SlideShare a Scribd company logo
1 of 43
Science = generalizable knowledge

Predictive and reliable information

Collected in some subjects and generalized to others

Sampling problems often exists, but not always
Qualitative research methods

Are only meaningful without sampling problems
(constant outcome, deterministic events)

or when sampling problems are irrelevant
(one observation is sufficient to reject the hypothesis)
Quantitative methods (statistical)
Are used to quantify sampling uncertainty

Sampling uncertainty is caused by variability

Statistics is primarily about variability
The confusing EQ-5D index
      A study of variability
Swedish Knee Arthoplasty Register
The distribution is important when discussing improvement




9/24/11
The Poisson distribution
                Defined by: λ




            Siméon Poisson (1781–1840)
9/24/11
The Gaussian distribution
          Defined by: µ and σ (the latter usually assumed constant)




                    Abraham de Moivre (1667–1754)
9/24/11
Empirical EQ-5D distribution




     Xie F Li S-C, Luo N, LO N-N,Yeo S-J,Yang KY, Fong KY, Thumboo J. Comparison
     of the EuroQol and Short Form 6D in Singapore Multiethnic Asian Knee
     Osteoarthritis Patients Scheduled for Total Knee Replacement. Arthritis &
     Rheumatism (Arthritis Care & Research) 2007;57:1043–1049

9/24/11
A Gaussian mixture distribution?
            Defined by: µ1,µ2,σ1,σ2 and w




9/24/11
Hospital A              Hospital B

                            87%
            13%                       70%     30%




      !

              Mean = 0.58           Mean = 0.58
              SD = 0.21             SD = 0.21

9/24/11
Studying change in EQ-5D
   It has been suggested that pairwise differences
   between pre- and postoperative EQ-5D values are
   normally distributed and can be meaningfully
   interpreted.




9/24/11
Studying change in EQ-5D
   It has been suggested that pairwise differences
   between pre- and postoperative EQ-5D values are
   normally distributed and can be meaningfully
   interpreted.

   It can easily be shown that this is not correct

   The sum of two bimodal distribution has a distribution
   with three modes, the difference four.




9/24/11
Empirical EQ-5D data from knee patients in Trelleborg 2007-2008




          Preop EQ-5D                        Postop EQ-5D




                                              Delta EQ-5D
          Delta EQ-5D
9/24/11
9/24/11
9/24/11
9/24/11
9/24/11
Additional problem with analyses of change
  Change is confounded by association with baseline

  X = pre-operative (baseline) value
  Y = postoperative (follow up) value

  Y-X correlates with X

  Solution

  When analyzing change, adjust for imbalance at baseline

  (This is an almost perfect case-mix adjustment!)




9/24/11
1. Stockholm


          2. Kronoberg   2. Gävleborg




                         3. Östergötland




9/24/11
2. Gävleborg

      1. Stockholm
                         3. Östergötland
          2. Kronoberg




9/24/11
9/24/11
EQ-5D Problems
Conventional analyses

Mean values not interpretable

Confidence intervals not reliable (Calculated assuming Gaussian
 distribution)

P-values not reliable (Student's t-test, ANOVA, etc. requires
  Gaussian distribution and homogeneous variance)




9/24/11
EQ-5D Problems cont'd

Non-parametric analysis

Median value may not exist.

Confidence intervals not reliable (calculated assuming Gaussian
 or binomial distribution).

P-values not reliable (Wilcoxon's MPSR-test requires a
  symmetrical distribution, Mann-Whitney U-test requires
  distributions with identical shape.




9/24/11
EQ-5D Problems cont'd
Adjusting for baseline

How meaningful is the outcome of an ANCOVA with
variables having non-Gaussian, multimodal distributions
(with different number of modes)? What do these
residuals look like?




9/24/11
EQ-5D Problems cont'd
Alternative analyses methods?

