Showing posts with label statistics. Show all posts
Showing posts with label statistics. Show all posts

Monday, June 10, 2019

Medical Cannabis Does Not Prevent Opioid Overdoses





The political aspects of medical cannabis are undeniable. The legalization of cannabis for recreational purposes had no traction with American politicians or voters until it was promoted as a miracle drug.  With that widespread promotion medical cannabis is now legal in 33 states and recreational cannabis is legal in ten.  The legalization arguments also suggested that the US was behind other countries of the world when there are only two countries – Canada and Uruguay – where it is completely legal for medical or recreational sale and purchase.  In the world, 22 of 195 countries have legalized medical cannabis with widely varying restrictions on its use. The Netherlands is often cited as an example of recreational cannabis use, but most Americans don’t realize that it is illegal for recreational use and tolerated for use and sale only in specially licensed coffee shops.  The promotion of cannabis as a solution to the opioid overuse and chronic pain problems can be seen as an extension of the political arguments for legalization that outpace any science to back them up.

There was probably no greater hype about the purported benefits of medical cannabis than early data suggesting that it might decrease the rate of opioid overdoses (1). The sequence of events was supposed to be opioid users tapering off of opioids or using lower equivalent amounts because of medical cannabis use.  The original study covered the time period from 1999-2010 and suggested that states with medical cannabis laws had a lower mean opioid overdose mortality and that the annual rates of overdose progressively decreased over time.  The authors conclusion was:  “Medical cannabis laws are associated with significantly lower state-level opioid overdose mortality rates.”

Despite the usual caveats suggested by the authors in the original study the results of that study were heavily hyped by all cannabis promoters as was the discussion of many Internet forums.  The lay press, public, and politicians saw it as another reason to promote medical cannabis and recreational cannabis by association.

A study came out today in PNAS (2), that is an extension of the original data and it no longer comes to the same conclusion.  In this new study the authors replicated the opioid mortality estimates from the original study but when the data was extended from 2010 to 2017 – the improved opioid overdose mortality rates not only did not stay constant but they reversed themselves to that they were now on the average from -21% to +23%.  They provide an even more valuable analysis of this effect as spurious rather than a true positive or negative effect based on the low penetration of medical cannabis in the population at large (2.5%).  The authors focus on the problem of ecological fallacy – that is conclusions about individuals are drawn from aggregate data across the entire population.They point out that the states with the medical cannabis laws have a number of characteristics separating them from other states.  A recent good example of this fallacy was the New England Journal of Medicine (3,4) report that per capita chocolate consumption correlated with the number of Nobel Laureates in a particular country.  

This is a valuable lesson in scientific analysis. The political approach to the problem is all that most of the public sees. That approach is to grab any information that seems to agree with your viewpoint and run with it.  Big Cannabis and cannabis promoters have been doing this for almost 20 years now. The process of science on the other hand is slower and more deliberate.  It is not a question of a right answer but a dialogue that hopefully produces the right pathway. The authors of this study have added a lot to the dialogue about cannabis but also statistics and how statistical descriptions may not be what they seem to be. 

George Dawson, MD, DFAPA


References:

1: Bachhuber MA, Saloner B, Cunningham CO, Barry CL. Medical Cannabis Laws and Opioid Analgesic Overdose Mortality in the United States, 1999-2010. JAMA Intern Med. 2014;174(10):1668–1673. doi:10.1001/jamainternmed.2014.4005 (full text)

2:  Shover CL, Davis CS, Gordon SC, Humphreys K.    Association between medical cannabis laws and opioid overdose mortality has reversed over time.  First published June 10, 2019 https://doi.org/10.1073/pnas.1903434116  (full text)

3: Messerli FH. Chocolate consumption, cognitive function, and Nobel laureates. NEngl J Med. 2012 Oct 18;367(16):1562-4. doi: 10.1056/NEJMon1211064. Epub 2012 Oct 10. PubMed PMID: 23050509.

4:  Pierre Maurage, Alexandre Heeren, Mauro Pesenti, Does Chocolate Consumption Really Boost Nobel Award Chances? The Peril of Over-Interpreting Correlations in Health Studies, The Journal of Nutrition, Volume 143, Issue 6, June 2013, Pages 931–933, https://doi.org/10.3945/jn.113.174813


Attribution:

Above figure is from the original article (reference 2): "This open access article is distributed under Creative Commons Attribution-Non Commercial No Derivatives License 4.0 (CC BY-NC-ND).y"  See this link for full conditions of this license.



