OHDSI Home | Forums | Wiki | Github

The dynamics of observational data

Researchers
#1

Just to follow up on the fascinating results discussed during the population-level estimation workgroup meeting yesterday, I’ve generated this plot:

days.png1000×500 4.89 KB

It simply shows the number of visits on each day divided by the number of people whose observation period included that day (in a US insurance claims DB). @Andrew: this graph shows the variations as a function of calendar time I was referring to, and the reason why matching on calendar time makes sense. Medical events are far more likely to be recorded on some calendar days than others, and failure to adjust for this could lead to (time-varying) bias. Some things worth pursuing are: distinguishing between weekdays and weekends in the SCCS, and matching subjects on day of the week (in addition to the propensity score) in a new-user cohort design.

I remember @Patrick_Ryan and I once generating similar plots, but showing days relative to birth and death instead of calendar time, which were equally revealing. Perhaps a nice topic for a paper?

2 Likes
[Patient Level Prediction Workgroup] Meetings
#2

I love this graph! It’s always seemed obvious that there would be weekend and Christmas effects, but it’s surprising also to see the difference between Monday and Friday.

#3

Wow. That’s huge. A much larger difference than I was imagining.
The increases you found in the effects of calibration when matching by day make perfect sense to me now.
That’s really fascinating. I think it would make a great paper.

In addition to the weekday/weekend and day of the week differences it looks as though it might be worth looking for consistent dips in summer, and around midwinter and spring holidays in other data sets.
Bias when matching on data from weeks-long look back periods wouldn’t be as significant as the day-level bias, but it would be nice to see whether they’re important enough to adjust for.

#4

Yes, I think I read a paper somewhere (can’t find it right now) showing that roughly 10% (I think) of concepts captured in clinical settings have clear seasonal patterns. I think there are time-varying patterns at multiple scales to be found in the data.

#5

Great job. On these time-varying patterns, is it possible to create reasonable metrics and include them in the propensity score matching for cohort studies? E.g., total visits (scaled by patients at risk) on the index day and total visits for the observation period? And season?

#6

Ha, you wanna see REAL healthcare seasonality? Go look at a European GP
database…the summer month off goes for both patient and providers:)

1 Like
#7

Ok, to be honest I don’t have time to write a paper about this topic in the foreseeable future, but I would be very supportive if someone else cared to take the lead. For now I’ll just add it to our list of papers we could write.

t