Cross sectional study design and analysis technique for evaluating impact of COVID 19 on radiotherapy quality of care indicators

We are planning a study to investigate if the COVID 19 epidemic impacted the radiotherapy quality of care indicators for cervical cancer patients treated with curative intent radiotherapy. This is a complex treatment where the sequencing of two different radiotherapy forms (external beam radiotherapy and brachytherapy) is required, along with concurrent administration of chemotherapy. As a result, several metrics that define the quality of care (quality indicators) have been suggested - these have a known association with outcomes (e.g., patients should receive brachytherapy). Most of the QIs are boolean (present or absent) - and we have about 19 such indicators. Treatment may have been done in adherence to QI or not.

Our patient group comprises patients who have received treatment at a single center with a unified protocol. So we expect minimal differences, but we would formally like to evaluate this. We expect to have around 90 patients treated in the pandemic period and an almost equal number who were treated before the pandemic (to avoid issues related to protocol changes, I do not want to have a longer pre-pandemic time period from which the control group will be drawn). We propose to use all patients to reduce selection bias (note that the act of receiving therapy during the pandemic period has an implicit selection bias which I cannot control for). Note that patients treated in the period of a pandemic may not have COVID at all.
My question is regarding the analytic strategy for this study. I will, of course, present the unadjusted and raw data regarding each QI (proportion of patients who adhered to QI and reasons why there was no adherence). I suspect this will be useful, but I would like advice on how we can enhance the conclusion drawn.

The variables which may affect adherence to QI in addition to the treatment period are, in my opinion:

  1. Age
  2. Performance status
  3. Nutritional status (assessed using BMI - also can use albumin - may have some missing data, though)
  4. If they developed COVID 19 - this can only happen during the pandemic period.
  5. Poor response to external beam radiotherapy - in which case brachytherapy cannot be given - will influence only a few QIs.
  6. Toxicities during external beam radiotherapy - usually these lead to a delay in starting brachytherapy rather than avoiding it - duration of treatment is a QI (boolean - < 56 days or not) - will impact a few QIs not others.
  7. Patient may have logistical or financial issues with continuing therapy (may be applicable in very few patients )

Note that I am not considering the hospital’s distance as the patients start radiotherapy only after they have secured a place to stay nearby. Else we always refer them to a place nearer to their home for treatment.

Since the design a cross-sectional study with the outcome variable being adherence to QI and exposure variable being treatment during or before the pandemic, I am planning to evaluate each QI with a logistic regression where the exposure variable will be included along with some important factors like age, performance status, stage of disease and baseline BMI - known to influence our therapy decisions.

I am not keen on using a PSM technique for matching primarily because of the small sample size and the fact that estimates I derive will be more imprecise if I lose a chunk of patients - but I am open to contrary advice.

I am also planning a Poisson regression where the dependant variable will be the QI count which were adhered to for each patient (total number 19 possible range 2 - 19). I am also open to advise if a proportional odds model will be better.

Any additional advice on the design/analysis of the study would be welcome. While I would love to do this as a multicentric study, we do not have the funding to support one.
Link to proposed QIs

Link to Protocol (work in progress)

Linking to Frank’s answer in another related question prompted us to use the proportional odds model for this question. Sample size requirements for a proportional odds model