Re: Main and interaction effect in Cox Prop Hazards
Locked
 
|
Hello,
===================== 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 |
|||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Re: Main and interaction effect in Cox Prop Hazards
Administrator
|
In Nabble, at least, your table was unreadable. But by coping it, pasting
into Word, and saving as a PDF, I was able to make a readable version. There should be a link to it below. (Members who do not use Nabble may need to view this thread in Nabble to get access to it: http://spssx-discussion.1045642.n5.nabble.com/Re-Main-and-interaction-effect-in-Cox-Prop-Hazards-td5738390.html.) Variables_in_the_Equation.pdf <http://spssx-discussion.1045642.n5.nabble.com/file/t7186/Variables_in_the_Equation.pdf> In a regression model, the terms that you are calling main effects are really simple effects. E.g., the coefficient for heavy_drinker shows the effect of heavy drinking when educ_lowses is set to its reference category. To get the effect of heavy drinking in the other category of educ_lowses, add the coefficient for the interaction. Similarly, the coefficient for low_ses shows the effect of low_ses when heavy_drinker is set to its reference category. To get the low_ses effect for the other category of heavy_drinker, add the coefficient for the interaction. I worked out those effects in Excel. Not sure how good the formatting will be, but I'll paste the results here: B Heavy Drinking Low SES Yes No Yes 0.299 0.393 No 0.566 0.472 If you exponentiate those numbers, you'll get the corresponding Exp(B) values: Exp(B) Heavy Drinking Low SES Yes No Yes 1.349 1.481 No 1.761 1.603 Finally, the test on the interaction is nowhere near significance (p = .418). So unless there is a good theory-based reason for keeping it in the model, I would consider removing it. HTH. jaycamh wrote > Hello, >> I need help. >> >> I am doing a Cox proportional hazards model in SPSS. I have Low SES and >> Heavy drinking as 2 main effects and LOW SES x Heavy drinking as >> interaction effect in the model. >> >> We have stronger (and significant) effects for each main effect but the >> interaction term is coming protective which is against the theory. Not >> sure if I am doing it wrong ... see the SPSS result below. >> >> I have, Low SES 0 is reference; 1=low SES >> >> Heavy drinking 0 is reference; 1=heavy drinking. >> >> >> in fact when I make interaction of alcohol and smoking it’s also coming >> protective... >> Is my Interpretation wrong or is there any other way my syntax should >> have been formed. >> >> > > Variables in the Equation > > B > SE > Wald > df > Sig. > Exp(B) > 95.0% CI for Exp(B) > > Lower > Upper > AGE > .085 > .000 > 3.719E4 > 1 > .000 > 1.089 > 1.088 > 1.090 > SEX - FEMALE > -.433 > .013 > 1.179E3 > 1 > .000 > .649 > .633 > .665 > BMI > .001 > .001 > 6.844 > 1 > .009 > 1.001 > 1.000 > 1.003 > YEAR > -.006 > .002 > 15.634 > 1 > .000 > .994 > .990 > .997 > smoker_cfn > > > 2.716E3 > 2 > .000 > > > > smoker_cfn(1) > .827 > .016 > 2.600E3 > 1 > .000 > 2.285 > 2.214 > 2.359 > smoker_cfn(2) > .223 > .015 > 221.735 > 1 > .000 > 1.250 > 1.214 > 1.287 > heavy_drinker > .566 > .110 > 26.427 > 1 > .000 > 1.761 > 1.419 > 2.186 > educ_lowses > .393 > .016 > 608.737 > 1 > .000 > 1.481 > 1.436 > 1.528 > heavy_drinker*educ_lowses > -.094 > .116 > .656 > 1 > .418 > .911 > .726 > 1. > >> >> >> >> > > ===================== > To manage your subscription to SPSSX-L, send a message to > LISTSERV@.UGA > (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. -- Sent from: http://spssx-discussion.1045642.n5.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
--
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/). |
Re: Main and interaction effect in Cox Prop Hazards
|
|
In reply to this post by jaycamh
(I'm reading the distribution from SPSSX-L directly, ,using Outlook, and the table is formatted fine for me.)
No - you do not have an effect worth mentioning for the interaction that you have coded.
In fact, it is somewhat impressive that the Wald chi-squared for it is less than 1; it is not
even picking up trivial, artifactual contributions from anywhere.
Do you know how to read those "Wald" numbers that end in E? For Age, Wald is 3.719E4.
"3.719E4" is a notation where "E4" stands for "10 to the 4th": move the decimal over
by 4 places, so the actual number is approximately 37,190. Apparently your sample size
is "hundreds of thousands" for it to generate a chi-squared type of test that is that large.
