1

I am currently conducting a study about the involvement of male in family planning. It is of interest to create an index for the involvement of male in FP. My variables were only categorical variables and mainly dichotomous variables. Is there anyone who is knowledgeable on creating index using logistic approach? or any suggestions for the creation of this index. Thanks so much

paualdemita
  • 55
  • 2
  • 4

1 Answers1

1

The general sequence for this kind of problem is:

  1. check your items actually scale together
  2. fit an IRT model to them
  3. confirm your model fits reasonably
  4. extract factor scores for cases

If you happen to be an R user then one can do 1. using the aisp function in the mokken package. This looks at your items and tries to tell you how many (if any) separate scales there are among your items. You can also compute Cronback alphas and other stats with this package.

If for 2. you want to fit an item response model -- I'm guessing this is what you mean by 'logistic approach' -- then you can use the grm function from the ltm package to fit a scale that has some mix of dichotomous or Likert items, and the factor.scores function to get scale scores for whatever subset of items you chose.

More practical detail and pointers to the relevant theory is provided by the package vignette. As it happens the example code all uses the built in data set WIRS which consists of 6 dichotomous items, so you can see how it works using that.

conjugateprior
  • 19,431
  • 1
  • 55
  • 83
  • Thank you very much for your help. Is this approach available in Stata or SPSS? In 2, how can I fit a model if I don't have a dependent variable(s)? Is there any reference/source for examples of such creation of this kind of index? – paualdemita Mar 17 '14 at 04:32
  • @paualdemita The answer to this [related question](http://stats.stackexchange.com/questions/15565/how-to-get-started-with-applying-item-response-theory-and-what-software-to-use) may be helpful. In Stata it seems that one uses GLAMM [here](http://www.gllamm.org/faqs/models/irtfitb.html) for binary, and [here](http://www.gllamm.org/faqs/models/irtfito.html) for ordinal items. – conjugateprior Mar 17 '14 at 08:45
  • 1
    About the model: the point of measurement modelling is to build a model of how the unobserved thing to be measured affects the observed items, usually with the assumption it affects them each separately (local independence). In your case the model will be some set of logistic regressions where the 'independent' variable is the thing to be measured and the dependent variable is an item. It might be worth working through the first link in my first comment link to get the idea. – conjugateprior Mar 17 '14 at 08:52
  • To be detailed: My variables in the study are: 1. Male's fertility preference 2. Desire more children 3. Male currently use contraceptive method 4. Male heard about CM and/or FP 5. Male knows side effects of CM 6. No. of ideal children 7. total no. of living children 8. knows where to obtain CM 9. Male knows wife's menstruation cycle 10. who decides using CM 11.heard STDs and somesocio-demographic variables as well as female variables(same as male variables listed above) With these variables, I want to create an index measuring the involvement of male in Family Planning. – paualdemita Mar 17 '14 at 11:23
  • With these variables, i can't fully grasp the idea of using logistic models to create the index since a dependent variable is needed to fit a regression model or IRT model.. Please help. – paualdemita Mar 17 '14 at 11:27
  • I assume that you want an index because you think that the responses to some of these item variables are caused by a single unobserved something which you're thinking of as 'involvement'. Moreover you can also imagine being able to order that subset of items from left to right such that if someone says yes to one item then they tend to also say yes to the item to the left of it (a stochastic dominance or 'Guttman' structure). e.g if they know the wife's cycle then they nearly always have heard about CM/FP but not vice versa. – conjugateprior Mar 17 '14 at 16:26
  • My step 1 is a automated check that all (rather than just some of) your subset of items have this dominance structure, at least probabilistically. Clearly if they don't you've got items that measure two or more things at once and any index or scale will be to some extent invalid. My step 2 reflects the idea that you wanted to fit an explicit measurement model once you had established the uni-dimensionality of the items. But from your questions I think maybe that's not what you meant by 'logistic approach'. (fyi IRT is a *latent* variable model so there is no 'dependent variable') – conjugateprior Mar 17 '14 at 16:35
  • accepted. :)sorry for the delay, im just new here so i don't know that I have to accept it. :) – paualdemita Apr 03 '14 at 18:06
  • @paualdemita not a problem - and you don't *have* to accept an answer, just if it helps move you along. You can always ask a bunch more questions on the same topic later as you get more into it. – conjugateprior Apr 05 '14 at 19:24