Archive | to choice your study’s design RSS feed for this section

Propensity scores

25 Jul

Propensity score gives the probability of a subject in a population to belong to a group of interest such as a treatment group.
Then comparing subjects with the same propensity scores across treatment and no-treatment groups enables the researcher to infer on the effect of the treatment regarding a given outcome even if he works on merely observational data.
But the researcher must beware of the unobserved differences between the group of interest and the comparison group created using the propensity score.
As always the relevance of the model depends on the nature of the covariates entered in it.

Garrido, M. M., Kelley, A. S., Paris, J., Roza, K., Meier, D. E., Morrison, R. S. and Aldridge, M. D. (2014), Methods for Constructing and Assessing Propensity Scores. Health Services Research. doi: 10.1111/1475-6773.12182


Case base study vs case control study

25 Jul

Unlike the case control studies the case base studies are well suited to the cross sectional extractions from the reimbursement data bases that we usually do.
The case base studies use the whole population of the database as a control group , including the subjects who are affected by the disease (ie the cases).
Thus, making no difference whether the subjects have the disease or not , the control group is far more easy to constitute.

Citation: Chui TT-T, Lee W-C (2013) A Regression-Based Method for Estimating Risks and Relative Risks in Case-Base Studies. PLoS ONE 8(12): e83275. doi:10.1371/journal.pone.0083275

Pitfalls of retrospective database studies

30 Mar

As you know a part of my work consists to participate in studies based on the extraction from retrospective databases and the analysis of the informations thus retrieved. The eligibility of the beneficiaries to the provision that represents the study’s outcome is always a major concern. There is two explanations for a beneficiary not having access to a care according to the data retrieved from the reimbursement base: either a real lack of access or a non eligibility of the care for a record in the reimbursement data base (for example if the insured is covered by another insurance or has lost his coverage and has exited from the health plan)*. I have always to keep in mind that I work on secondary data which are only a reflection of the primary data the reality of which I try to apprehend.
The dilemma is pretty well addressed in this article:

*as always there is a third possibility: the data concerning the care has been erased from or not yet recorded in the base. The timeline of the refreshment of the base (ie the loading and the purifying of the data) must be precisely described in the methodology of the study.

Article cited:
1)- Motheral, B., Brooks, J., Clark, M. A., Crown, W. H., Davey, P., Hutchins, D., Martin, B. C. and Stang, P. (2003),

A Checklist for Retrospective Database Studies—Report of the ISPOR Task Force on Retrospective Databases.

Value in Health, 6: 90–97. doi: 10.1046/j.1524-4733.2003.00242.x

Two other articles address the pitfalls of inferring from secondary data extracted from a retrospective data base:

2)- Berger M, Mamdani M, Atkins D, Johnson M.

Good Research Practices for Comparative Effectiveness Research: Defining, Reporting and Interpreting Nonrandomized Studies of Treatment Effects Using Secondary Data Sources: The ISPOR Good Research Practices for Retrospective Database Analysis Task Force Report—Part I.

Value in Health 2009 ; 12(8) :1044-52

The use of claims databases for outcomes research : Rationale, challenges, and strategies. Annual international meeting of the Association for Pharmacoeconomics and Outcome Research.

Philadelphia, Pennsylvania (USA), 1996/05/12. CLINICAL THERAPEUTICS, vol. 19, n° 2, 1997, pages 346-366, 74 réf., ISSN 0149-2918, USA. MOTHERAL (B.R.) *, FAIRMAN (K.A.). Outcomes Research. Express Scripts. Inc. Maryland Heights. USA

Full text of the article here:


Historical cohorts: strengths and weaknesses

8 Jul

Historic cohort study, generally means to take a look back at events that already have taken place.


database (Photo credit: Sean MacEntee)

With the huge data bases containing Millions of lines of historic of several years of reimbursements of health care and health conditions now at the disposal of nation wide health care insurances like CNAMTS or RSI in France or Kaiser Permanente or Veteran Affairs in the USA , historical cohorts such as the one that is described in the article referenced below are very easy to implement provided that researchers have access to the data base and use the appropriate software to extract accurately the information to transform rough data in a relevant medical information. Personally I am a fan of SAS enterprise guide (no funding by SAS to disclose for this Blog).
But nothing being perfect in this world the weakness of such historical cohorts retrospectively rebuilt is that they can only put in evidence associations without absolutely no hint but the possible causation process involved in the association. Their force is of course the number of subjects analyzed (usually huge) and the provenance of the subjects (community and real life subjects as opposed as the carefully selected subjects of the controlled randomized trials).
But at the end of the day, to conclude like the study referenced below does, that high doses of ACE treatment causes a lowering of the mortality rate and the readmission rate is obviously going beyond the proper results of the study. Indeed no observational historic cohort, whatever the size of the analysed sample is, has the power to demonstrate a causality link. One possible explanation of the association unveiled by the study is that prescribers could be more reluctant to give high doses of ACE to the more fragile groups and comorbidity incurring groups of the studied population.

More content and referenced study:


Improved Outcomes in Heart Failure Treated With High-Dose ACE Inhibitors and ARBs: A Population-Based Study: full text research letter

Immortal Time Bias

14 Dec

What the hell is an Immortal Time Bias?