Friday, July 24, 2015

Depression and the Genetics Of Large Combinations










from:  CONVERGE consortium.  Nature. 2015 Jul 15. doi: 10.1038/nature14659. [Epub ahead of print] - see complete reference 1 below.         



This is an interesting effort from a large number of researchers looking at candidate genes in major depression. The authors studied major depressive disorder (MDD) in 5,303 Han Chinese women selected for recurrent major depression compared with 5,337 Han Chinese women screened to rule out MDD. The depressed subjects were all recruited from provincial mental health centers and psychiatric departments of general hospitals in China. The controls were recruited from patients undergoing minor surgical procedures in general hospitals or from local community centers. All of the subjects were Han Chinese women between the ages of 30 and 60 with four Han Chinese grandparents. The MDD sample had two episodes of MDD by DSM-IV criteria. The diagnoses were established by computerized assessments conducted by postgrad medical students, junior psychiatrists, or senior nurses trained by the CONVERGE team. The interview was translated into Mandarin. Exclusion criteria included other serious medical of psychiatric morbidity (see details in ref 1). 

Whole genome sequences were acquired from the subjects and 32,781, 340 SNPs were identified, 6,242,619 were included in genome-wide association studies (GWAS). Figure 1 above is the quantile-quantile plot for the GWAS analysis resulting from "a linear mixed model with genetic relatedness matrix (GRM) as a random effect and principle components from eigen-decomposition of the GRM as fixed effect covariates." I won't pretend to know what that methodology is, even after reading the Methods, Supplementary Notes section. I expect that it would take a more detailed explanation and in the era of essentially unlimited online storage capacity, I would like to see somebody post it with examples. Without it, unless you are an expert in this type of analysis you are forced to accept it at face value. I am skeptical of manipulations of data points that provide a hoped for result and can cite any number of problems related to this approach. On the other hand information of this magnitude probably requires a specialized approach. 

In this case the authors found two loci on chromosome 10 that contributed to the risk of MDD. They replicated the findings in an independent sample. 



One of the features that I liked about this paper was the focus on patients with severe depression. I have lost count of the number of papers I have read where the depression rating scores were what I consider to be low to trivial. Many rating systems used in clinics seem to use these same systems for determining who gets an antidepressant and who does not.  Whenever I see that, I am always reminded of the "biological psychiatry versus psychotherapy" debates that existed when I was in training in the 1980s.  Once of my favorite authors at the time was Julien Mendlewicz and anything he would publish in the Journal of Clinical Endocrinology and Metabolism (4-6).  There is a table in one of his studies with the HAM-D scores of the patients with unipolar depression he was seeing that ranged from 30-57 with a mean of 41+/- 10.  For bipolar patients in the same study the range was 30-43 with a mean of 36 +/- 5.  One of those patients could not be rated initially because of severe psychomotor retardation.  These are levels of depression that are not typically seen in depression research from either the standpoint of basic science and probably never for psychopharmacological research.  Much of the research that I am aware of allows for the recruitment of patients with HAM-D scores in the high teens and low 20s.  I don't think that is the best way to run experiments on biologically based depressions or antidepressant medications, but there is rarely any commentary on it.  The CONSORT group in this paper finally comments on this factor as being a useful experimental approach even though Mendlewicz was using it in the 1980s.

The second issue that crops up in the paper is replication.  The authors validate their original work by running a second sample for validation.  That is the approach we would use in analytic chemistry.  If we were using a new technique we would run samples in triplicate or in extreme cases in sets of 5 to make sure we could replicate the analysis.  It reminded of one of the first great genetic marker papers in the field that was published in the New England Journal of Medicine by Elliot Gershon's lab in 1984 (2).  It was an exciting proposition to consider that fibroblasts could be grown from a skin biopsy and the muscarinic cholinergic receptor in those fibroblasts would be a marker for familial affective disorder.   The general observation in this pilot study of 18 patients was that they had an increased muscarinic receptor density in fibroblasts compared to controls and that the relatives with histories of minor depression had receptor densities that were more similar to the subjects with mood disorders than normal controls.  The subjects with familial affective disorder were defined as subjects with bipolar I, bipolar II, or major depression according to Research Diagnostic Criteria (RDC).  No rating of depression severity was made acutely or on a historical basis.  These findings could not be replicated, in the end even by the original lab.  That process played out in the pages of the New England Journal of Medicine (3) and the original findings were withdrawn.  It would be interesting to look at how often a similar debate occurs in a prestigious journal these days.  Estimates of non-replicable findings by the pharmaceutical industry suggests that it should happen a lot more often.   

