|
|
شما اينجا هستيد: صفحه اصلي » تبادل نظر » تئوريك » روشهاي آماري در اپيدميولوژي
| ا16/2/1386 ا10:15 صبح |
|
|
دكتر محمد انصاري پور
|
با سلام
پروپوزال با موفقيت ثبت شد. اما يك سوال هم دارم: براي adjustment به روش M-H در يك جدول 2*2 پس از ايجاد strata و محاسبه OR(MH در صورت معني دار بودن آن چگونه در مورد interaction قضاوت مي كنيم؟ |
| ا16/2/1386 ا11:47 صبح |
|
|
irzamin
|
Hello! Let 's say you have an outcome and two other variables. You would like to assess the effect of one of the variables, let's say an exposure for example, on the outcome. You would like to see if the effect of the exposure on the outcome is modified by the effect of the third variable or not. Effect modifiers (in statistical world are called "interactions") are always superior to confounders in epidemiologic studies. So If you see the presence of an interaction, you must NOT calculate the Mantel-Haenszel odds ratio, which is a weighted average odds ratio. Instead you should calculate and report two separate odds ratios for each strata (here you said you have two). But if the interaction (effect modifier) is not significant, then you can go a head and calculate the M-H odds ratio. To my understanding your question is not quite correct, you have to asses interaction first, and if there is not any interaction you can asses confounders and report M-H OR. Good luck |
| ا16/2/1386 ا2:43 بعدازظهر |
|
|
دكتر محمد انصاري پور
|
Hi
thank you for your answer. let me ask my question again giving an example.
if in a case control study crude OR= 2.41 , and in the stratoms OR1=2.30 , OR2=1.95, OR3=2.10 and ORmh=2.06 with CI ( 1.34 , 3.15 ) and P value < .001 what is your opinion? is there any interaction or confounding? was it necessary to calculate ORmh? why? |
| ا16/2/1386 ا3:38 بعدازظهر |
|
|
irzamin
|
The stratum-specific odds ratios (OR1, OR2, and OR3) are quite different from each other suggesting interaction (i.e. the variable is an effect modifier). So when a factor is an effect modifier, we make no further discussion whether it is a confounder or not. We just stop at this point and will report the 3 stratum-specific odds ratios (if this be the case calculation of M-H OR is wrong).
Let's assume the stratum-specific ORs were almost identical, then you could have compared the stratum-specific ORs with the crude (unadjusted) OR to see if they differ by about 10%. Again in the absence of an interaction, if the crude and adjusted estimates for OR be different for more than 10% of each other, you will conclude that factor (variable) is a confounder, and then you can report M-H OR. Unlike interactions, We should never decide whether a factor is a confounder or not based on formal statistical testing (because interaction is a statistical concept while confounding is a biologic concept), we should only do simple comparison between crude and adjusted ORs (after being sure that there is no interaction) to see if they differ by more than an arbitrary value usually 10%.
The last thing I would like to add is that because the stratum-specific ORs are close to each other in your example, (for being more confident) you can run a logistic regression, and include the interaction term between your exposure and the questionable factor, to see if the interaction term in your logistic regression analysis is significant or not. If the interaction is significant, then stop and report only the stratum-specific ORs
|
| ا16/2/1386 ا8:15 بعدازظهر |
|
|
آمار دوست
|
من فكر مي كنم منظور دكتر انصاري پور محاسبه Odds يك كاسه شده باشد. در چنين مواردي حدود اعتماد حاصل براي اين Odds نشان خواهد داد كه با حذف اثر يك متغير احتمالا مداخله گر متغير بعدي بر متغير پاسخ اثر معني داري داشته يا خير. بهرحال جهت بررسي وجود يا عدم وجود اثر متقابل و همچنين كنترل اثرات مداخله گرانه متغيرها استفاده از رگرسيون لجستيك ارجح است. البته بررسي مقادير OR در زيرگروهها و OR يك كاسه شده و OR كل سر نخ هايي در مورد وجود يا عدم وجود اثر مداخله گرانه متغيرها و نيز وجود اثر متقابل خواهد داد. |
| ا17/2/1386 ا9:49 صبح |
|
|
irzamin
|
Yes, I always use logistic regression rather than stratified analysis. Stratified analysis may work well when you have very few factors to study, but logistic regression can easily handle the situations when you have multiple factors especially in case-control studies |
|
|
|
|
