Ann Lazar

Transcription

1 Ann Lazar Department of Preventive and Restorative Dental Sciences Department of Epidemiology and Biostatistics University of California, San Francisco February 21, 2014

2 Motivation Rheumatoid Arthritis case-control High School and Beyond - survey CAN DO FV Phase 3 RCT The Johnson-Neyman (J-N) Examples Final Comment Lazar AA, PhD 2

3 Motivation: RA Case-Control Example Autoantibodies are present years prior to the onset of symptoms of rheumatoid arthritis (RA) and may be highly predictive of the development of RA Interest in determining when the autoantibodies are elevated prior to diagnosis of RA Lazar AA, PhD 3

4 Motivation: RA Case-Control Example Autoantibodies are present years prior to the onset of symptoms of rheumatoid arthritis (RA) and may be highly predictive of the development of RA Interest in determining when the autoantibodies are elevated prior to diagnosis of RA Lazar AA, PhD 4

5 RA Case Control Study SERA study retrospective case-control study of RA subjects years (4994 days) prior to diagnosis Study Sample 166 subjects 83 controls and 83 cases 486 measurements 1 to 4 measurements per subject Deane KD, O'Donnell CI, Hueber W, Majka DS, Lazar AA, Derber LA, Gilliland WR, Edison JD, Norris JM, Robinson WH, Holers VM. The number of elevated cytokines and chemokines in preclinical seropositive rheumatoid arthritis predicts time to diagnosis in an age-dependent manner. Arthritis Rheum Nov; 62(11): Lazar AA, PhD 5

6 RA Case Control Study Lazar AA, PhD 6

7 RA Case Control Example Lazar AA, PhD 7

8 High School and Beyond Survey Study the relationship between mathematics achievement score (MA) and socioeconomic status score (SES) between sectors: Catholic (n=70) and public (n= 90) high schools SES = a composite score of parental income, education, occupation and achievement A subset of the data from the High School and Beyond Survey (HSBS) of 7,185 students nested within 160 high schools, which averaged 45 students per school This study can address important questions: 1) What range of SES does the mathematics achievement scores statistically differ between private and public high schools 2) Which school sector does better, Catholic or Public High Schools? Lazar AA, PhD 8

9 High School and Beyond Survey Lazar AA, PhD 9

10 High School and Beyond Survey Lazar AA, PhD 10

11 CAN DO Fluoride Varnish (FV) We are interested in detecting and exploring patterns of heterogeneity of treatment effect (e.g., odds ratio). Identify people who respond differently to treatment. Identify who may benefit the most (or the least) from a treatment so treatment can be tailored to the individual We proceed by studying Differences (heterogeneity) between groups for varying values of a baseline covariate based on a model Lazar AA, PhD 11

12 CAN DO Fluoride Varnish (FV) Variables Primary Outcome (Dichotomous): Any Caries Incidence Treatment: FV + parental counseling vs. counseling alone Baseline Covariate (Continuous): Salivary Mutans Streptococci (MS) (CFU/ml)) Study Sample 249 caries-free children w/ baseline MS 0 FV: n=89; 1F/Yr : n=78 ; 2FV/Yr : n=82 Goal At what values of MS does FV treatment result in statistically different (heterogeneity) caries incidence? Who benefits (most) from treatment? Lazar AA, PhD 12

13 CAN DO Fluoride Varnish (FV) Logistic Regression Model of Caries Incidence v. No Caries Incidence with Associated Standard Error Lines Lazar AA, PhD 13

14 J-N Overview Recent Work Main Results Johnson-Neyman (J-N) Approach Lazar AA, PhD 14

15 J-N Overview Recent Work Main Results The J0hnson-Neyman Approach Aim: What range of the independent variable is the difference in the outcome variable statistically significant among groups ( Significance Region )? Johnson-Neyman (J-N) (1936) in ANCOVA when the regression lines are not parallel Solves the problem of identifying regions of significance or Significance Region Tests whether the difference in means for a particular value of the covariate between two groups is statistically significance comparing computed F statistics to the critical value of Fisher s F distribution Assumes normality and homogeneity of variance Lazar AA, PhD 15

16 J-N Overview Recent Work Main Results Continuous Outcome Continuous Covariate Lazar AA, PhD 16

17 xtensions to J-N J-N Overview Recent Work Main Results Improvements to J-N Approach Huitema (1980) explicit solution, a simple formula that involves solving a quadratic equation, available for comparing several groups but only when the number of covariates is one For two or more covariates, the explicit solution is thought to be intractable Hunka and Leighton (1997) developed a solution for any number of possible covariates by casting the equation within a general linear model framework. But this approach cannot be solved directly, symbolic processing capabilities of computational software (e.g., Mathematica_ Lazar AA, PhD 17

18 xtensions to J-N J-N Overview Recent Work Main Results Improvements to J-N Approach Hunka and Leighton (1997) 3 variables of interest Two Parabloids Lazar AA, PhD 18