- Mixture distribution analysis (mixdist library for R)

- Multi-state Markov analysis (msm library for R)




9/24/11
9/24/11
9/24/11
9/24/11
9/24/11
“This is about clinical improvement, not science”




9/24/11
Swedish law defines clinical improvement work
(CIW) as “not research”

Some CIW projects include experiments on patients

- No ethics approval is required (or can be applied for)
- No informed consent
- No scientific planning or evaluation of the experiments
- No formal publication of studies and results




9/24/11
Regression analysis

  - Adjusting for baseline

  - Models only including statistically significant factors

  - Stepwise regression methods




9/24/11
What factors should be included in a
 linear model (ANCOVA)?
 Y = b0 + b1X1 + b2X2 + … + bnXn + e

 This is a multiple or multivariable analysis but not multivariate.

 Xi is a variable (factor or covariate)
 bi is the effect on Y of one unit change in Xi

 Assume that Y is blood pressure and X1 an indicator of anti-
 hypertensive treatment. bi will then estimate the treatment effect in
 terms of blood pressure reduction.




9/24/11
Linear models
 Answer

 It depends on a) the purpose of the study and b) the study design
 used.

 1. Purpose: (black-box) prediction

 Any variable can be included as long as it increases the sensitivity and
 specificity of the prediction, and as long as results (bi) are not
 interpreted in terms of causal effects.

 2. Purpose: effect estimation

 The variables needed to produce valid (bi and their s.e.) should be
 included.



9/24/11
Linear models
 1. Common for all designs

 Include baseline when analyzing change in a continuous variable.

 2. Randomized trial

 Include randomization stratification factors (for valid standard errors).

 3. Observational study

 Include potential confounding factors (for valid regression coefficients).




9/24/11
Linear models
 How should confounding factors be included?

 1. By the investigator's reasoning.

 2. By reviewing other publications on the same endpoint.

 3. By performing sensitivity analyses.

 4. But not by using hypothesis testing or stepwise regression analysis.




9/24/11
Parsons et al. A systematic survey of the quality
   of research reporting in general orthopaedic
   journals. J Bone Joint Surg Br 2011;93-B,1154-9




9/24/11
9/24/11
9/24/11

More Related Content

Viewers also liked

8 steps to success
8 steps to success8 steps to success
8 steps to success
Winalite
 

Viewers also liked (16)

Winalite Opportunity
Winalite OpportunityWinalite Opportunity
Winalite Opportunity
 
Winalite business briefing (wbb)
Winalite business briefing (wbb)Winalite business briefing (wbb)
Winalite business briefing (wbb)
 
Love Moon
Love MoonLove Moon
Love Moon
 
Lund 2010
Lund 2010Lund 2010
Lund 2010
 
8 steps to success
8 steps to success8 steps to success
8 steps to success
 
Open source-options-v1
Open source-options-v1Open source-options-v1
Open source-options-v1
 
Breastfeeding Education
Breastfeeding EducationBreastfeeding Education
Breastfeeding Education
 
Integrating php withrabbitmq_zendcon
Integrating php withrabbitmq_zendconIntegrating php withrabbitmq_zendcon
Integrating php withrabbitmq_zendcon
 
Datavalidering jr1
Datavalidering jr1Datavalidering jr1
Datavalidering jr1
 
Cedera olahraga
Cedera olahragaCedera olahraga
Cedera olahraga
 
Cloud Foundry Bootcamp
Cloud Foundry BootcampCloud Foundry Bootcamp
Cloud Foundry Bootcamp
 
Cloud Messaging With Cloud Foundry
Cloud Messaging With Cloud FoundryCloud Messaging With Cloud Foundry
Cloud Messaging With Cloud Foundry
 
Scaling webappswithrabbitmq
Scaling webappswithrabbitmqScaling webappswithrabbitmq
Scaling webappswithrabbitmq
 
Interoperability With RabbitMq
Interoperability With RabbitMqInteroperability With RabbitMq
Interoperability With RabbitMq
 
Writing testable code
Writing testable codeWriting testable code
Writing testable code
 
Malmo 17.10.2008
Malmo 17.10.2008Malmo 17.10.2008
Malmo 17.10.2008
 

Similar to Actalecturerungsted

anovappt-141025002857-conversion-gate01 (1).pdf
anovappt-141025002857-conversion-gate01 (1).pdfanovappt-141025002857-conversion-gate01 (1).pdf
anovappt-141025002857-conversion-gate01 (1).pdf
GorachandChakraborty
 
