Extension of McNemar's test
Locked
 
|
Dear SPSS List Members,
We are looking for an extension of McNemar's test in order to detect differences in pre-post changes between two groups. Here's what the data look like: (1) Group: Treatment & Control The control group members were eligible for treatment but put on a waiting list. The members were not matched individually to any treatment group cases. (2) We measured employment status (Yes/No) at two points in time. We have assessed change in employment status (time 1 to time 2) *within* each group separately via McNemar's. We'd now like to see if the pattern or rate of change is different between the two groups. I've read that the Breslow-Day test is appropriate for testing for differences between two odds ratios, however, I'm not certain that forming odds ratios on a pre-post group even makes sense. Any suggestions are truly welcome. Hunter ===================== To manage your subscription to SPSSX-L, send a message to [hidden email] (not to SPSSX-L), with no body text except the command. To leave the list, send the command SIGNOFF SPSSX-L For a list of commands to manage subscriptions, send the command INFO REFCARD |
Administrator
|
This reminds me of something I helped a colleague with a couple years ago. They had a binary outcome measured at 3 time points. The conventional approach in the field was to do a separate logistic regression model at each time point. But with GENLIN, we were able to include all 3 time points in one model. If I left out the interaction with Time, I was able to duplicate exactly the results of the individual logistic regressions. But the advantage was we could then throw in a Treatment x Time interaction to ask if the effect of Treatment varied over time.
Here's a modification of the code I used for that, with 2 time points instead of 3. I think the Treat*T2 term in this model addresses your question. * Data file has 2 rows per subject (times 1 and 2). * T2 = a 0/1indicator for time point 2. * Treat = the treatment variable (with two values). * EmpStat = employment status (2 values). * Generalized Estimating Equations. GENLIN EmpStat (REFERENCE=FIRST) WITH treat t2 /MODEL treat t2 treat*t2 INTERCEPT=YES DISTRIBUTION=BINOMIAL LINK=LOGIT /CRITERIA METHOD=FISHER(1) SCALE=1 MAXITERATIONS=100 MAXSTEPHALVING=5 PCONVERGE=1E-006(ABSOLUTE) SINGULAR=1E-012 ANALYSISTYPE=3(WALD) CILEVEL=95 LIKELIHOOD=FULL /REPEATED SUBJECT=mrnum WITHINSUBJECT=t2 SORT=YES CORRTYPE=unstructured ADJUSTCORR=YES COVB=ROBUST MAXITERATIONS=100 PCONVERGE=1e-006(ABSOLUTE) UPDATECORR=1 /MISSING CLASSMISSING=EXCLUDE /PRINT CPS DESCRIPTIVES MODELINFO FIT SUMMARY SOLUTION (EXPONENTIATED). HTH.
--
Bruce Weaver bweaver@lakeheadu.ca http://sites.google.com/a/lakeheadu.ca/bweaver/ "When all else fails, RTFM." PLEASE NOTE THE FOLLOWING: 1. My Hotmail account is not monitored regularly. To send me an e-mail, please use the address shown above. 2. The SPSSX Discussion forum on Nabble is no longer linked to the SPSSX-L listserv administered by UGA (https://listserv.uga.edu/). |
|
|
In reply to this post by Anton-24
To be clear, Bruce is proposing that you test the difference between
groups in log(odds) change from time 1 to time 2 by fitting a generalized linear model. On a side note, if you wanted the GENLIN procedure to output the log(odds) change from time 1 to time 2 for each group in the "Parameter Estimates" table, you could parameterize the model with the fixed effects, group (categorical variable) and group-BY-time, excluding the grand intercept. Ryan On Thu, Mar 31, 2011 at 1:08 PM, Hunter <[hidden email]> wrote: > Dear SPSS List Members, > > We are looking for an extension of McNemar's test in order to detect > differences in pre-post changes between two groups. > > Here's what the data look like: > > (1) Group: Treatment & Control > The control group members were eligible for treatment but put on a waiting > list. The members were not matched individually to any treatment group > cases. > > (2) We measured employment status (Yes/No) at two points in time. > > We have assessed change in employment status (time 1 to time 2) *within* > each group separately via McNemar's. We'd now like to see if the pattern or > rate of change is different between the two groups. > > I've read that the Breslow-Day test is appropriate for testing for > differences between two odds ratios, however, I'm not certain that forming > odds ratios on a pre-post group even makes sense. > > Any suggestions are truly welcome. > > Hunter > > ===================== > To manage your subscription to SPSSX-L, send a message to > [hidden email] (not to SPSSX-L), with no body text except the > command. To leave the list, send the command > SIGNOFF SPSSX-L > For a list of commands to manage subscriptions, send the command > INFO REFCARD > ===================== To manage your subscription to SPSSX-L, send a message to [hidden email] (not to SPSSX-L), with no body text except the command. To leave the list, send the command SIGNOFF SPSSX-L For a list of commands to manage subscriptions, send the command INFO REFCARD |
|
|
In reply to this post by Anton-24
Many thanks to Bruce & Ryan for their help. It looks like the GEE procedure