Just like with any ANOVA table, your emphasis should be on the larger tests values when
their sizes vary by orders of magnitude. Why? For example: If your "age" is at all confounded with
BMI (which it probably is), then it is easy to imagine that the /observed/ BMI effect, Wald= 6.844,
will disappear if there were a more complete control for Age. It would not be unreasonable,
IMHO, to "look at" a re-run of the test where Age is entered as dozens of separate years, as dummy
categories, so that there are K-1 d.f. for age variables (including "linear") where K is the number of
separate ages. - Looking at the effects on other coefficients will give you /illustrations/ of how
the largest effects can create artifacts.
I'm not sure what Bruce was saying about coefficients.
--
Rich Ulrich
From: SPSSX(r) Discussion <[hidden email]> on behalf of i Jay <[hidden email]>
Sent: Friday, September 20, 2019 12:53 PM To: [hidden email] <[hidden email]> Subject: Re: Main and interaction effect in Cox Prop Hazards Hello,
=====================
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 |
|||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Re: Main and interaction effect in Cox Prop Hazards
|
Administrator
|
Rich Ulrich wrote
> I'm not sure what Bruce was saying about coefficients. I'll try again. Here are the B values for the two variables of interest and their interaction: B(drink) = 0.566 = ln(HR) for drink when ses = 0 (i.e., reference category) B(ses) = 0.393 = ln(HR) for ses when drink = 0 (i.e., reference category) B(interaction) = -0.094 Therefore: 0.566 -0.094 = 0.472 = ln(HR) for drink when ses = 1 0.393 -0.094 = 0.299 = ln(HR) for ses when drink = 1 The interaction term tests the differences between 0.566 and 0.472 (for drink) and between 0.393 and 0.299 (for ses). And of course, those ln(HR) values are adjusted for the other variables in the model. Is this clearer? ----- -- 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. -- Sent from: http://spssx-discussion.1045642.n5.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
--
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/). |
|
|
Thank you Bruce.
Sent from my iPhone > On Sep 20, 2019, at 6:02 PM, Bruce Weaver <[hidden email]> wrote: > > Rich Ulrich wrote >> I'm not sure what Bruce was saying about coefficients. > > I'll try again. Here are the B values for the two variables of interest and > their interaction: > > B(drink) = 0.566 = ln(HR) for drink when ses = 0 (i.e., reference category) > B(ses) = 0.393 = ln(HR) for ses when drink = 0 (i.e., reference category) > > B(interaction) = -0.094 > > Therefore: > 0.566 -0.094 = 0.472 = ln(HR) for drink when ses = 1 > 0.393 -0.094 = 0.299 = ln(HR) for ses when drink = 1 > > The interaction term tests the differences between 0.566 and 0.472 (for > drink) and between 0.393 and 0.299 (for ses). > > And of course, those ln(HR) values are adjusted for the other variables in > the model. > > Is this clearer? > > > > > ----- > -- > 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. > > -- > Sent from: http://spssx-discussion.1045642.n5.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 |
Re: Main and interaction effect in Cox Prop Hazards
|
|
In reply to this post by Bruce Weaver
Bruce,
I hate it when interactions have to be interpreted, especially if it
is possible that the program generated them and I don't have control.
Also, telling a program that this is a "factor" and not a continuous
"variable" raises the problem that, for 0/1, if the program takes "1" as
the reference group, it reverses the sign of the B obtained. PITA.
Further: if two variables are coded as 0/1 and the interaction is their
simple product - instead of the "centered" product - then the interaction's
B is correlated with the main-effects' B's, and affects their magnitude in
the equation that has them all.
I recognize that this OP needs basic help with interpretation of the model,
which you provided. I was skipping that, because the main thing relevant
to the question is that there /is/ no interaction effect worth worrying
about, according the Wald test.
--
Rich Ulrich
From: SPSSX(r) Discussion <[hidden email]> on behalf of Bruce Weaver <[hidden email]>
Sent: Friday, September 20, 2019 6:02 PM To: [hidden email] <[hidden email]> Subject: Re: Main and interaction effect in Cox Prop Hazards Rich Ulrich wrote
> I'm not sure what Bruce was saying about coefficients. I'll try again. Here are the B values for the two variables of interest and their interaction: B(drink) = 0.566 = ln(HR) for drink when ses = 0 (i.e., reference category) B(ses) = 0.393 = ln(HR) for ses when drink = 0 (i.e., reference category) B(interaction) = -0.094 Therefore: 0.566 -0.094 = 0.472 = ln(HR) for drink when ses = 1 0.393 -0.094 = 0.299 = ln(HR) for ses when drink = 1 The interaction term tests the differences between 0.566 and 0.472 (for drink) and between 0.393 and 0.299 (for ses). And of course, those ln(HR) values are adjusted for the other variables in the model. Is this clearer? ----- -- 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. -- Sent from: http://spssx-discussion.1045642.n5.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 |
Re: Main and interaction effect in Cox Prop Hazards
|
Administrator
|
Fair points, Rich. It would be helpful if the OP posted the syntax.