1 lab basicstatisticsfall2013
1 lab basicstatisticsfall20131 lab basicstatisticsfall2013
1 lab basicstatisticsfall2013
TAMUK
 
Chapter 11 Chi-Square Tests and ANOVA 359 Chapter .docx
Chapter 11 Chi-Square Tests and ANOVA  359 Chapter .docxChapter 11 Chi-Square Tests and ANOVA  359 Chapter .docx
Chapter 11 Chi-Square Tests and ANOVA 359 Chapter .docx
bartholomeocoombs
 
Metanalysis Lecture
Metanalysis LectureMetanalysis Lecture
Metanalysis Lecture
drmomusa
 
Approximate ANCOVA
Approximate ANCOVAApproximate ANCOVA
Approximate ANCOVA
Stephen Senn
 

Similar to Actalecturerungsted (20)

Javier Garcia - Verdugo Sanchez - Six Sigma Training - W2 Simple Variance Ana...
Javier Garcia - Verdugo Sanchez - Six Sigma Training - W2 Simple Variance Ana...Javier Garcia - Verdugo Sanchez - Six Sigma Training - W2 Simple Variance Ana...
Javier Garcia - Verdugo Sanchez - Six Sigma Training - W2 Simple Variance Ana...
 
Analysis of Covariance.pptx
Analysis of Covariance.pptxAnalysis of Covariance.pptx
Analysis of Covariance.pptx
 
Oac guidelines
Oac guidelinesOac guidelines
Oac guidelines
 
anovappt-141025002857-conversion-gate01 (1).pdf
anovappt-141025002857-conversion-gate01 (1).pdfanovappt-141025002857-conversion-gate01 (1).pdf
anovappt-141025002857-conversion-gate01 (1).pdf
 
Theory of Probability-Bernoulli, Binomial, Passion
Theory of Probability-Bernoulli, Binomial, PassionTheory of Probability-Bernoulli, Binomial, Passion
Theory of Probability-Bernoulli, Binomial, Passion
 
Thesis Defense
Thesis DefenseThesis Defense
Thesis Defense
 
2.AA.anova sesion applied biostat III (2).ppt
2.AA.anova sesion applied biostat III (2).ppt2.AA.anova sesion applied biostat III (2).ppt
2.AA.anova sesion applied biostat III (2).ppt
 
A Study of Some Tests of Uniformity and Their Performances
A Study of Some Tests of Uniformity and Their PerformancesA Study of Some Tests of Uniformity and Their Performances
A Study of Some Tests of Uniformity and Their Performances
 
Hypothesis Test _Two-sample t-test, Z-test, Proportion Z-test
Hypothesis Test _Two-sample t-test, Z-test, Proportion Z-testHypothesis Test _Two-sample t-test, Z-test, Proportion Z-test
Hypothesis Test _Two-sample t-test, Z-test, Proportion Z-test
 
1 lab basicstatisticsfall2013
1 lab basicstatisticsfall20131 lab basicstatisticsfall2013
1 lab basicstatisticsfall2013
 
Variance component analysis by paravayya c pujeri
Variance component analysis by paravayya c pujeriVariance component analysis by paravayya c pujeri
Variance component analysis by paravayya c pujeri
 
CABT SHS Statistics & Probability - Mean and Variance of Sampling Distributio...
CABT SHS Statistics & Probability - Mean and Variance of Sampling Distributio...CABT SHS Statistics & Probability - Mean and Variance of Sampling Distributio...
CABT SHS Statistics & Probability - Mean and Variance of Sampling Distributio...
 
Undergraduate Research work
Undergraduate Research workUndergraduate Research work
Undergraduate Research work
 
Chapter 11 Chi-Square Tests and ANOVA 359 Chapter .docx
Chapter 11 Chi-Square Tests and ANOVA  359 Chapter .docxChapter 11 Chi-Square Tests and ANOVA  359 Chapter .docx
Chapter 11 Chi-Square Tests and ANOVA 359 Chapter .docx
 
Metanalysis Lecture
Metanalysis LectureMetanalysis Lecture
Metanalysis Lecture
 
Approximate ANCOVA
Approximate ANCOVAApproximate ANCOVA
Approximate ANCOVA
 
Anova ppt
Anova pptAnova ppt
Anova ppt
 
2.0.statistical methods and determination of sample size
2.0.statistical methods and determination of sample size2.0.statistical methods and determination of sample size
2.0.statistical methods and determination of sample size
 