In terms of the original paper, the sheer amount of information involved in the genetic code is staggering.  Just looking at the 130 millions base pairs on Chromosome 10 and thinking about combinations of 2, 3, 4, 5, or 6 base pairs yields the numbers in the table below entitled "Combinations of 130 million base pairs."  The exponential notation ranges from 1015 to 1045 or a quadrillion  to a quattuordecillion combinations.  Figuring out the best way to determine which combinations are relevant in illnesses with polygenic inheritance will be an interesting process.
  

George Dawson, MD, DFAPA



References:

1:  CONVERGE consortium. Sparse whole-genome sequencing identifies two loci for major depressive disorder. Nature. 2015 Jul 15. doi: 10.1038/nature14659. [Epub ahead of print] PubMed PMID: 26176920.

2:  Nadi NS, Nurnberger JI Jr, Gershon ES. Muscarinic cholinergic receptors on skin fibroblasts in familial affective disorder. N Engl J Med. 1984 Jul 26;311(4):225-30. PubMed PMID: 6738616.

3:  Failure to Confirm Muscarinic Receptors on Skin Fibroblasts.  N Engl J Med 1985 Mar 28; 312: 861-862  PubMed PMID: 3974670.

4:  Linkowski P, Mendlewicz J, Kerkhofs M, Leclercq R, Golstein J, Brasseur M,Copinschi G, Van Cauter E. 24-hour profiles of adrenocorticotropin, cortisol, and growth hormone in major depressive illness: effect of antidepressant treatment. J Clin Endocrinol Metab. 1987 Jul;65(1):141-52. PubMed PMID: 3034952.

5:  Linkowski P, Mendlewicz J, Leclercq R, Brasseur M, Hubain P, Golstein J, Copinschi G, Van Cauter E. The 24-hour profile of adrenocorticotropin and cortisol in major depressive illness. J Clin Endocrinol Metab. 1985 Sep;61(3):429-38. PubMed PMID: 2991318.

6:  Mendlewicz J, Linkowski P, Kerkhofs M, Desmedt D, Golstein J, Copinschi G, Van Cauter E. Diurnal hypersecretion of growth hormone in depression. J Clin Endocrinol Metab. 1985 Mar;60(3):505-12. PubMed PMID: 4038712.


Attribution:

Extended Data Figure 1 is from: CONVERGE consortium. Sparse whole-genome sequencing identifies two loci for major depressive disorder. Nature. 2015 Jul 15.  With Permission from Nature Publishing Group  © 2015.  License number 3672900044284.

Supplementary 1:





Tuesday, January 13, 2015

JAMA Psychiatry Suicide Article, Statistics and AI

Suicide Rates - Selected OECD Countries




Suicide is a very important problem for psychiatrists.  Even though it is a rare event, it seems like most of our time is focused on preventing suicide.  There are many days where many high risk patients and patients with chronic suicidal ideation are seen in clinics and hospitals.  Most of them are treated in outpatient settings and very few are treated on an involuntary basis in hospital settings.  Since suicide is diametrically opposed to self preservation it is assumed that any rational person would want to get help with those thoughts and impulses.  Like most things in psychiatric practice it is almost never than simple.  Psychiatrists encounter a wide range of of reasons for suicidal thinking.  At times, the suicidal thinking was not obvious until it was declared after a suicide attempt.  Many people decide to see psychiatrists after a first suicide attempt.  Even at that point it is common to find a person who is disappointed that they did not succeed.  It is more common to find a person greatly relieved that they survived but even then that does not assure the cooperation necessary to prevent another attempt.