The appellation is very poetic indeed. What is hidden behind?

I searched on the internet and found three papers which unveiled the mystery.

An immortal time bias occurs when individuals of one of the two groups that are compared (e.g.treated patients and control patients) are guaranteed for one period (called immortal time) to be alive if the outcome of interest is all cause mortality (or not to incur the condition of interest if the outcome is a disease). The period of immortal time must be situated after the cohort entry and before the end point (i.e. during the follow up time).

For example if date of birth is the date of cohort entry and death is the outcome of interest, Popes or Oscar Winners live longer than others. The explanation of this life time discrepancy is only the immortal time bias. You have to be alive long enough (and thus not to die) to become Pope or to win Oscars!

Below are the full text of the three papers I found on the internet treating of  the immortal time bias scope:





Bad science (part two)

26 Nov
Cover of "Bad Science"

Cover of Bad Science

Look at this incredible video. The man who is talking is a medical practitioner who has become an epidemiologist. He belongs to a Cochrane group. The Cochrane groups chase false evidence-based medicine by reviewing all the published studies on a particular public health issue (like preventive behaviors, cancer screening, pharmaceutical industry lobbying and so on). It’s an awesome one man show made by a Doctor!

Vodpod videos no longer available.

Evaluation of Scientific Publications

29 Oct
Coverage Probability of Clopper-Pearson confid...

Image via Wikipedia

In his blog named “OH-world” John Cherrie from Edinburgh, United Kingdom, signaled us an interesting series of seventeen articles freely available in full text on PubMedCentral. The first of the series is entitled Critical Appraisal of Scientific Articles; Part 1 of a Series on Evaluation of Scientific Publications.

The title of the following ones are listed below:

1. Critical Appraisal of Scientific Articles

2. Study Design in Medical Research

3. Types of Study in Medical Research

4. Confidence Interval or P-Value?

5. Requirements and Assessment of Laboratory Tests: Inpatient Admission Screening

6. Systematic Literature Reviews and Meta-Analyses

7. The Specification of Statistical Measures and Their Presentation in Tables and Graphs

8. Avoiding Bias in Observational Studies

9. Interpreting Results in 2×2 Tables

10. Judging a Plethora of p-Values: How to Contend With the Problem of Multiple Testing

11. Data Analysis of Epidemiological Studies

12. Choosing statistical tests

13. Sample size calculation in clinical trials

14. Linear regression analysis

15. Survival analysis

16. Concordance analysis

17. Randomized controlled trials

An other way to be able to evaluate a scientific article in medicine is to read the fourteen articles constituting the Clinical Chemistry Guide to Scientific Writing. The first article is entitled The Title Says It All.
The following articles are listed below:
Part 1. The Title Says It All

Part 2. The Abstract and the Elevator Talk: A Tale of Two Summaries

Part 3. “It was a cold and rainy night”: Set the Scene with a Good Introduction

Part 4. Who, What, When, Where, How, and Why: The Ingredients in the Recipe for a Successful Methods Section

Part 5. Show Your Cards: The Results Section and the Poker Game

Part 6. If an IRDAM Journal Is What You Choose, Then Sequential Results Are What You Use

Part 7. Put Your Best Figure Forward: Line Graphs and Scattergrams

Part 8. Bars and Pies Make Better Desserts than Figures

Part 9. Bring Your Best to the Table

Part 10. The Discussion Section: Your Closing Argument

Part 11. Giving Credit: Citations and References

Part 12. How to Write a Rave Review

Part 13. Top 10 Tips for Responding to Reviewer and Editor Comments

Part 14. Passing the Paternité Test

We thank Hervé Maisonneuve for having signaled this Guide in his blog.

How to evaluate a media report of a medical innovation?

15 Jul

Evaluation of the alleged medical innovation needs to respond to 10 questions:

1) availability

2) cost

3) disease mongering (exaggeration of the disease)

4) evidence (how strong are the proofs)

5) harms caused  by the innovation (how often)

6) false novelty (off label use of an old technique)

7) way of reporting benefits (relative or absolute results)

8 ) are sources appropriately quoted (who are the experts?)

9) who is promoting (conflict of interest)

10) does an alternative to this innovation exist?

Responses to these ten questions resume what an health care consumer needs to know when reading to news that report on health innovation and breaking news research findings.

Adherence to guidelines: over estimation due to social desirability bias.

4 Jul

1) Guidelines are published but in some cases are retracted because of potential conflict of interest bias.

2) Practitioner’s adherence to guidelines are published but in some cases should be examined with caution because of potential social desirability bias and interviewer bias.

3) In conclusion: nobody in this world is perfect neither those who write recommendations nor those who pretend to respect them.

We thank Pr H MaisonneuveSharib Ziya Khan,  H Sharma, and A Hakeem who gave us the idea of this post.

Make your epidemiological studies with Google®

9 Jun

Correlation is not causation but Google® which is launching Google Correlate offers to searchers the possibility of data mining in the variation across time and seasonal trends of and correlation between the search terms submitted to their search engine. Have idea to use and experiment it  (answer is reason but question is imagination) waiting for the next pandemic flu…
See also here and the publication in Nature that analysed the Google query data to detect influenza epidemics.

%d bloggers like this: