Journal Club Q+A: May 2016

Association of emergency department opioid initiation with recurrent opioid use. (NCBI Article Link)

Check out table 3.  What characteristics appear to be protective against developing recurrent opioid use?  Does their definition of “recurrent opioid use” have face validity?  (i.e. – does it make sense that these characteristics are associated with people who are less likely to develop recurrent use habits?)

Self-pay, Hispanic patients presenting with painful “injury,” extremity, chest, or ENT/dental complaints are actually less likely to develop opioid dependence than patients with insurance, other races, or other complaints.  One could argue that patients with no insurance to help pay for narcotics would be less likely to seek meds they are less likely to afford, so the finding that self-pay patients may support the validity of the definition.  Furthermore, it is possible that cultural factors in Hispanic patients make them less inclined to utilize prescription medications in pain relief, so this also could support the face validity of the definition.  However the fact that pain in certain different anatomic locations is associated with decreased risk of developing dependence doesn’t make sense.  Mu receptors are not just localized in some parts of the body.  This is an inexplicable finding, so there is probably something wrong with the authors’ definition.

Naproxen With Cyclobenzaprine, Oxycodone/Acetaminophen, or Placebo for Treating Acute Low Back Pain: A Randomized Clinical Trial. (NCBI Article Link)

The authors note that patients who took the study drug more than once and who were randomized to oxycodone reported moderate to severe pain less often than patients in other arms of the study.  However, they caution that this finding should be “interpreted cautiously because of the large number of analyses we performed.”  Why would “a large number of analyses” call a result into question?  It might help to consider some of the problems that arise from the practice of data mining.

On page 1578, the authors note that this finding was “one of multiple exploratory comparisons.”  Essentially, they were breaking their data set into smaller groups and comparing the groups in order to identify a new pattern, which is a definition of data mining.  Data mining can be used to find previously unanticipated relationships, however it is almost as likely to discover spurious correlations.  A finding such as the one these authors report is typically followed up with a targeted prospective study to determine if the pattern is reproducible or spurious.  (Check out http://www.tylervigen.com/spurious-correlations for good examples of this phenomenon.)

Oligoanalgesia in blunt geriatric trauma. (NCBI Article Link)

Univariate analysis is used in this study to identify possible variables that influence analgesic administration independent of the variable of interest (age).  Which possible confounders did they identify?  Are these confounders independent or interrelated?  Logistic regression analysis was used to control for these confounders – is this a statistically valid test in this case?

Table 2 shows the results of the univariate analysis, in which “pain scale not recorded,” GCS, and alcohol were discovered to be possible confounders.  In other words, the administration of analgesia seemed to depend on these factors regardless of age.  It is easy to see how all three of these variables could be interrelated.  A significantly intoxicated person might be expected to have a reduced GCS and to offer challenges in extracting a pain rating.  Since a multivariate logistic regression assumed the variables being controlled for are all independent, this is not the most valid statistical test to perform in this case.  ANOVA (analysis of variants) is likely a more appropriate test in this case.