Desperate for help => Poisson regression or simple non-parametric test?

[PB_1989]

Dear Community,

Hopefully, an easy question to answer! Thanks for picking up this thread. I'm doing an evaluation on whether the use of a digital health tool (MyHealth app) changes the utilisation of healthcare services and am stumped on how to do the analysis. We have 4 groups, based on usage of the app: low, medium, high and a control group.

Hypothesis: patients who used the MyHealth app miss fewer hospital appointments / the more patients use the MyHealth app, the less likely they are to miss hospital appointments

I have data for a 12-month period in the following format as mocked-up in the table below:



I've also got aggregated data in this format:



Is there a statistical test that can tell me if there is a significant difference in the 4 groups rate of missed hospital appointments? Or do I need to do a test on the raw data (table 1)? I think a poisson regression might be in order but I can't figure out the periodicity e.g. if I need to take averages of hospital appointments over a time frame e.g. units of months.

Thank you so much for any and all help!

 

[hlsmith]

Describe the four groups please, Was the control group randomly selected to not get the app? So those patients are exchangeable with the app patients given background characteristics (e.g., overall health, disease severity). Then those that got the app you then grouped them based on usage. What if there was a confounding variable for their usage (app usage <- mild cognitive disorder -> need for additional visit. If you don't control for confounders you would ascribe the effect of something like cognitive disorder to the effect of the app. So how these patients got into these four groups is very important!

 

[Karabiner]

Why "desperate"? Which important circumstances motivate the usage of such a
strong emotional signal? I suppose we have to take it into account when trying
to assist you in solving your statistical problems?

With kind regards

Karabiner

 

[PB_1989]

Describe the four groups please, Was the control group randomly selected to not get the app? So those patients are exchangeable with the app patients given background characteristics (e.g., overall health, disease severity). Then those that got the app you then grouped them based on usage. What if there was a confounding variable for their usage (app usage <- mild cognitive disorder -> need for additional visit. If you don't control for confounders you would ascribe the effect of something like cognitive disorder to the effect of the app. So how these patients got into these four groups is very important!

Describe the four groups please, Was the control group randomly selected to not get the app? So those patients are exchangeable with the app patients given background characteristics (e.g., overall health, disease severity). Then those that got the app you then grouped them based on usage. What if there was a confounding variable for their usage (app usage <- mild cognitive disorder -> need for additional visit. If you don't control for confounders you would ascribe the effect of something like cognitive disorder to the effect of the app. So how these patients got into these four groups is very important!

Thank you very much, @hlsmith!

This analysis is retrospective i.e. we are looking at data from August 2018 - July 2019 (1 year snapshot). The healthcare provider I work for runs three similar hospitals (hospital A, B and C), serving similar populations in the same state. Patients were enrolled onto the MyHealth app at outpatient appointments at hospital A (all patients who had an appointment were enrolled). The control group is a randomly selected group of patients who had outpatient appointments in the same 1 year snapshot at hospitals B and C. We know from previous data that hospitals A, B and C have the same rate of missed appointments.

The groups are:

• Control — randomly selected patients who had at least one outpatient appointment at hospital B or C
• Low users: patients enrolled onto MyHealth app who logged on 1-5 times
• Medium users: patients enrolled onto MyHealth app who logged on 6-15 times
• High users: patients enrolled onto MyHealth app who logged on 15+ times — log on counts being within the one year timeframe

We know how many appointments patients had at an individual and at the group-level (as per the two tables above), and are specifically interested if patients who use the MyHealth app more miss fewer hospital appointments. I should have led with the fact that one of the key functionalities of the app is being able to book, view, cancel and rebook appointments. Thanks again for picking up this thread!

 

[PB_1989]

Why "desperate"? Which important circumstances motivate the usage of such a
strong emotional signal? I suppose we have to take it into account when trying
to assist you in solving your statistical problems?

With kind regards

Karabiner

Oh sorry! I should have qualified my use of such a pleading tagline. We are submitting a business case next week to try and get more funding to continue rolling out the app for more patients. As a team of non statisticians, we have googled and read around how to put our hypothesis to the test but can’t agree on the right approach — hence my post on this forum, in search of some sharper statistical minds. Thank you!

 

[Dason]

Do you still have the raw login counts for each user? Is there a particular reason you bucketed them into these categories?

 

[PB_1989]

Do you still have the raw login counts for each user? Is there a particular reason you bucketed them into these categories?

Thanks a lot, Dason. I currently only have data that describes the profile (low, medium or high user) of the patient. As far as I know, getting raw login counts may be possible but highly unlikely in time for our deadline next week , so trying to work with what I have (which I concede could be better).

 

[hlsmith]

@Dason hit on the next point i was going to make. You should procure the counts! If there is a linear relationship you would be able to see it. If not, you would want to know the shape of the relationship prior to possibly breaking the reltionships into pieces. You could be misrepresenting data by bucketing and definite!y losing information.

you also havent mentioned the patients with '0' usage? What about them? The issue being, how many of them are there and how do you know the random people are comparable to the app people if you are not familiar with the '0' people?