A Closer Look At The Pfizer Enigma
Using multivariate logistic regression to investigate under-reporting bias
In part 4 of my Digging Deeper Into EudraVigilance series we noted an extraordinarily curious result whereby Pfizer jabs were going unlisted as a suspected drug exclusively in the 18 - 64y age group. This morning I decided to run analyses once again but using records for this age group alone.
Pasting another bunch of similar tables didn’t fill me with as much joy this time round so I decided to reach for logistic regression with PfizerZERO as my binary dependent1 and the following as my remaining six independents:
Economic Region
Primary Source Qualification
COVID-19 Diagnosis
Concomitant Drug Use
Total Drug Entries
Total Reactions Recorded
For good measure I also threw in all interactions between these six variables up to the level of 3-way (35 interactions in total), the whole being subject to forward selection using the likelihood ratio. Herewith some stats output to munch through:
ROC curve (area under curve = 0.853)
Thinking…
This isn’t a bad fitting model given the impossible task of predicting when Cominarty is going to be left off the suspected drug listing, though Hosmer & Lemeshow returned a disappointing Chi-square of 64.2 on 8 d.f. (p<0.001). Decent levels of classification boosted my confidence, along with the ROC analysis, though we have to accept we are skating on thin ice.
As for variables in the equation I note that only one 3-way interaction made it through, this being Total reactions recorded by EU region by Total drug entries (OR = 0.93, 95%CI[0.90;0.96], p<0.001). This refinement indicates a reduction in the tendency for Cominarty to go unsuspected in non-EEA regions when cases are particularly complex.
I am hoping that subscribers are sufficiently knowledgeable to unpick the gems within the final predictive equation otherwise I am going to have to write several paragraphs! A treat that caught my eye is COVID-19 diagnosis by Report source (OR = 0.41, 95%CI[0.21;0.80], p=0.009) which indicates a marked reduction in the tendency for Cominarty to go unsuspected for COVID cases within non-EEA regions; that is to say non-EEA regions are hot on nailing Cominarty as a suspected drug if folk fall foul of COVID after vaccination but not otherwise! Why Cominarty and only Cominarty is anyone’s guess. Another treat is Report source by EU region (OR = 10.39, 95%CI[5.65;19.09], p<0.001); there’s your culprit right there: non-healthcare professionals in non-EEA regions who tend to overlook listing of Cominarty as a suspected drug when compared to healthcare professionals in EEA regions. The good news is that I am able to correct those 2,085 unlisted reactions to Pfizer’s product and (hopefully) produce less biased analyses.
Kettle on!
[0] - Listed; [1] = Unlisted