Advice needed about research study design (please!)

[Jumb0]

---

Members don't see this ad.

I am conducting a retrospective cohort study on recurrence rates following treatments for a type of cancer. I obtained the raw data by searching a single hospital's database of pathology reports to find all cases of the cancer that were diagnosed OR treated there during a 10 year period. If the case was a recurrence of a cancer originally treated prior to the 10 year period, it was excluded. However, I did include cases of cancers that were originally diagnosed and treated at outside hospitals within the 10 year time frame which then presented as recurrences at our hospital , where they were re-treated...I did this only when we had access to the full clinical records of the original case. To be clear, I did NOT count the recurrent cancer as a primary. Rather, I simply traced the history back to the original, primary cancer and then following it from there until the end-point (recurrence) occurred.

Is this considered a systematic / methodological error? Would it be a "purer" experiment if I only included cases that were originally diagnosed AND treated at our hospital? It dawned on me that including those outside cases might be artificially inflating the apparent recurrence rate of this cancer at our hospital since by definition the only reason these outside cases ended up in our system is because they recurred.

Is this true or am I not thinking clearly?

For the record, I can easily find the cases in question in our database and delete them if necessary. I should also mention that the primary objective of this study is to compare the recurrence rate between two treatment modalities...So, in that sense, the aforementioned issue wouldn't necessarily affect the validity of the experiment as it pertains to this aim...

Still, can the experiment rightfully be called a single institution study if I keep the methodology as is?

Thank you!

 

[IslandStyle808]

I can definitely see several 'potential' biases in your statement. However, I think you should should state what your hypothesis is first. I think this will give a clearer idea of there are any errors in methodology.

 

[aSagacious]

Hmm... interesting question. My guess is that including referred patients would introduce selection bias and artificially increase the apparent recurrence rate. Regarding your second question, the label "single institution study" is vague and arbitrary. In either case you'll enumerate your inclusion/exclusion criteria to specify.

 

[CaliforniaDreamer]

Are you trying to report the general reoccurrence rate of this type of cancer? Or just your institutions rate? What is the purpose of the study?

 

[Tots]

So if I am not mistaken: Your goal was to do a retrospective cohort study to look at recurrence rates in this particular cancer with the exposure of interest being treatment.

As described you have 2 populations:

1. Those patients treated for this cancer at your institution within 10 years, that you then classified as recurred or not. This is a standard retrospective cohort population, could be analyzed with many time-to-event methods (cox regression), and with it comes all the standard biases of this type of study.

* The one to be really cautious about is that you are missing recurrence if a patient who recurred went to another hospital instead of yours. The literature is full of ways to handle this. In addition a cancer treatment analysis like this raises a few interesting/unique problems that could be glossed over in other time-to-event analyses:

• Immortal time bias: may or may not be relevant depending on what the treatments are and and other elements of your study. It can be subtle though so best be aware of it.
• Competing Risks: If someone dies before recurring (and death is related to treatment, which invariably it is) then your analysis will need to account for this. Analyzing cancer recurrence & cancer specific mortality are the textbook examples of where competing risks really matter. There are many ways to account for this in an analysis but it can get complicated (ie multi-state models).

2. The second population you describe --> Individuals you identified as having recurred and then included if you could get information on their original diagnosis. This sounds like a case-control population (you identified people based on the outcome of interest). This is a different population than population 1. I would not recommend including the two populations together. This is a whole additional can of worms to put it lightly.

Edit: And yes no matter how you slice it (even if you included both populations which I do not recommend) this is a single institution study.

​

 

[Jumb0]

Are you trying to report the general reoccurrence rate of this type of cancer? Or just your institutions rate? What is the purpose of the study?

The purpose is to compare the recurrence rate following 2 different treatment modalities. There is a milder treatment option that we suspect has the same outcomes than the more aggressive treatment, so we are putting that to the test with a time-to-recurrence cox model analysis.

So if I am not mistaken: Your goal was to do a retrospective cohort study to look at recurrence rates in this particular cancer with the exposure of interest being treatment.

As described you have 2 populations:

1. Those patients treated for this cancer at your institution within 10 years, that you then classified as recurred or not. This is a standard retrospective cohort population, could be analyzed with many time-to-event methods (cox regression), and with it comes all the standard biases of this type of study.

* The one to be really cautious about is that you are missing recurrence if a patient who recurred went to another hospital instead of yours. The literature is full of ways to handle this. In addition a cancer treatment analysis like this raises a few interesting/unique problems that could be glossed over in other time-to-event analyses:

• Immortal time bias: may or may not be relevant depending on what the treatments are and and other elements of your study. It can be subtle though so best be aware of it.
• Competing Risks: If someone dies before recurring (and death is related to treatment, which invariably it is) then your analysis will need to account for this. Analyzing cancer recurrence & cancer specific mortality are the textbook examples of where competing risks really matter. There are many ways to account for this in an analysis but it can get complicated (ie multi-state models).

2. The second population you describe --> Individuals you identified as having recurred and then included if you could get information on their original diagnosis. This sounds like a case-control population (you identified people based on the outcome of interest). This is a different population than population 1. I would not recommend including the two populations together. This is a whole additional can of worms to put it lightly.

Edit: And yes no matter how you slice it (even if you included both populations which I do not recommend) this is a single institution study.

​

Wow, thanks for the reply. I think you really hit the nail on the head. For the record, we are also keeping track of the dates of metastasis and death, so we have all the data needed to do proper censoring. The main issue is competing risk since the average age of our cohort is very high. Indeed, it is a cancer that disproportionately affects the elderly.

Anyway, I think I concur with the idea of removing those cases. It just seems like a can of worms that I would rather not have to explain in our paper. It seems much more parsimonious to strictly include cases diagnosed and treated at our center. Thank you again for your detailed reply!