Hi! I am working on an epidemiological database in which data is expressed in terms of frequency and not as individual cases. The frequencies are too much (hundreds of thousands) to reconvert in subject-cases.
How can I performe a binary logistic regression in SPSS with such data?
Hello Ainitze,
You could use the frequencies as a weighting variable and run an ordinary LR in SPSS (or your software of choice). In SPSS, from the main menu bar, select Data / Weight Cases, then click the radio button in the subsequent dialog box for "Weight cases by" and copy the frequency variable in. Internally, the program behaves as if you did convert the grouped profiles into individual subject-cases even though your data worksheet only shows a limited number of "cases" (each with its own weighting/frequency value, of course).
Good luck with your work.
Full Disclosure, I am not thankfully an spss user. However, I do not think it is binary logistic and weights Is the way to go. Binary logistic is what it says on the can, binary with individual cases. And the weights are survey weights, that is certain individuals have been deliberately oversampled eg young black men, complex survey data will provide this with the data.
You want instead the genlin procedure which deals with frequencies and varying denominators , that is binomial regression not Bernoulli. I hope I am wrong as the genlin procedure is quite complicated, as it does GEE as well.
https://www.ibm.com/support/pages/using-spss-genlin-generalized-linear-models-procedure-fit-binary-logistic-regression-model
https://www.ibm.com/support/knowledgecenter/SSLVMB_24.0.0/spss/advanced/syn_genlin_overview.html
If if I am correct, it would make a lot of sense for a lot of confusion amongst posters on rgate I have been arguing with.
https://www.researchgate.net/post/what_is_the_difference_between_linear_regression_and_logistics_regression
As they see everything in binary terms.
For the first time that I can recall, I disagree with Kelvyn Jones. :-O ;-)
Kelvyn, the GENLIN functionality you are referring to, I think, is the Events-of-Trials specification. It is not intended for dealing with survey weights.* Rather, I would use that approach when each independent subject (or unit of observation) has X events (or successes) in T trials--i.e., a binomial proportion as the outcome. But that is not what I see in the document Ainitze Labaka uploaded. Rather, I see frequency weights that indicate how many independent subjects have each combination of variables (including the outcome variable). And so, assuming the outcome is dichotomous, I would use the WEIGHT CASES approach described by David Morse.
Cheers,
Bruce
* The CSLOGISTIC command would be used for dealing with survey weights. (I have no experience with it.)
https://www.ibm.com/support/knowledgecenter/en/SSLVMB_26.0.0/statistics_reference_project_ddita/spss/complex_samples/syn_cslogistic_overview.html
Apologies for the confusion.
To be clear I was not suggesting that GENLIN should be used for taking account of survey weights. Rather I thought that this procedure was for dealing a binomial proportion; exactly as Bruce Weaver says in which there is a numerator (the number of cases) and a denominator (number of case plus number of non cases). And I thought that this was the data that Ainitze Labaka had if we were given a longer extract. Or more correctly could be readily be turned into.
I have therefore taken the liberty of extending Ainitze's data to include the Non-Cancer cases with fictitious values and showing how to turn it into the "R of N" form required by GENLIN. (see document attached). You are then modelling the proportion of total cases which are cancers, Binomial not Bernoulli.
I would be interested in knowing if this gives the same results as WEIGHT CASES approach as recommended by
David Morse
My confusion was not helped by this explanation of Weighting cases https://libguides.library.kent.edu/SPSS/WeightCases#:~:text=Weighting%20cases%20in%20SPSS%20works,Click%20OK.
" In SPSS, weighting cases allows you to assign "importance" or "weight" to the cases in your dataset. Some situations where this can be useful include:
- Your data is in the form of counts (the number of occurrences) of factors or events. The "weight" is the number of occurrences.
- Your data requires adjustments to correct for over- or under-representation of certain characteristics in your sample. (This often happens with large surveys: a "weighting" variable is developed to adjust a sample's composition to be reflective of the population's composition, or to control for over- or under-reporting from a certain group.)
- The Pew Research Center often makes their raw survey data available online to the public. Many of these surveys include weighting as a part of the study methodology.
Weighting cases in SPSS works the same way for both situations. "
And the syntax LOGISTIC REGRESSION procedure had nothing about weights. So I do not know how the weights feed through to the binary logistic.
It is really good to have these conversations with researchers who are regular users of SPSS. I have found that SPSS is brilliant at many things, but for somethings the documentation and interface are very problematic .
- Badges
- Science topic
- Similar topics
- Medicine
- Epidemiology