is going to work well. I created* the following syntax that produces the same results as Bruce's syntax below: genlin EmpStat (reference = first) with Time Treat /model Time Treat Treat*Time distribution = binomial link = logit /repeated subject = mrnum corrtype = unstructured /print solution (exponentiated) modelinfo. My question is this: Will the above approach properly assess the Time by Treatment interaction when there are more than 2 time periods involved? I believe the above is treating the Time periods as if they are continuous, rather than categorical. Again, I'm grateful for any and all advice. Hunter * The above syntax was modified from an example on UCLA's fantastic website: http://www.ats.ucla.edu/stat/spss/library/gee.htm On Thu, 31 Mar 2011 11:45:16 -0700, Bruce Weaver <[hidden email]> wrote: >This reminds me of something I helped a colleague with a couple years ago. >They had a binary outcome measured at 3 time points. The conventional >approach in the field was to do a separate logistic regression model at each >time point. But with GENLIN, we were able to include all 3 time points in >one model. If I left out the interaction with Time, I was able to duplicate >exactly the results of the individual logistic regressions. But the >advantage was we could then throw in a Treatment x Time interaction to ask >if the effect of Treatment varied over time. > >Here's a modification of the code I used for that, with 2 time points >instead of 3. I think the Treat*T2 term in this model addresses your >question. > > >* Data file has 2 rows per subject (times 1 and 2). >* T2 = a 0/1indicator for time point 2. >* Treat = the treatment variable (with two values). >* EmpStat = employment status (2 values). > >* Generalized Estimating Equations. >GENLIN EmpStat (REFERENCE=FIRST) WITH treat t2 > /MODEL treat t2 treat*t2 INTERCEPT=YES > DISTRIBUTION=BINOMIAL LINK=LOGIT > /CRITERIA METHOD=FISHER(1) SCALE=1 MAXITERATIONS=100 MAXSTEPHALVING=5 >PCONVERGE=1E-006(ABSOLUTE) > SINGULAR=1E-012 ANALYSISTYPE=3(WALD) CILEVEL=95 LIKELIHOOD=FULL > /REPEATED SUBJECT=mrnum WITHINSUBJECT=t2 SORT=YES CORRTYPE=unstructured >ADJUSTCORR=YES > COVB=ROBUST MAXITERATIONS=100 PCONVERGE=1e-006(ABSOLUTE) UPDATECORR=1 > /MISSING CLASSMISSING=EXCLUDE > /PRINT CPS DESCRIPTIVES MODELINFO FIT SUMMARY SOLUTION (EXPONENTIATED). > > >HTH. > > > >Anton-24 wrote: >> >> Dear SPSS List Members, >> >> We are looking for an extension of McNemar's test in order to detect >> differences in pre-post changes between two groups. >> >> Here's what the data look like: >> >> (1) Group: Treatment & Control >> The control group members were eligible for treatment but put on a waiting >> list. The members were not matched individually to any treatment group >> cases. >> >> (2) We measured employment status (Yes/No) at two points in time. >> >> We have assessed change in employment status (time 1 to time 2) *within* >> each group separately via McNemar's. We'd now like to see if the pattern >> or >> rate of change is different between the two groups. >> >> I've read that the Breslow-Day test is appropriate for testing for >> differences between two odds ratios, however, I'm not certain that forming >> odds ratios on a pre-post group even makes sense. >> >> Any suggestions are truly welcome. >> >> Hunter >> >> ===================== >> To manage your subscription to SPSSX-L, send a message to >> [hidden email] (not to SPSSX-L), with no body text except the >> command. To leave the list, send the command >> SIGNOFF SPSSX-L >> For a list of commands to manage subscriptions, send the command >> INFO REFCARD >> > > >----- >-- >Bruce Weaver >[hidden email] >http://sites.google.com/a/lakeheadu.ca/bweaver/ > >"When all else fails, RTFM." > >NOTE: My Hotmail account is not monitored regularly. >To send me an e-mail, please use the address shown above. > >-- >View this message in context: >Sent from the SPSSX Discussion mailing list archive at Nabble.com. > >===================== >To manage your subscription to SPSSX-L, send a message to >[hidden email] (not to SPSSX-L), with no body text except the >command. To leave the list, send the command >SIGNOFF SPSSX-L >For a list of commands to manage subscriptions, send the command >INFO REFCARD ===================== To manage your subscription to SPSSX-L, send a message to [hidden email] (not to SPSSX-L), with no body text except the command. To leave the list, send the command SIGNOFF SPSSX-L For a list of commands to manage subscriptions, send the command INFO REFCARD |
|
Administrator
|
Hi Hunter. In SPSS lingo, it is BY factor and WITH covariate, where factor = a categorical variable, and covariate = a continuous variable. So you are right, your syntax treats Time as continuous. Try this instead:
genlin EmpStat (reference = first) by Time Treat /model Time Treat Treat*Time distribution = binomial link = logit /repeated subject = mrnum corrtype = unstructured /print solution (exponentiated) modelinfo. Assuming Treat has only two levels, you should get the same results with: genlin EmpStat (reference = first) by Time with Treat /model Time Treat Treat*Time distribution = binomial link = logit /repeated subject = mrnum corrtype = unstructured /print solution (exponentiated) modelinfo. Although the signs on the coefficients may be switched--I think the default is to use the last category of a factor as the reference category. HTH.
--
Bruce Weaver bweaver@lakeheadu.ca http://sites.google.com/a/lakeheadu.ca/bweaver/ "When all else fails, RTFM." PLEASE NOTE THE FOLLOWING: 1. My Hotmail account is not monitored regularly. To send me an e-mail, please use the address shown above. 2. The SPSSX Discussion forum on Nabble is no longer linked to the SPSSX-L listserv administered by UGA (https://listserv.uga.edu/). |
«
Return to SPSSX Discussion
|
1 view|%1 views
| Free forum by Nabble | Edit this page |