Dear Expert Peers,
I am conducting statistical tests to propose a cytokine as a predictor of tumor recurrence after surgery. First, I found the difference in cytokine levels in 2 groups (recurrence group/1 and no recurrence group/0) by Mann Whitney Test. Next, I analyzed ROC curve and yielded a significant result with AUC>0.8. Now, I want to perform univariate and multivariate analyses using SPSS. Since, I am new to SPSS and statistics both, will you kindly explain step-wise how can I run this test in SPSS. Do I need to analyze Kaplan-Meyer or COX-regression? I have data of: operation date, date of recurrence, last follows up day. The cohort is composed of 40 patients (N), and 17 patients progressed recurrence. Step-wise explanation would be really very helpful.
Unfortunately, this answer is probably going to be rather more complex than you had hoped.
1) Although it was a reasonable first step, I must point out that testing the difference in cytokine levels between the "recurrence" group and the "no recurrence" group has a serious flaw. Why? Because the next part of your answer indicates that the 40 patients have variable amounts of follow-up time, so not all patients had equal opportunity to experience "recurrence" - but this analysis gives them all an equal weight (so, by chance, if the patients with the highest cytokine levels happened to be the patients followed for the longest period of time in the study, we might see a false "relationship" between cytokine level and "recurrence" that was actually due to the patients with higher cytokine levels being followed for longer).
2) The same basic problem applies to "analyzing an ROC curve" (I assume this means that you performed logistic regression?). An ROC curve is generated by distilling your outcome down to a binary 0/1 response, but recurrence is a variable that's also dependent on how long each patient has been followed. The fact that your AUC>0.8 and is a "significant" result (more on this in a moment) means little.
3) This is why we perform survival analysis (as you allude to, asking about Kaplan-Meier or Cox regression). When you have a variable such as "recurrence" (unless you reduce the outcome to a fixed length of time like "30 day recurrence" or "1 year recurrence") it should only be analyzed with survival-analysis techniques that take into account the time-to-event, appropriately handling patients that have unequal amounts of follow up time.
Even within this, there are some serious complexities. The first problem is that, to create Kaplan-Meier curves analyzing the relationship between cytokines (which are presumably continuous variables) and recurrence, you will have to group the patients according to their cytokine levels. Many people will do this using above / below the median, or grouping patients into quartiles, but that is overly simplistic and might miss a true effect if there is some "threshold effect" at which the risk of recurrence increases substantially. It also is very difficult to consider multivariable analyses with that approach, which brings us to ANOTHER problem: a study with 40 patients and 17 recurrences is essentially limited to "descriptive" analyses only. Multivariable analysis on a cohort of that size is inappropriate (yes, you can torture SPSS or other statistical programs into doing it, but the models are likely to be wildly inaccurate).
My honest recommendation, although it may not be what you want to hear, is that you should find a qualified statistician with expertise in epidemiology or clinical-outcomes research to help you work through this problem. Even a "regular" statistician may be unfamiliar with the specific nuances of this type of project, and (although perhaps an unpopular opinion) my experience has shown that a large majority of MD's are woefully lacking the skills to perform even the most basic analysis.
Unfortunately, this answer is probably going to be rather more complex than you had hoped.
1) Although it was a reasonable first step, I must point out that testing the difference in cytokine levels between the "recurrence" group and the "no recurrence" group has a serious flaw. Why? Because the next part of your answer indicates that the 40 patients have variable amounts of follow-up time, so not all patients had equal opportunity to experience "recurrence" - but this analysis gives them all an equal weight (so, by chance, if the patients with the highest cytokine levels happened to be the patients followed for the longest period of time in the study, we might see a false "relationship" between cytokine level and "recurrence" that was actually due to the patients with higher cytokine levels being followed for longer).
2) The same basic problem applies to "analyzing an ROC curve" (I assume this means that you performed logistic regression?). An ROC curve is generated by distilling your outcome down to a binary 0/1 response, but recurrence is a variable that's also dependent on how long each patient has been followed. The fact that your AUC>0.8 and is a "significant" result (more on this in a moment) means little.
3) This is why we perform survival analysis (as you allude to, asking about Kaplan-Meier or Cox regression). When you have a variable such as "recurrence" (unless you reduce the outcome to a fixed length of time like "30 day recurrence" or "1 year recurrence") it should only be analyzed with survival-analysis techniques that take into account the time-to-event, appropriately handling patients that have unequal amounts of follow up time.
Even within this, there are some serious complexities. The first problem is that, to create Kaplan-Meier curves analyzing the relationship between cytokines (which are presumably continuous variables) and recurrence, you will have to group the patients according to their cytokine levels. Many people will do this using above / below the median, or grouping patients into quartiles, but that is overly simplistic and might miss a true effect if there is some "threshold effect" at which the risk of recurrence increases substantially. It also is very difficult to consider multivariable analyses with that approach, which brings us to ANOTHER problem: a study with 40 patients and 17 recurrences is essentially limited to "descriptive" analyses only. Multivariable analysis on a cohort of that size is inappropriate (yes, you can torture SPSS or other statistical programs into doing it, but the models are likely to be wildly inaccurate).
My honest recommendation, although it may not be what you want to hear, is that you should find a qualified statistician with expertise in epidemiology or clinical-outcomes research to help you work through this problem. Even a "regular" statistician may be unfamiliar with the specific nuances of this type of project, and (although perhaps an unpopular opinion) my experience has shown that a large majority of MD's are woefully lacking the skills to perform even the most basic analysis.
Dear Professor thak you very much for your valuable suggestions, and i completely agree with the first statement
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics