EDUCATIONAL MATERIAL

Transcription

1 ERS Annual Congress Amsterdam September 2015 EDUCATIONAL MATERIAL Meet the expert 11 Research methodology and data analysis Thank you for viewing this document. We would like to remind you that this material is the property of the author. It is provided to you by the ERS for your personal use only, as submitted by the author by the author Tuesday, 29 September :00 14:00 Room G111 RAI

2 You can access an electronic copy of these educational materials here: To access the educational materials on your tablet or smartphone please find below a list of apps to access, annotate, store and share pdf documents. iphone / ipad Adobe Reader - FREE With the Adobe Reader app you can highlight, strikethrough, underline, draw (freehand), comment (sticky notes) and add text to pdf documents using the typewriter tool. It can also be used to fill out forms and electronically sign documents. UPAD Lite - FREE UPAD Lite is an advanced note-taking application with annotation features. You can handwrite notes, highlight text, add sticky notes and reference images and export any type of document as a PDF or PNG file by or to cloud services. Noteability - $4.99 Noteability uses CloudServices to import and automatically backup your PDF files and allows you to annotate and organise them (incl. special features such as adding a video file). On ipad, you can bookmark pages of a note, filter a PDF by annotated pages, or search your note for a keyword. Android Adobe Reader - FREE The Android version of Adobe Reader lets you view, annotate, comment, fill out, electronically sign and share documents. It has all of the same features as the ios app like freehand drawing, highlighting, underlining, etc. iannotate PDF - FREE You can open multiple PDFs using tabs, highlight the text and make comments via handwriting or typewriter tools. iannotate PDF also supports Box OneCloud, which allows you to import and export files directly from/to Box. ez PDF Reader - $3.99 With the ez PDF reader you can add text in text boxes and sticky notes; highlight, underline, or strikethrough texts or add freehand drawings. Add memo & append images, change colour / thickness, resize and move them around as you like.

3 Regular physical activity improves quality of life and fitness in healthy individuals and people living with lung conditions. In partnership with: Farbe/colour: PANTONE 288 CV ERS/ELF ER Journal 210x280 ad_v2 AW.indd 2 29/07/ :22

4 Research Methodology and Data Analysis Confounding, stratification and interaction Prof. Dr H. Marike Boezen Dept of Epidemiology UMCG & University of Groningen Hanzeplein RB Groningen NETHERLANDS h.m.boezen@umcg.nl Dr Judith M. Vonk Dept. of Epidemiology UMCG & University of Groningen Hanzeplein RB Groningen NETHERLANDS j.m.vonk@umcg.nl AIMS: To review the different methodological procedures in scientific studies and to explain the statistical procedures that can be used to analyse the data of scientific studies. TARGET AUDIENCE: Pulmonologists, thoracic surgeons, nurses, paediatricians, and clinical researchers. AIMS Identify confounding Test differential associations between groups SUMMARY Many studies aim to identify associations between a specific exposure (e.g. smoking or treatment) and a specific outcome (e.g. FEV 1). However, the problem is that this association can be biased due to other variables that are associated with both the exposure and the outcome. These variables are called confounders. A clear example of a confounder in the association between smoking and FEV 1 is sex since males are more often smokers and males have a higher FEV 1 than females. So, if the variable sex is not taken into account ( adjusted for ) in the analysis the result will not be correct. Another problem in studying the association between smoking and FEV 1 is that this association may be different between males and females. This is called effect modification or interaction. The specific associations for males and females separately can be analyzed in a stratified analysis. However, this does not show if there is a statistically significant difference in the association between smoking and FEV 1 between males and females. To identify such a difference an interaction term should be incorporated in an unstratified model and the significance of the interaction term should be determined. In the current presentation the concept of confounding and effect modification are explained and methods to identify confounding and effect modification are presented. Also the use of stratification is discussed. Pros and Cons of the different approaches are presented and discussed. With some clear examples the interpretation of the regression coefficients are explained. Given the targeted audience, the presentation concerns basic (statistical) methods to identify and deal with confounding and interactions. Participants are encouraged to be actively involved and pose questions, also about their own research. 4

5 REFERENCES There is a large amount of literature and educational material on websites, with different terminology and focus. Here are some suggestions of introductory papers as well as more advanced (note: these papers cover a wide range of aspects; not all will be discussed during the meeting). 1. Chinn S. Statistics for the European Respiratory Journal. Eur Respir J Aug; 18(2): A paper on basic statistics and how to present this in a paper. 2. Kamangar F. Confounding variables in epidemiologic studies: basics and beyond. Arch Iran Med Aug; 15(8): Extensive review on confounding and interaction and related subjects such as mediation. 3. Kamangar F. Effect modification in epidemiology and medicine. Arch Iran Med Sep; 15(9): Extensive review on confounding and interaction and related subjects such as mediation. 4. Knol MJ, VanderWeele TJ. Recommendations for presenting analyses of effect modification and interaction. Int J Epidemiol Apr;41(2): Advanced paper on additive and multiplicative interactions and how to present this in a paper 5. Knol MJ, van der Tweel I, Grobbee DE, Numans ME, Geerlings MI. Estimating interaction on an additive scale between continuous determinants in a logistic regression model. Int J Epidemiol Oct; 36(5): Advanced paper on additive and multiplicative interactions and how to present this in a paper 6. Larose TL, Brumpton BM, Langhammer A, Camargo CA Jr, Chen Y, Romundstad P, Mai XM. Serum 25-hydroxyvitamin D level, smoking and lung function in adults: the HUNT Study. Eur Respir J Aug; 46(2): A paper from the August issue of the European Respiratory Journal. This paper is an example of adjusting for confounding and stratified analysis. Interactions are not analyzed in this paper. 7. de Jong K, Boezen HM, Kromhout H, Vermeulen R, Postma DS, Vonk JM; LifeLines Cohort study. Pesticides and other occupational exposures are associated with airway obstruction: the LifeLines cohort study. Occup Environ Med Feb; 71(2): A paper from our own group in which both confounding, stratification and interaction are analyzed and presented. 8. de Jong K, Boezen HM, Kromhout H, Vermeulen R, Postma DS, Vonk JM. Association of occupational pesticide exposure with accelerated longitudinal decline in lung function. Am J Epidemiol Jun 1; 179(11): A paper from our own group in which both confounding, stratification and interaction are analyzed and presented. EVALUATION 1. If a specific variable is associated with the exposure but not with the outcome then this variable is not a confounder. a. True b. False 2. If a specific variable is associated with the exposure but not with the outcome then this variable cannot be an effect modifier a. True b. False 3. Including a statistical interaction term between two variables in a linear regression model can only be done if: a. both variables are categorical (0/1) variables b. one variable is a categorical variable (0/1) and the other is a continuous variable c. both variables are continuous variables d. all combinations of variable types are possible 5

6 4. Stratified analysis shows if the association is significantly different between the stratification groups. a. True b. False 6

7 7

8 RESEARCH METHODOLOGY AND DATA ANALYSIS Confounding, stratification and interaction MARIKE BOEZEN & JUDITH VONK Dept. of Epidemiology UMCG & University of Groningen Netherlands 8

9 I have no real or perceived conflicts of interest that relate to this presentation: Affiliation/Financial Interest Grants/research support: Commercial Company Honoraria or consultation fees: Participation in a company sponsored bureau: Stock shareholder: Spouse/partner: Other support/potential conflict of interest: This event is accredited for CME credits by EBAP and EACCME and speakers are required to disclose their potential conflict of interest. The intent of this disclosure is not to prevent a speaker with a conflict of interest (any significant financial relationship a speaker has with manufacturers or providers of any commercial products or services relevant to the talk) from making a presentation, but rather to provide listeners with information on which they can make their own judgments. It remains for audience members to determine whether the speaker s interests, or relationships may influence the presentation. The ERS does not view the existence of these interests or commitments as necessarily implying bias or decreasing the value of the speaker s presentation. Drug or device advertisement is forbidden. 9

10 Introduction AIMS Aim 1 Identify confounding Aim 2 Test differential associations between groups 10

11 Sources of bias Sources of bias in epidemiological research: Selection bias Unrepresentative nature of sample Information (misclassification) bias Errors in measurement of exposure of disease Confounding bias Distortion of exposure - disease relation by some other factor Types of bias not mutually exclusive 11

12 Confounding A confusion of effects The apparent association of the exposure of interest on the outcome is distorted because the association of an extraneous factor is mistaken for or mixed with the actual exposure effect The distortion can be large and lead to overestimation or underestimation of the association (or no association) between the exposure and the outcome, or change the apparent direction of the association 12

13 Confounding Confounder A third factor which is related to both exposure and outcome, and which accounts for some/all of the observed relationship between the two 13

14 Confounder Exposure Outcome Potential confounder Associated with exposure - without being the consequence of exposure Associated with outcome - independently of exposure (not an intermediary) 14

15 Confounding Mean FEV 1 : Never smoker Ex-smoker Current smoker 291 cl 325 cl 319 cl FEV 1 (cl) = *ex + 28*current Vlagtwedde/Vlaardingen cohort study; survey 1976/78, n=

16 Confounding Numbers of smokers according to sex FEV 1 according to sex 16

17 Confounder Smoking FEV 1 Sex Associated with exposure - without being the consequence of exposure Associated with outcome - independently of exposure (not an intermediary) 17

18 Confounding FEV 1 (cl) = *ex + 28*current Is sex a confounder of the association between smoking and FEV 1? FEV 1 (cl) = 270-3*ex - 5*current + 84*sex Yes, the association of smoking and FEV 1 changes substantially after adjustment for sex 18

19 Confounding vs interaction Is the association between X and Y influenced by another variable Z? yes confounding Is the association between X and Y different for different values of Z? yes interaction 19