19 J-N Overview Recent Work Main Results Improvements to J-N Approach Miyazaki and Maier (M-M) (2005) developed a procedure for determining the significance regions for correlated Gaussian distributed data involves fitting hierarchical linear mixed models (HLM) solved for significance region using the symbolic processing capabilities (e.g., Mathematica) Above solutions are available for comparing groups Lazar AA, PhD 19

20 Recent Extension in the Spirit of J-N Main Results Improvements to J-N Approach Significance regions (solutions) for both correlated Gaussian and non-gaussian distributed data suitable for generalized linear (mixed) models (GL(M)M), which includes HLMs. The proposed solution can : Compare subjects and groups Adjust for covariates Account for Multiple Testing J-N Overview Recent Work Makes use of one software package to implement the model and solution, as follows Lazar AA, PhD 20

21 GLMM J-N Overview Recent Work Main Results The GLMM Framework The conditional mean of Y depends on the fixed (β) and random effects (u) via the linear predictor η = Xβ + Zu, where g E Y u = η = Xβ + Zu H 0 : θ sx1 = C sx(p+q) β px1 u qx1 = C 1 β + C 2 u = 0 H 0 : θ = 0 can be tested with a t- or F-statistic. Let V θ = CLC and L is the empirical covariance matrix of β β and u u; s=the rank of C, the contrast matrix t θ = θ V θ and F θ = θ 1xs V θ sxs 1 θ sx1 s Lazar AA, PhD 21

22 Significance Region: Simultaneous Multiple Testing Solutions J-N Overview Recent Work Main Results Let h identify subsets of linear function of β u H 0 : hθ = 0 rejected (hθ 0) t hθ > sf 1 α,s,v Or we can use our explicit solution to identify the significance region h x θθ sf 1,s,v V θ h x > 0 SAS macro available to determine the significance regions Lazar AA, PhD 22

23 J-N Overview The Significance Region (Explicit Solution) Recent Work Main Results Suppose θ C represent the difference in intercepts; θ A denote the difference in slopes, then θ = θ C θ A Let V θ(c), V θ(a), and Cov θ(a)θ(c) be the empirical prediction error variance (V) and covariance (Cov) of θ C and θ A 2 A V Cov Cov C C A C C A 1 X sf C 2 A 1,s,v V A C A 1 X 0 Lazar AA, PhD 23

24 J-N Overview Recent Work Main Results Let A = θ A 2 sf 1 α,s,v V θ A, B = 2 θ A θ C sf 1 α,s,v Cov θ A θ C, C = θ C 2 sf 1 α,s,v V θ C, and the discriminant, D = B 2 4AC Let X 1 B 2A D and X 2 B 2A D Case I (A>0 implies D>0) : X< X 1 or X > X 2 Case II (A<0 and D>0) : X 2 < X< X 1 Case III (D < 0 and A < 0): no region A proof of these 3 cases has been derived Lazar AA, PhD 24

25 Examples-RA Case Control Study Lazar AA, PhD 25

26 Examples- RA Case Control Example Lazar AA, PhD 26

27 Step 1: Determine ( x ) Lazar AA, PhD 27

28 Step 2: Re-write ( x ) Lazar AA, PhD 28

29 Step 3: Determine the set of x s for which the times response curves differ Lazar AA, PhD 29

30 Lazar AA, PhD 30

31 Lazar AA, PhD 31

32 Lazar AA, PhD 32

33 Lazar AA, PhD 33

34 High School and Beyond Survey 1) What range of SES does mathematics achievement scores statistically differ between private and public high schools 2) Which school sector does better, Catholic or Public High Schools? Lazar AA, PhD 34

35 High School and Beyond Survey What range of SES does the mathematics achievement scores statistically differ between the Catholic high school sector and a particular public high school (for example, school 1296) Does the Catholic high school sector have higher math scores than a particular public high school? Lazar AA, PhD 35

36 Fit generalized linear mixed models (GLMM) of mathematics achievement score, y ij, for the ith student in the jth cluster (school) y ij = β 0j + β 1j SES ij SES j + e ij β 0j = γ 00 +γ 01 SES j +γ 02 sector j + u 0j β 1j = γ 10 +γ 11 SES j +γ 12 sector j + u 1j SES j denotes the school-averaged student SES sector equals 1 for Catholic (C) schools or 0 for public (P) schools e ij represents the unobservable random vector of normally distributed errors e ij were assumed uncorrelated with the multivariate normally distributed random effects,u ij, of the school mean,(u 0j ), and SES-achievement slope (u 1j ). Lazar AA, PhD 36

37 E C Y ij E P Y ij = γ 02 + γ 12 SES ij SES j, where 1 x h 1 0 C 0 1 γ 02 γ 12 β = 1 x γ 02 γ 12 = hθ, Results in (over) 3839 null hypotheses of no difference in mathematical achievement scores between Catholic and public schools for each of the 3839 distinct values of relative SES, SES ij SES j, or X. Lazar AA, PhD 37