Chapter6.pdf.pdf
Chapter6.pdf.pdfChapter6.pdf.pdf
Chapter6.pdf.pdf
 
"intelligent" intelligence testing: Evaluating wihtin CHC domain test score ...
"intelligent" intelligence testing:  Evaluating wihtin CHC domain test score ..."intelligent" intelligence testing:  Evaluating wihtin CHC domain test score ...
"intelligent" intelligence testing: Evaluating wihtin CHC domain test score ...
 

More from Jonas Ranstam PhD (20)

The SPSS-effect on medical research
The SPSS-effect on medical researchThe SPSS-effect on medical research
The SPSS-effect on medical research
 
Sof stat issues_pro
Sof stat issues_proSof stat issues_pro
Sof stat issues_pro
 
Sof klin forsk_stat
Sof klin forsk_statSof klin forsk_stat
Sof klin forsk_stat
 
Rcsyd pres nara
Rcsyd pres naraRcsyd pres nara
Rcsyd pres nara
 
Odense 2010
Odense 2010Odense 2010
Odense 2010
 
Oarsi jr1
Oarsi jr1Oarsi jr1
Oarsi jr1
 
Oac beijing jr
Oac beijing jrOac beijing jr
Oac beijing jr
 
Norsminde 2009
Norsminde 2009Norsminde 2009
Norsminde 2009
 
Nara guidelines-jr
Nara guidelines-jrNara guidelines-jr
Nara guidelines-jr
 
Malmo 30 03-2012
Malmo 30 03-2012Malmo 30 03-2012
Malmo 30 03-2012
 
Lund 2009
Lund 2009Lund 2009
Lund 2009
 
London 2008
London 2008London 2008
London 2008
 
Lecture jr
Lecture jrLecture jr
Lecture jr
 
Karlskrona 2009
Karlskrona 2009Karlskrona 2009
Karlskrona 2009
 
Copenhagen 2008
Copenhagen 2008Copenhagen 2008
Copenhagen 2008
 
Brussels 2010
Brussels 2010Brussels 2010
Brussels 2010
 
Amsterdam 2008
Amsterdam 2008Amsterdam 2008
Amsterdam 2008
 
Abc4
Abc4Abc4
Abc4
 
Umeapresjr
UmeapresjrUmeapresjr
Umeapresjr
 
Stockholm 6 7.11.2008
Stockholm 6 7.11.2008Stockholm 6 7.11.2008
Stockholm 6 7.11.2008
 

