Advice regarding a Directed Acyclic Graph and further analysis Plan

[S_Chakraborty]

We are planning a study to evaluate the impact of diabetes on acute toxicities experienced by patients with cervical cancer who are receiving concurrent chemotherapy and radiotherapy. Acute toxicities refer to toxicities that occur during treatment. Clinically this is an important issue as patients with severe acute toxicity often end up receiving less treatment or breaks in treatment.

In preparation for the analysis of the study we have prepared a DAG which is shown here in the image (prepared in excalidraw hence not exactly a proper diagram)

To explain the relationships:

1. Old age and obesity increase the risk of having a diabetes
2. Patients who are older may have a poorer tolerance to radiotherapy and chemotherapy (evidence is modest) which leads to increased risk of GI toxicity
3. Radiotherapy and chemotherapy together can cause GI Toxicity
4. Chemotherapy can cause Renal Damage.
5. Renal damage results in dyselectrolytemia and some forms of dyselectrolytemia result in further exacerbation of renal damage.
6. GI Toxicity often results in renal damage due to dehydration resulting from diarrhea and vomiting. This can also cause dyselectrolytemia.
7. Diabetes causes malnutrition and inflammation and these states may also be associated with increased radio-sensitivity and thereby worse radiation toxicities

Note that in the diagram the radio-sensitivity link is to GI Toxicity and not directly to renal damage as during radiotherapy we ensure a very low dose of radiation to kidney which is associated with a low clinically relevant renal toxicity during the treatment.

They study we are planning is a retrospective cross sectional study of patients treated for cervical cancer. The link to the registered study protocol in OSF is OSF. While the primary endpoint is the prevalance of toxicities in the patients with diabetes, we would also like to ascertain the association using a multivariable regression model to obtain adjusted estimates.

We need advice regarding the appropriate analytical strategy to use so that we can get reasonably reliable estimates given the complex relationships (as an aside we are selecting patients treated with a homogenous treatment protocol where radiotherapy is delivered using a single technique to same dose levels).
Current Strategy:

We are planning to develop two regression models where acute toxicity (the dependant variable) will be evaluated. Method one will use a maximum grade method and use a ordinal regression model. Method two will use a toxicity index (weighted sum of ordered toxicity scores) and will also use ordinal regression.

The following independent variables are planned to be included in the model:

1. Age in years
2. Height in cm
3. Weigh in kg
4. Baseline serum albumin (marker of nutrition)
5. Baseline serum hemoglobin (anemia can be nutritional and associated with chronic disease).
6. Baseline total leucocyte count (marker of nutrition and immune function)
7. Baseline serum creatinine (marker of baseline renal function)
8. Baseline neutrophil count (as with leucocyte count marker of nutrition and immune function)
9. Volume of the plannig target (The high dose radiation volume)
10. Use of simultaneous integrated boost (yes / no)
11. Comorbidity

Comorbidity will be coded as follows:

1. No comorbidity
2. DM only
3. DM + Other comorbidity
4. Other comorbidity without DM

As old age is associated with diabetes we have planned an interaction term with age and comorbidity. Continous variables like age, height, weight, serum albumin, hemogllobin etc will be modelled with restricted cubic splines.

The dependant variable - toxicity will be modelled as an ordinal variable to retain the maxium information and we will develop models for toxicity as a cumulative as well seperately for toxicities which belong to families:

1. GI : Diarrhea, nausea, vomiting
2. Dyselectrolytemia : Various electrolyte abnormalities
3. Hematological toxicities
4. Infections

We estimate we will have around 350 - 400 patients available for analysis. Given the nature of treatment, almost everyone is expected to have some degree of toxicity with approximately 10 - 20% having severe toxicity (which will reflect in higher scores).

Note that this study would be interesting as India is currently the diabetes capital of the world and we estimate that about 25 - 30% of our patients have diabetes !

[davidcnorrismd]

What patient-centered aims would this study serve? What (prospective) clinical decision-making would it influence? Are there scientific questions this work might help answer? I ask these basic questions because I have the impression that purely technical questions have gotten ahead of what ought to be the ultimate ends of such work.

[S_Chakraborty]

Thank you for these excellent questions. Diabetes is a very important comorbidity and our empirical observation has been that these patients are at a higher risk of developing adverse effects during chemoradiation. A future prospective study would be designed based on this study to understand a comprehensive diabetes management strategy during the treatment of these patients. We need to understand the toxicities that need particular attention so that we can define the strategies which would probably make the greatest difference. We may also design studies where additional dose constraints are used for radiotherapy planning specifically for these patients based on the information gleaned from this study.

[ESMD]

Would it be fair that the goals of your study could be summarized this way?

1. Conduct a observational study to substantiate your clinical impression/suspicion that diabetic patients with cervical cancer seem to suffer more treatment-related toxicity and treatment interruptions than non-diabetic patients;
2. Use the study to quantify the burden and spectrum of treatment-related toxicities among diabetic patients;
3. If the study suggests a high prevalence of particular treatment-related toxicities among diabetic patients (e.g., GI toxicity), then you would consider (subsequently) conducting an RCT in which diabetic patients were randomized to two different treatment approaches (e.g., reduced chemo or RT doses vs standard doses, or prophylactic treatment with medications designed to prevent GI toxicity vs no prophylactic treatment), with the goal of reducing GI toxicity while at the same time showing that you could maintain treatment efficacy (e.g., ensure that overall mortality remains unaffected by treatment modification)?

[S_Chakraborty]

Thanks. That would be something to aspire to in the future. Given the fact that these are locally advanced disease we would be investigating ways which can ameliorate toxicities more than dose de-escalation of therapy. Additionally we are hoping we will be able to come with a more comprehensive support strategy for these patients emphasizing pre-treatment diagnostic workup, nutritional optimization, glycemic control and monitoring as well as increased emphasis on integration of physical activity and exercise. These are of course just nascent thoughts at the moment and we are hoping that the study will enable to get together stakeholders from endocrinology, nutritional sciences, nephrology, and physical therapy to design these strategies in a manner which will be affordable for the patients in our part of the world.

[ESMD]

@S_Chakraborty

These are of course just nascent thoughts at the moment

Can you flesh out your aspirations a bit more? It feels like thinking through concrete next steps, given various potential findings from your observational study would be important in guiding your study design.

For example, what if your study identified an “association” between diabetes and treatment-related GI toxicity. Reasonable questions would be:

• What would you consider to be a clinically compelling association, and why? Does your definition of “clinically compelling” jibe with that of other important stakeholders?

• What strength of association would health care systems find sufficiently compelling to pay for any treatment modifications or special treatment for diabetic patients going forward (have you asked them?)

• Do you already have in mind certain treatment modifications that you might want to apply if your study were to confirm that diabetic patients seem to have more e.g., treatment-related GI (or other) toxicity? If so, 1) how much do they cost?; 2) if the current barriers to implementing these modifications empirically (i.e, without any further study) are financial, then will policy-makers be satisfied with observational evidence of “associations” between treatment toxicity and diabetes, OR would they only be likely to pay for treatment modifications for diabetic patients if RCT evidence were to become available to show an incremental benefit from such modifications? If so, would your group be able to secure funding for such a study?; 3) if you are unlikely to be able to secure future RCT funding to provide “proof” to funders that they should pay for special treatment for patients with diabetes, then would you feel that your non-RCT evidence of diabetes-related treatment toxicity would be sufficiently compelling for you to ask patients to pay for these modifications out of their own pocket?

Your observations and work sound very important. It’s just that it very often feels that observational research is conducted before “next steps” are fully thought through, and this often makes it less useful that it otherwise might have been…

[S_Chakraborty]

Thank you once more. This is actually very helpful.

1. I would consider a clinically compelling case if an odds ratio of 1.5 - 2 or higher is seen in patients with diabetes.
2. The healthcare system in India actually does not have a single payer - for vast majority of cases, patients pay out of pocket or through private insurance.This is a very important distinction as compared to other healthcare systems. For us it is actually very important to reduce acute toxicities simply because the costs of managing these acute toxicities through hospitalisation is prohibitively high. To illustrate the point the cost differential between 3-dimensional conformal radiotherapy and intensity modulated radiotherapy in our setting is about Rs 40,000 - Rs 80,000 (depending on weather the patient’s treatment is being partly subsidized). The cost of admission with 3rd generation antibiotic for toxicity management for 3 days would be about double this. We would be actually able to get the direct cost component from the hospital billing records which should allow us to do the costing exercises after the study.
3. We have certain modifications in mind - starting off with ensuring that all patients with diagnosed diabetes are seen by a nutritionist and endocrinologist (the data will help us to emphasize to the administration that we need these people for the patients). Also starting a exercise program for these patients would be a reasonably low hanging fruit to tackle. Better glycemic control by emphasizing compliance to medications and glycemic monitoring during treatment would be possibly important. On the radiotherapy part further dose reduction to the bowel and bone marrow can be technically feasible.
4. Funding for RCT actually would be possible if we can actually demonstrate the potential burden of toxicity and formulate a “package” which includes all of the above elements - and given the rich event rate it would possibly be a very doable trial.

Practically we are working in a cancer center and this is a disadvantage as we do not have an academic endocrinology department per se. I am also hoping that this study will give us some data with which we can approach academic medical colleges with dedicated endocrinology departments where I can probably find out better strategies for generating this package intervention.

[Pavlos_Msaouel]

It may be easier to use software such as dagitty for your DAGs. You also may find of use the practical examples on how to use DAGs for oncology research here. Note that the graph in your original post is not a DAG as it includes a bidirectional arrow.

The DAG encodes your causal assumptions and helps you identify which variables to include in your statistical model. See also here for more details. Since they are non-parametric, DAGs will not help you on deciding whether to use a linear vs ordinal model etc.

As a practical example for your purposes let’s take the following hypothesis you expressed above:

@S_Chakraborty

Diabetes is a very important comorbidity and our empirical observation has been that these patients are at a higher risk of developing adverse effects during chemoradiation.

To make it more tractable and pertinent to this specific question, I have collapsed in dagitty your original graph into the following DAG to test the hypothesis that diabetes influences the development of adverse events during chemoradiation:

It becomes therefore clear that “old age” is a confounder of the relationship between diabetes and toxicity and should definitely be included in your statistical model. Radiosensitivity is a mediator and, for this specific question, if you include it in a standard regression model then you will actually block part of the effect you are interested in. It will also help to include the type of radiotherapy and chemotherapy in your model to reduce outcome heterogeneity and thus increase power.

Conversely, including obesity in your statistical model may actually reduce power or even increase bias if either you do not include “old age” in your model or there are other unaccounted confounders for the relationship you are investigating.

Hope this helps. Good luck in your study!