The standard of practice for assessing suicidal thinking or ideation and potential risk is risk factor analysis.  This has been the standard of practice for as long as I have practiced over the past 30 years.  To do this analysis, it requires making a diagnosis or a series of diagnoses and looking at associated factors and how the patient describes his/her mental state at the time.  Major psychiatric diagnoses like major depression, schizophrenia, bipolar disorder, panic disorder, borderline personality disorder and chronic substance use disorders all have significant lifetime prevalences of suicide varying from 3 to 15%.  Psychological autopsies of series of suicides find that nearly all of the patients who have suicided in these studies had a significant psychiatric disorder.  There are also studies done from a social science perspective that emphasize the social risk factors for suicide including sex, martial and relationship status, economic factors and loss.

Suicide is a widely misunderstood problem sometimes even for the patients who are experiencing the thoughts.  It is common for example to encounter people with suicidal thinking who say that their only deterrent to suicide is that they don't "have the guts" to do it.  An associated worry might be that it is "too painful."  They feel a need to explain why they cannot carry out an irrational act.  I take this to mean that at some point in time, the suicidal person's conscious state has changed.  They are no longer a rational person and that is why they must explain away the fact that they cannot carry out an irrational act.  Another common observation that speaks to the conscious state is that many people will say "I never understood how a person could be suicidal until I finally felt that way."  That suggests that the altered conscious state is associated with a mood state of depression or many times a mixture of depression, anger, and anxiety resulting in an agitated state that led to the understanding about suicidal thoughts.  A final observation is one of the most stressful parts of psychiatric practice and that is:  "Can I believe this person when they tell me they are not going to kill themselves?"  Much of acute care psychiatry hinges on that ultimate question.  The risk factor analysis is essentially nullified if the patient is in an emergency department and their diagnosis and past suicide attempts are known.  The only thing left to go on are the standard questions about current state of mind, deterrents, safety plans and whether the person seems reliable and says they will not kill themselves.   It is widely known that people kill themselves after leaving emergency departments and hospitals.  People have killed themselves in hospitals while under direct observation.

Many of these assessments become adversarial.  By the time a psychiatrist sees a patient in a hospital, a lot has already happened. In all of the hospitals where I have practiced, crisis teams, paramedics, and the police have assessed the person in the community and brought them in to the hospital.  Very few people were under psychiatric care at the time of that intervention.  Friends and family members of the patient were the people who called the first responders.  The patient is usually there out of some concern for their welfare that they may not be aware of.  The psychiatrist comes around sometime in the next 24 hours and the interaction unfolds.  Very few people seem interested in the fact that they might kill themselves.  Getting out of the hospital may be the priority.  Their approach might be one of non-disclosure or denial: "I really did not say I was suicidal." or "I did not mean it",  or "I was drunk or high at the time".  Even those responses can vary from very unlikely (as in a patient with a serious self inflicted gunshot wound) to unlikely (a patient with delusional depression stopped in the midst or a suicide attempt) to possible (the intoxication history with no suicidal ideation while sober).  The interview dynamic is also quite variable.  A person may be sullen, irritated, and not wanting to discuss much information.  They may express concerns about self incrimination: "I know what I can and cannot say to psychiatrists.  I know if I say the wrong thing you will lock me up and throw away the key."  They may blame their problems on the psychiatrist: "Look - I know you don't care about me.  The only thing you care about is covering your ass.  You are going to do whatever you want to do."  They may be more hostile and sarcastic: "Look if I was really going to kill myself I wouldn't be sitting here talking to you.  I'd be dead.  I wouldn't be talking about it."

All of these statements ignore the fact that the person is sitting in front of the psychiatrist as the result of the actions of several other people including persons affiliated with them and having their best interests at heart.  That situation is so intense and uncomfortable that it prevents physicians from going into psychiatry.  I  have had many physicians tell me they could not go into psychiatry because:  "Guessing about whether or not a person will kill themselves is too stressful."  There are many ways to reduce the guesswork involved but the point I am trying to make here is that all of these behaviors are consistent with the patient having undergone a change in their conscious state.  They are no longer acting like a person interested in self preservation, but they are now a person who is contemplating self destruction and taking active measures to hide that thought pattern.  That is the main reason why psychiatrists can't predict suicide over long periods of time with any degree of certainty.  When a person's conscious state changes that completely, their actions are less predictable even to the point that they may be potentially self destructive and want to cover it up.

