observations to calculate proportions across individuals. If one individual (e.g. with DV=0) is considered to be an outlier, then all individuals with DV=0 are, and no proportion can be calculated anymore.
Case 2:
If you have multiple measurements per individual (e.g., each participant in an experiment makes 10 choices, each can be yes = 1 and no =0) then each participant can either have 0%, 10% ... 90%, 100% proportion of yes choices. If the average across all individuals is 15%, and some individuals have 100%, those might be considered as outliers, ... if there are theoretical reasons for it (e.g., response bias might be a problem). But this does not sound like the design you have. Also, usually, one would consider zero-inflation models (to account for too much zero-observations), or even beta-regression models (which allow to model bimodal proportion distributions).
Case 3:
If each individual has more than 30 observations (0 or 1), you can count them together (ranging individual scores between 0 and 30+; assuming 'some' independence), and with large numbers binomial distributions approach normality. Then you can do normal regression stuff. :)
My suggestion. Forget about the outliers here :)
Best
René
- Badges
- Science topic
- Similar topics
- Publishing