Hi Lucinda,

Thanks very much for your response. You have certainly helped me to think more clearly about the issues surrounding this problem and I'll be re- reading your reply in order to help me fathom out what's going on with this data.

Thanks again,

Lou

On Thu, 15 Jun 2006 13:06:37 -0700, LUCINDA M TEAR <lucindatear@msn.com> wrote:

>Hi, Lou. I agree with all that Keith has said. I might add that the non >significant interaction using categorical variables could be due either to >the fact that by lumping together the Y responses over a range of X inputs >you created a categorical variable whose variance is large enough that it is >not possible to detect any interaction and/or that the endpoints of the bin >categories you are using occur at points in the data that obscure the >interaction you found using the continuous data. > >In some cases, it may actually serve you to have a model without an >interaction effect - it is possible, however, that the confidence intervals >around such a model will be larger than they would be from a model with an >interaction. On the other hand, using the continuous data apparently >allowed you to detect some underlying "process" (the interaction you found). >If you are trying to understand what creates the patterns you see in your >data, both models give you information about the resolution at which certain >processes are revealed or obscured. Apparently lumping the way you have >obscures the interaction. You might want to try binning your x variables >differently than the previous report did, just to see if there is a way to >categorize the x variables such that an interaction is detected. You could >probably use plots from your continuous model to give you an idea about >where appropriate bins thresholds might lie. I tend to be one who likes to >use models as a way of revealing the "scale" at which the data should be >approached in order to answer the question at hand. A different question >about the same data could require a different type of model. Models also >help you find out if the scale you are looking at is missing information >about some underlying effects that could effect the application of the >results. > >Just some thoughts. > >Lucinda > > > >----- Original Message ----- >From: "Statisticsdoc" <statisticsdoc@cox.net> >Newsgroups: bit.listserv.spssx-l >To: <SPSSX-L@LISTSERV.UGA.EDU> >Sent: Thursday, June 15, 2006 12:36 PM >Subject: Re: Logistic regression help > > >> Keith Starborn >> www.statisticsdoc.com >> >> Lou, >> >> I bet most of the people on this listerserv have faced a similar dilemma >> at some time in their careers. Which one is best from the point of view >> of using the data to answer your questions and generate information that >> you can act on? Probably, keeping the variables continuous is better from >> that point of view. >> >> As to the politics of the situation, in your position, I would run the >> analyses both ways (continuous and categorized) in order to: a.) show that >> I did the analysis the way I was told to; and b.) found something else >> that works better. You know the situation best of all. >> >> HTH, >> >> KS >> >> ---- Lou <charl_bean@YAHOO.CO.UK> wrote: >> > Dear Keith, >> > >> > Thanks for your advice which was very helpful. I feel a bit stuck as to >> > know what to do about this really. My boss (who knows rougly zero about >> > statistics) is insisting that I categorise these variables since I am >> > comparing results with a previous report which did the same. Does it >> > take >> > meaning away from the analysis if I discuss results obtained using the >> > original continuous variables and then discuss results separately using >> > the categorised versions (i.e. generate two separate models)? Not sure >> > if >> > this really defies logic too much and how I would justify this in the >> > final report. Although I have a lot to learn in this field, the report >> > that this work is being based on has a lot of dubious findings with >> > regards to the stats, so I'm very keen to ensure that the one I produce >> > is >> > accurate!! >> > >> > Many thanks, >> > >> > Lou >> > >> > On Thu, 15 Jun 2006 11:36:45 -0400, Statisticsdoc >> > <statisticsdoc@cox.net> >> > wrote: >> > >> > >Keith Starborn >> > >www.statisticsdoc.com >> > > >> > >Dear Lou, >> > > >> > >Categorizing continuous variables into categorical variables can result >> > is a considerable loss of statistical power because the test for the >> > categorized version of the variable uses more degrees of freedom that >> > the >> > test for the continuous variable. In addition, categorizing a >> > continuous >> > variable can result in a loss of predictive information. >> > > >> > >HTH, >> > > >> > >KS >> > > >> > >---- Lou <charl_bean@YAHOO.CO.UK> wrote: >> > >> Dear list >> > >> >> > >> I am trying to carry out a logistic regression analysis and have a >> > >> quick >> > >> question with regards to the best way to input my independent >> > >> variables. >> > >> I have three input variables: ethnicity (5 groups), age and >> > >> deprivation >> > >> score. Although age and deprivation score are continuous variables, >> > >> I >> > >> have also been asked to split them into groups (4 for age and 5 for >> > >> deprivation) which are pre-determined by previous work on this >> > >> subject >> > >> matter. The dependent variable is simply whether or not a person >> > >> took a >> > >> particular test. >> > >> >> > >> I have tried generating models both with the age and deprivation >> > variables >> > >> as they are and also with the new categorical age and deprivation >> > >> variables. However, when looking at interaction terms, I find that >> > >> the >> > >> interaction between age and deprivation is significant when they are >> > input >> > >> as the continuous variables but not significant when I used the >> > >> categorical versions. Why would this happen? Furthermore, which is >> > >> the >> > >> best way to go? I have read information on logistic regression until >> > >> my >> > >> head hurts, but still don’t feel completely satisfied as to how I >> > should >> > >> determine the best model possible. >> > >> >> > >> Any advice would be appreciated please! >> > >> >> > >> Thanks >> > >> >> > >> Lou >> > > >> > >-- >> > >For personalized and experienced consulting in statistics and research >> > design, visit www.statisticsdoc.com >> >> -- >> For personalized and experienced consulting in statistics and research >> design, visit www.statisticsdoc.com >>