That is also why risk factor analysis is so imperfect.  In the case of the diagnosis, a lot of clinicians are under the impression that if a person satisfies some written criteria for a diagnosis that provides a lot of critical information about the potential for suicide.  Many clinicians seem to miss the point that a patient can have the exact same written criteria for major depression with psychotic features and the same chronic markers on a suicide risk assessment and suddenly be much more likely to attempt suicide.  The only thing that has changed has been the patient's conscious state and their awareness that suicide is an unwanted state.  The evidence that this happens is clinical and ample.  Patients will report back to their psychiatrists that they were in this conscious state and the psychiatrist did or did not miss it.  Either way, there is no clinician in this situation who could make the correct call.  Without any clear markers, there is no way to figure out if this change in conscious state has occurred.  The patient usually recognizes it only in retrospect.

This clinical information on the assessment of suicide is what makes this JAMA Psychiatry article interesting.  In this article the authors attempt to determine predictors of suicide by soldiers in the year following psychiatric hospitalization within the Veteran's Administration hospital system over a 6 year period.  That was a total of 40,820 hospitalizations or 0.9% of the total Army personnel in any 12 month period.  During that time there were a total of 68 deaths by suicide.  That is number is 12% of all US Army suicides.  The authors consider a long list of potential risk factors that are largely demographic in nature to determine concentration of risk of suicide.  That list includes a law enforcement data base that clinicians do not have access to.  Their overall goal was to determine of it was practical identify high risk patients for post hospitalization intervention and whether that might be a cost effective way to prevent suicide.  They were able to identify the highest risk group - the 5% of hospitalizations in which 52.9% of the suicides occurred.  Like many similar studies the authors also comment on  how their "actuarial" methods usually trump clinicians making the same predictions.  I found very limited commentary on that fact that it is generally possible to illustrate what you want with enough variables or as we used to say "a large enough spreadsheet".  In this case they looked at a large number of variables to come up with 421 predictors for further analysis.  I have reviewed hospital records consisting of the printout of the electronic health record where there were scarcely 421 words and it was usually impossible to determine an admission or discharge date.  Any information on even a short term assessment of suicide risk is scant and it frequently says basically that the patient told us he or she was not going to make a suicide attempt.  In some cases a rating scale approach like the Columbia is used.  Clinicians using these scales are often surprised about how few variables change after the initial rating and how the numerical risk does not necessarily reflect an inpatient versus and outpatient population.

As I read through the article, I was also impressed with the amount of alien statistics and fairly esoteric statistical terms.  If JAMA Psychiatry wants to include these methods, I think an example of the calculations and a bibliography of additional reading would be a minimal requirement.  The addition of statistical reviewers' comments or an independent statistical discussion of the pros and cons of these methods would only enhance the quality of the discussion.  One of my concerns is that as the statistical methods get more abstract and vague notions about big data are more accepted, clinical complexity and wisdom are completely diluted down and out.  I saw a headline the other day that Internet sellers know more about your "personality" than your spouse.  It should be fairly obvious from all of the healthcare research done that is based on HEDIS (The Healthcare Effectiveness Data and Information Set) information, that demographic variables and product choices are not the same thing as clinical assessment and treatment.

If the headlines about artificial intelligence replacing doctors ever comes true, it will only happen if the machine can implement the required knowledge.  The performance of computers sifting through text based findings and diagnostic criteria has been know for 20 years (reference 3).   Those data points were generally far superior to demographics.  I owned 2 of those programs and they don't bother to sell them anymore.  In terms of the assessment and treatment of suicide a knowledge base included in the Harvard  Medical School Guide To Suicide Assessment and Intervention might be a step in the right direction.  A lot of that knowledge depends on the skill of a particular clinician and that includes the personality factors of clinicians who continue to do this impossible job day after day.      

Trying to predict suicide and prevent it can't currently be done with an algorithm.  If I see an algorithm I will consider why the high risk people aren't being seen in follow up from the hospital rather than who should get an intervention.   And I would not mind errors on the false positive side.