Also, if I'm using a command that allows factor variables, if it also has EMMEANS, I like to generate the fitted values (and pairwise contrasts) corresponding to the interaction term to make sure I've got things right. Rich Ulrich wrote > Bruce, > I hate it when interactions have to be interpreted, especially if it > is possible that the program generated them and I don't have control. > Also, telling a program that this is a "factor" and not a continuous > "variable" raises the problem that, for 0/1, if the program takes "1" as > the reference group, it reverses the sign of the B obtained. PITA. > > Further: if two variables are coded as 0/1 and the interaction is their > simple product - instead of the "centered" product - then the > interaction's > B is correlated with the main-effects' B's, and affects their magnitude in > the equation that has them all. > > I recognize that this OP needs basic help with interpretation of the > model, > which you provided. I was skipping that, because the main thing relevant > to the question is that there /is/ no interaction effect worth worrying > about, according the Wald test. > > -- > Rich Ulrich ----- -- 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. -- Sent from: http://spssx-discussion.1045642.n5.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
--
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/). |
|
|
Hi Bruce: Here is the syntax. COXREG event /STATUS=dead(1) /CONTRAST (sex)=Indicator(1) /CONTRAST (smoker_cfn)=Indicator /CONTRAST (HEAVY_DRINKER)=Indicator(1) /CONTRAST (educ_Lowses)=Indicator(1) /CONTRAST (HED)=Indicator(1) /METHOD=ENTER age sex bmi year smoker_cfn educ_lowses HEAVY_DRINKER HED HEAVY_DRINKER*educ_lowses /PRINT=CI(95) /CRITERIA=PIN(.05) POUT(.10) ITERATE(20). Here is how the reference groups are set.
OUTPUT:
On Sun, Sep 22, 2019 at 8:58 AM Bruce Weaver <[hidden email]> wrote: Fair points, Rich. It would be helpful if the OP posted the syntax. |
|||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Re: Main and interaction effect in Cox Prop Hazards
|
|
In reply to this post by jaycamh
By the way - I would like to add - do not present the Age effect, when you get that far,
as the B coefficient = 1.089. That is the increase in the OR for risk for a /single/ year. As a
"risk", it is not nearly commensurate with the values of the risks you produce for your
dichotomous items.
There are a couple of approaches for handling this. One that does not depend on the
range of ages in the sample is run the run with an alternate version of age, like, AGE10 = AGE/10 .
The test-statistics remain exactly the same, but the new "effect" is the risk per decade of age,
and will be 2.346 (if I've picked the right formula - raise 1.089 to the power of the number
of years, which is 10).
The approach which gives values that are even more commensurate to the dichotomies is
to compare the risk at the range of plus-and-minus one standard deviation of the age range.
How many years? - raise 1.089 to that power. That one would do more to emphasize that
Age is by far the strongest effect according to the Wald tests.
--
Rich Ulrich
|
Re: Main and interaction effect in Cox Prop Hazards
|
|
In reply to this post by Bruce Weaver
Bruce "RTFM" The Manual says that the best way forward is to use the Google Group "MedStats". (I'm biased....I'm the Founder :)) Kind Regards Martin Freelance Medical Statistician If you can't explain it simply, you don't understand it well enough.....Einstein Concise Encyclopedia of Biostatistics for Medical Professionals Martin P. Holt Linked In: https://www.linkedin.com/in/martin-holt-3b800b48?trk=nav_responsive_tab_profile
On Saturday, 21 September 2019, 00:15:41 BST, Bruce Weaver <[hidden email]> wrote:
Rich Ulrich wrote > I'm not sure what Bruce was saying about coefficients. I'll try again. Here are the B values for the two variables of interest and their interaction: B(drink) = 0.566 = ln(HR) for drink when ses = 0 (i.e., reference category) B(ses) = 0.393 = ln(HR) for ses when drink = 0 (i.e., reference category) B(interaction) = -0.094 Therefore: 0.566 -0.094 = 0.472 = ln(HR) for drink when ses = 1 0.393 -0.094 = 0.299 = ln(HR) for ses when drink = 1 The interaction term tests the differences between 0.566 and 0.472 (for drink) and between 0.393 and 0.299 (for ses). And of course, those ln(HR) values are adjusted for the other variables in the model. Is this clearer? ----- -- 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. -- Sent from: http://spssx-discussion.1045642.n5.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 |
«
Return to SPSSX Discussion
|
1 view|%1 views
| Free forum by Nabble | Edit this page |