20 Interaction Is the association between age and FEV 1 different between males and females? LEEFTIJD age (yrs) males man females vrouw Vlagtwedde/Vlaardingen cohort study; survey 1989/90, n=799 20

21 Stratification and interaction Research Question: Is the association between an exposure and a specific outcome different between groups 21

22 Stratification and interaction Research Question: Is the association between an exposure and a specific outcome statistically significant different between groups 22

23 Stratification and interaction Research Question: Is the association between smoking and FEV 1 statistically significant different between males and females 23

24 Stratification and interaction Vlagtwedde/Vlaardingen cohort study; survey 1989/90, n=

25 Stratification and interaction Stratification: Perform the analysis in the two groups separately 25

26 Stratification and interaction Stratification: Perform the analysis in the two groups separately Results linear regression: Model: FEV 1 %PP = intercept + β1 * ex smoker + β2 * current smoker B SE P Female intercept <.01 ex smoker current smoker <.01 Male intercept <.01 ex smoker <.01 current smoker <.01 26

27 Stratification and interaction 27

28 Stratification and interaction Stratification: Perform the analysis in the two groups separately Results linear regression (ex smoker as reference): Model: FEV 1 %PP = intercept + β1 * never smoker + β2 * current smoker B SE P Female intercept <.01 never smoker current smoker <.01 Male intercept <.01 never smoker <.01 current smoker <.01 28

29 Stratification and interaction 29

30 Stratification and interaction Stratification: Perform the analysis in the two groups separately So now we know in which group there is a significant association with smoking BUT. We still do not know if there is a statistically significant difference between the groups 30

31 Stratification and interaction Interaction: Perform the analysis in the total population and add an interaction term between smoking and male gender in the regression model 31

32 Stratification and interaction Interaction: add interaction between smoking and gender Model: FEV 1 %PP = intercept + β1 * male + β2 * ex smoker + β3 * current smoker + β4 * ex smoker * male + β5 * current smoker * male B SE P intercept <.01 male ex smoker current smoker <.01 ex smoker * male <.01 current smoker * male <.01 32

33 Stratification and interaction Interaction: add interaction between smoking and gender (ex smoker as reference) Model: FEV 1 %PP = intercept + β1 * male + β2 * never smoker + β3 * ex smoker + β4 * never smoker * male + β5 * ex smoker * male B SE P intercept <.01 male <.01 never smoker <.01 current smoker <.01 never smoker * male <.01 current smoker * male

34 Stratification and interaction Stratified: Female ex smoker B=0.3 Male ex smoker B=-5.7 Interaction: ex smoker ex smoker * male B=0.3 B=

35 Stratification and interaction Stratified: Female ex smoker B=0.3 Male ex smoker B=-5.7 Interaction: ex smoker ex smoker * male B=0.3 B=

36 Stratification and interaction Stratified: Female current smoker B=-4.6 Male current smoker B=-11.3 Interaction: current smoker current smoker * male B=-4.6 B=

37 Stratification and interaction Stratified: Female current smoker B=-4.6 Male current smoker B=-11.3 Interaction: current smoker current smoker * male B=-4.6 B=

38 Stratification and interaction Stratified: Female current (vs ex) smoker B=-4.9 Male current (vs ex) smoker B=-5.6 Interaction: current (vs ex) smoker current (vs ex) smoker * male B=-4.9 B=

39 Stratification and interaction Stratified: Female current (vs ex) smoker B=-4.9 Male current (vs ex) smoker B=-5.6 Interaction: current (vs ex) smoker current (vs ex) smoker * male B=-4.9 B=

40 Stratification and interaction Interaction: add interaction between smoking and gender Caution: The effect of the interacting variable changes if you take another reference B SE P intercept <.01 male ex smoker current smoker <.01 ex * male <.01 current * male <.01 B SE P intercept <.01 male <.01 never smoker <.01 current smoker <.01 never * male <.01 current * male

41 Summary Confounding: control for a variable that is associated with both the exposure and the outcome Stratification: associations within a group Interaction: differences in associations between groups 41

42 Faculty disclosures There are no faculty disclosures for this session. 42

43 Answers to evaluation questions Please find all correct answers in bold below Research Methodology and Data Analysis - Confounding, stratification and interaction Prof. Dr H. Marike Boezen, Dr Judith M. Vonk 1. If a specific variable is associated with the exposure but not with the outcome then this variable is not a confounder. a. True b. False 2. If a specific variable is associated with the exposure but not with the outcome then this variable cannot be an effect modifier a. True b. False 3. Including a statistical interaction term between two variables in a linear regression model can only be done if: a. both variables are categorical (0/1) variables b. one variable is a categorical variable (0/1) and the other is a continuous variable c. both variables are continuous variables d. all combinations of variable types are possible 4. Stratified analysis shows if the association is significantly different between the stratification groups. a. True b. False