George Dawson, MD, DFAPA

1:  Kessler RC, Warner CH, Ivany C, Petukhova MV, Rose S, Bromet EJ, Brown M 3rd, Cai T, Colpe LJ, Cox KL, Fullerton CS, Gilman SE, Gruber MJ, Heeringa SG, Lewandowski-Romps L, Li J, Millikan-Bell AM, Naifeh JA, Nock MK, Rosellini AJ, Sampson NA, Schoenbaum M, Stein MB, Wessely S, Zaslavsky AM, Ursano RJ; Army STARRS Collaborators. Predicting Suicides After Psychiatric Hospitalization in US Army Soldiers: The Army Study to Assess Risk and Resilience in Servicemembers (Army STARRS). JAMA Psychiatry. 2015 Jan 1;72(1):49-57. doi: 10.1001/jamapsychiatry.2014.1754. PubMed PMID: 25390793.

2:  Douglas G. Jacobs, editor.  Harvard  Medical School Guide To Suicide Assessment and Intervention.  Jossey-Bass Inc., San Francisco, CA, 1998.

3:  Berner ES, Webster GD, Shugerman AA, Jackson JR, Algina J, Baker AL, Ball EV,Cobbs CG, Dennis VW, Frenkel EP, et al. Performance of four computer-based diagnostic systems. N Engl J Med. 1994 Jun 23;330(25):1792-6. PubMed PMID: 8190157.

Thursday, February 23, 2012

Antidepressants - the limited analysis of a polarized argument


The current President John Oldham and President-elect Jeffrey Lieberman of the American Psychiatric Association came out with this press release today on a 60 Minutes episode characterizing antidepressants as no better than placebo.  They describe this characterization as “irresponsible and dangerous reporting” and “a message that could potentially cause suffering and harm to patients with mood disorders.”

It is good to see the APA finally taking a stand on this issue.  Antidepressants and the psychiatrists who prescribe them have been taking a pounding in the popular press for years.  The main proponent here was also featured in a Newsweek headline story two years ago.  This is a prototypical example of how the media and special interest groups can distort science and facts and politicize the discussion that must be nuanced.  The problem is that you have to know something and be fairly free of bias to participate in a nuanced discussion.  Like most issues pertaining to psychiatry, the issue is always polarized and poorly discussed in the media.

I got involved in this issue as a managing editor of an Internet journal and I solicited a paper from a world renowned epidemiologist to get his current view on antidepressant meta-analyses. In order to present the entire argument I also solicited response from a world renowned psychopharmacologist with broad expertise in this field. Both articles are available online for free and I think if they are both read in total they represent the most accurate picture of antidepressant response.  Both references are listed at the bottom of this page.

Rather than get into the specific details at this point I will say that it was extremely difficult to find a anyone willing to provide a rebuttal to the to the original article by Ioannidis, but anyone who reads that paper by Davis, et al and who follows the antidepressant literature will have a greater appreciation of the effectiveness of these medications.  I hope to post some information on the statistical analysis as well.  At some level people tend to view statistics as a hard mathematical science and there is plenty of room for interpretation.  The use of meta-analysis is a common approach to these problems and a detailed look at the shortcomings of meta-analysis are seldom discussed.  That might explain why one meta-analysis shows minimal effects and another shows that there might be some antidepressants with unique effectiveness (see Cipriani, et al)

A final dimension that is critical in the analysis of any source is potential conflicts of interest.  The only conflict of interest that is typically discussed is the financial interests of authors and pharmaceutical companies in producing positive trials.  That ignores the fact that many of these trials have been very public failures and that post trial surveillance limits the use of some of these compounds.  There are other conflicts of interest to consider when an author is selling a viewpoint and can potentially profit from it – either financially or politically.

The APA could provide a valuable service here in making the documents from the FDA and the EMA widely available for public discussion and analysis.

George Dawson, MD



from a thousand randomized trials? Philos Ethics Humanit Med. 2008 May 27;3:14.

Davis JM, Giakas WJ, Qu J, Prasad P, Leucht S. Should we treat depression with drugs or psychological interventions? A reply to Ioannidis. Philos Ethics Humanit Med. 2011 May 10;6:8.

Cipriani A, Furukawa TA, Salanti G, Geddes JR, et al.  Comparative efficacy and acceptability of 12 new-generation antidepressants: a multiple-treatments meta-analysis.  The Lancet - 28 February 2009 ( Vol. 373, Issue 9665, Pages 746-758 )