Actalecturerungsted

  • 1. Science = generalizable knowledge Predictive and reliable information Collected in some subjects and generalized to others Sampling problems often exists, but not always
  • 2. Qualitative research methods Are only meaningful without sampling problems (constant outcome, deterministic events) or when sampling problems are irrelevant (one observation is sufficient to reject the hypothesis)
  • 3.
  • 4. Quantitative methods (statistical) Are used to quantify sampling uncertainty Sampling uncertainty is caused by variability Statistics is primarily about variability
  • 5. The confusing EQ-5D index A study of variability
  • 6.
  • 7.
  • 8.
  • 9. Swedish Knee Arthoplasty Register The distribution is important when discussing improvement 9/24/11
  • 10. The Poisson distribution Defined by: λ Siméon Poisson (1781–1840) 9/24/11
  • 11. The Gaussian distribution Defined by: µ and σ (the latter usually assumed constant) Abraham de Moivre (1667–1754) 9/24/11
  • 12. Empirical EQ-5D distribution Xie F Li S-C, Luo N, LO N-N,Yeo S-J,Yang KY, Fong KY, Thumboo J. Comparison of the EuroQol and Short Form 6D in Singapore Multiethnic Asian Knee Osteoarthritis Patients Scheduled for Total Knee Replacement. Arthritis & Rheumatism (Arthritis Care & Research) 2007;57:1043–1049 9/24/11
  • 13. A Gaussian mixture distribution? Defined by: µ1,µ2,σ1,σ2 and w 9/24/11
  • 14. Hospital A Hospital B 87% 13% 70% 30% ! Mean = 0.58 Mean = 0.58 SD = 0.21 SD = 0.21 9/24/11
  • 15. Studying change in EQ-5D It has been suggested that pairwise differences between pre- and postoperative EQ-5D values are normally distributed and can be meaningfully interpreted. 9/24/11
  • 16. Studying change in EQ-5D It has been suggested that pairwise differences between pre- and postoperative EQ-5D values are normally distributed and can be meaningfully interpreted. It can easily be shown that this is not correct The sum of two bimodal distribution has a distribution with three modes, the difference four. 9/24/11
  • 17. Empirical EQ-5D data from knee patients in Trelleborg 2007-2008 Preop EQ-5D Postop EQ-5D Delta EQ-5D Delta EQ-5D 9/24/11
  • 22. Additional problem with analyses of change Change is confounded by association with baseline X = pre-operative (baseline) value Y = postoperative (follow up) value Y-X correlates with X Solution When analyzing change, adjust for imbalance at baseline (This is an almost perfect case-mix adjustment!) 9/24/11
  • 23. 1. Stockholm 2. Kronoberg 2. Gävleborg 3. Östergötland 9/24/11
  • 24. 2. Gävleborg 1. Stockholm 3. Östergötland 2. Kronoberg 9/24/11
  • 26. EQ-5D Problems Conventional analyses Mean values not interpretable Confidence intervals not reliable (Calculated assuming Gaussian distribution) P-values not reliable (Student's t-test, ANOVA, etc. requires Gaussian distribution and homogeneous variance) 9/24/11
  • 27. EQ-5D Problems cont'd Non-parametric analysis Median value may not exist. Confidence intervals not reliable (calculated assuming Gaussian or binomial distribution). P-values not reliable (Wilcoxon's MPSR-test requires a symmetrical distribution, Mann-Whitney U-test requires distributions with identical shape. 9/24/11
  • 28. EQ-5D Problems cont'd Adjusting for baseline How meaningful is the outcome of an ANCOVA with variables having non-Gaussian, multimodal distributions (with different number of modes)? What do these residuals look like? 9/24/11
  • 29. EQ-5D Problems cont'd Alternative analyses methods? - Mixture distribution analysis (mixdist library for R) - Multi-state Markov analysis (msm library for R) 9/24/11
  • 34. “This is about clinical improvement, not science” 9/24/11
  • 35. Swedish law defines clinical improvement work (CIW) as “not research” Some CIW projects include experiments on patients - No ethics approval is required (or can be applied for) - No informed consent - No scientific planning or evaluation of the experiments - No formal publication of studies and results 9/24/11
  • 36. Regression analysis - Adjusting for baseline - Models only including statistically significant factors - Stepwise regression methods 9/24/11
  • 37. What factors should be included in a linear model (ANCOVA)? Y = b0 + b1X1 + b2X2 + … + bnXn + e This is a multiple or multivariable analysis but not multivariate. Xi is a variable (factor or covariate) bi is the effect on Y of one unit change in Xi Assume that Y is blood pressure and X1 an indicator of anti- hypertensive treatment. bi will then estimate the treatment effect in terms of blood pressure reduction. 9/24/11
  • 38. Linear models Answer It depends on a) the purpose of the study and b) the study design used. 1. Purpose: (black-box) prediction Any variable can be included as long as it increases the sensitivity and specificity of the prediction, and as long as results (bi) are not interpreted in terms of causal effects. 2. Purpose: effect estimation The variables needed to produce valid (bi and their s.e.) should be included. 9/24/11
  • 39. Linear models 1. Common for all designs Include baseline when analyzing change in a continuous variable. 2. Randomized trial Include randomization stratification factors (for valid standard errors). 3. Observational study Include potential confounding factors (for valid regression coefficients). 9/24/11
  • 40. Linear models How should confounding factors be included? 1. By the investigator's reasoning. 2. By reviewing other publications on the same endpoint. 3. By performing sensitivity analyses. 4. But not by using hypothesis testing or stepwise regression analysis. 9/24/11
  • 41. Parsons et al. A systematic survey of the quality of research reporting in general orthopaedic journals. J Bone Joint Surg Br 2011;93-B,1154-9 9/24/11