38 A = γ F 0.95,2,7022 V γ12 = = 2.353; B = 2 γ 02 γ 12 2F 0.95,2,7022 V γ02 γ 12 = = ; C = γ F 0.95,2,7022 V γ02 = = ; D = B 2 4AC = = Lazar AA, PhD 38

39 After plugging in A, B, C and D for Case I; x < or x > Catholic better > 0 Public Better <0 Lazar, A.A. and Zerbe, GO (2011). Solutions for Determining the Significance Region Using the Johnson-Neyman Type Lazar AA, Procedure PhD in Generalized Linear (Mixed) Models. Journal of Education and Behavioral Statistics 36(6):

40 Example2: High School and Beyond High School and Beyond Survey (U.S. High School Students) Significance Region: Relative SES : and 0.737, students from the Catholic high school sector (n=70) have greater odds of higher mathematics achievement scores than the public school 1296 What range of SES does the mathematics achievement scores statistically differ between the Catholic high school sector (n=70) and a particular public high school (n=1)(for example, school 1296) and Figure: SAS macro output OR Lazar, A.A. and Zerbe, GO (2011). Solutions for Determining the Significance Region Using the Johnson- Lazar AA, PhD 40 Neyman Type Procedure in Generalized Linear (Mixed) Models. Journal of Education and Behavioral Statistics 36(6):

41 Efficacy of Different Fluoride Varnish Application Frequencies in Preventing Early Childhood Caries Incidence Log Odds No FV FV *MS(CFU/mL) *MS (CFU/mL) Log 10 Salivary Mutans Streptococci (CFU/ml) Lazar AA, PhD 41

42 Efficacy of Different Fluoride Varnish Application Frequencies in Preventing Early Childhood Caries Incidence Odds Ratio N Caries Events Any Intended Fluoride No Fluoride <MS (CFU/ml) < Log 10 Salivary Mutans Streptococci (CFU/ml) Lazar, A.A., Gansky, S.A., Halstead, D.D., Slajs, A., Weintraub, J.A. Improving patient care using the Johnson-Neyman analysis of heterogeneity of treatment effects according to individuals baseline characteristics. Journal of Dental, Oral and Craniofacial Epidemiology, accepted for publication. Lazar AA, PhD 42

43 CAN DO FV - children with higher baseline MS values who were randomized to receive fluoride varnish had the poorest dental caries prognosis and may have benefitted most from the preventive agent Lazar AA, PhD 43

44 Final Comment J-N may be an important tool in the field of personalized health care While J-N explicit solution assumes that the covariate has linear effects, the grid solution does not require linearity in the covariates J-N can make more detailed and substantial descriptions about the significance region Lazar AA, PhD 44

45 References for Johnson-Neyman J-N References Johnson, P. O., and Neyman, J. (1936). Tests of certain linear hypothesis and their application to some educational problems. Statistical Research Memoirs 1, Hunka, S., and Leighton, J. (1997). Defining Johnson-Neyman regions of significance in the three-covariate ANCOVA using Mathematica. Journal of Education and Behavioral Statistics 22, Raudenbush, S. W., and Byrk, A. S. (2002). HLMS--Hierarchical linear models: Applications and data analysis methods (2nd ed.). Newbury Park, CA: Sage Miyazaki, Y., and Maier, K. (2005). Johnson-Neyman Type Technique in Hierarchical Linear Models. Journal of Education and Behavioral Statistics 30, Lazar, A.A. and Zerbe, GO (2011). Solutions for Determining the Significance Region Using the Johnson-Neyman Type Procedure in Generalized Linear (Mixed) Models. Journal of Education and Behavioral Statistics 36(6): Lazar, A.A., Gansky, S.A., Halstead, D.D., Slajs, A., Weintraub, J.A. (in-press, 2014) Improving patient care using the Johnson-Neyman analysis of heterogeneity of treatment effects according to individuals baseline characteristics. Journal of Dental, Oral and Craniofacial Epidemiology, accepted for publication. Weintraub JA, Ramos-Gomez F, Jue B, Shain S, Hoover CI, Featherstone JD, Gansky SA. (2006) (Original CAN DO FV report) Deane KD, O'Donnell CI, Hueber W, Majka DS, Lazar AA, Derber LA, Gilliland WR, Edison JD, Norris JM, Robinson WH, Holers VM. The number of elevated cytokines and chemokines in preclinical seropositive rheumatoid arthritis predicts time to diagnosis in an age-dependent manner. Arthritis Rheum Nov; 62(11): Lazar AA, PhD 45

46 Acknowledgements Acknowledgements Patients and participants of the RA study, High School and Beyond, CAN DO FV The Center to Address Disparities in Children's Oral Health (CAN DO) NIDCR P30 DE & NCI T32 CA (AAL) NIH NIDCR P60DE013058, U54DE014251, U54DE (CAN DO) Thank you for your attention. Lazar AA, PhD 46