META-ANALYSIS OF FORMALDEHYDE AND LEUKEMIA

Craig Steinmaus, MD, MPH School of Public Health University of California, Berkeley Formaldehyde Science Conference, Madrid, April 19-20 2012

Schwilk et al., JOEM 2010 52(9):878-886

Discussion Points Basic Principles of Meta-analysis Goals of Our Meta-analysis Methods Overall Results Evaluation of Causal Inference

META-ANALYSIS
A quantitative approach for systematically assessing the results of previous research in order to arrive at conclusions about the body of research. Pettiti

RRstudy 1 RRstudy 2 RRstudy 3 RRstudy 4 RRstudy 5

RRSummary
*weighted by precision

3 MAIN GOALS
Summarize the current epidemiologic literature on formaldehyde exposure and leukemia Identify sources of heterogeneity in the literature (if present) Evaluate the major tenets of causal inference (Bradford Hill Criteria*): Magnitude of the Association Finding Due to Chance? Consistency (e.g. Heterogeneity) Bias (e.g. exposure or outcome misclassification) Confounding Dose-response
*Hill, Proc R Soc Med 1965, 58 295-300

MAJOR STEPS
1. Define research question 2. Formal inclusion and exclusion criteria (a priori) 3. Literature search: PubMed, Embase, Biosis, Google Scholar, Cochrane, Toxnet, dissertations, bibliographies, meeting abstracts 4. Evaluate the studies and select ones most likely to be valid 5. Calculate a weighted average summary relative risk: weighted by precision (i.e. study size) 7. Evaluate causal inference

INCLUSION EXCLUSION CRITERIA

Epidemiologic studies on formaldehyde exposure and leukemia Case-control and cohort studies Peer reviewed scientific literature or edited books (Government reports evaluated in sensitivity analyses) Relative risks and estimates of variance (e.g. 95% confidence intervals) provided Independent studies: high exposure group, largest number of cases, most recent Exposure assessed similarly in cases and controls (e.g. Partanen et al. excluded)

METHODS: Myeloid leukemia RRs used if available

Based on prior results: Zhang et al. 2009 Mutat Res 681:150-168 Wanted to avoid including unrelated subtypes: bias to the null Impact of adding unassociated outcomes: example of silicosis and cancer:
Meta-analysis of silicosis and lung cancer
Study Study 1 Study 2 Study 3 RR 1.6 2.1 1.7 CIlow 0.7 1.1 1.1 CIup Cancer type

Add an unrelated cancer (prostate)

Study Study 1 Study 2 Study 3 Study 4 Study 5 Study 6 Study 7 Study 8 Study 9 RR 1.6 2.1 1.7 1.0 0.9 1.1 0.9 1.0 0.8 CIlow 0.7 1.1 1.1 0.7 0.2 0.3 0.1 0.6 0.5 CIup 3.1 3.5 2.5 1.4 2.0 2.2 3.0 1.4 1.1 Cancer type Lung cancer Lung cancer Lung cancer Prostate cancer Prostate cancer Prostate cancer Prostate cancer Prostate cancer Prostate cancer

3.1 Lung cancer 3.5 Lung cancer 2.5 Lung cancer

RRs = 1.78 (95% CI, 1.32 2.42)

RRs =1.15 (95% CI, 0.97 1.37)

Methods: High exposures selected over all exposed

All exposed (high and low exposed people)

Other meta-analyses Only highly exposed people

Beane-Freeman 2009 (JNCI, 101:751-761)

Our meta-analyses

Why focus on high exposures?

RELATIONSHIP BETWEEN POWER AND RELATIVE RISK

Relative risk has greater impacts on power than sample size

Methods: When different exposure metrics were given RRs were selected in the following order
1. Highest Peak Exposure 2. Highest Average Exposure Intensity 3. Longest Exposure Duration 4. Work Prior to 1970s 5. Cumulative Exposure Average intensity, cumulative exposure and duration of exposure are all susceptible to inaccuracy due to periods of high intermittent exposure - Bachand et al. Formaldehyde Council Inc. For other carcinogens (arsenic, smoking) average exposure intensity is a stronger determinant of risk than cumulative exposure or exposure duration (Lubin et al., 2008 EHP 116:1661-5)

HETEROGENEITY It is not evil It is to be expected in observational epidemiology If bias is introduced, what is the direction of the bias?

Effects of adding poor or unrelated exposure metrics

RRs = 1.78
(95% CI, 1.32 2.42)

Add studies using poor exposure metrics poorly or unrelated to leukemia

RRs =1.15
(95% CI, 0.97 1.37)
bad = non-differential misclassification of exposure

Methods: Fixed Effects Model

Weighting is based on study precision (i.e. variance or size)

Wi = 1 / variancei variancei = (log CIup - log CIlow / 3.92)2 Wi = 1 / (log CIup - log CIlow / 3.92)2

Adjusted for heterogeneity using: Shore et al., Br J Ind Med 1993, 50, 971-97

We are essentially weighting on study size

Larger study

Narrower confidence interval

Smaller variance Variancei = ((ln CIup - ln CIlow) / 3.92)2

Larger the weight Wi = 1/Variancei2

Random Effects Model Not Used

1 Wi* = [ ((1/si2)(bi ( (bi /si2) / (1/si2)))2) (N 1)] x (1/si2) si2 + 2 2 22 ((1/si )) ((1/si ) )

s = standard error; b = logRR; N = number of studies

STUDIES USED

Notes: 13 cohort/1 ncc; chemical plants/other industries/anatomists/funeral directors; 4 myeloid, 10 all leukemia; some high/some all exposed; 9/14 RRs above 1.0

Overall RR = 1.53 95% CI, 1.11 2.11 p = 0.005

Andjekovich 1995: 2 cases < 4% total weight Harrington 1975 (pathologists): 1 case Harrington 1975 (lab techs): 1 case Stellman: ACS cohort, population based (?low exposures) Coggon 2003: ***MAJOR SOURCE OF HETEROGENEITY****

SUBGROUP ANALYSES: BIAS AND OTHER ISSUES

Myeloid leukemia only: assessing bias from including unassociated subtypes

N = 14

N=4

N = 16

DOSE-RESPONSE: Decreased RRs from including lower exposure workers

Definition of high exposure: Peak exposure > 2 ppm (STEL)* Average exposure > 0.75 ppm (PEL)* Average duration 10 years Work prior to 1961

DOSE-RESPONSE INFORMATION
3.5 3.0 2.5

p = 0.018
Exposure category
Low

Relative risk

2.0 1.5 1.0 0.5 0.0 Medium High Highest

Study
Andjelkovich et al, 1995 1 Andjelkovich et al, 1995 2 Andjelkovich et al, 1990 2 Andjelkovich et al, 1990 3 Band et al, 1997 4 Beane Freeman et al, 2009 Beane Freeman et al, Beane Freeman et al, 5 Hauptmann et al, 2003 5 Hauptmann et al, 2003 6 Blair et al, 1986 6 Blair et al, 1986 7 Bertazzi et al, 1986 8 Coggon et al, 2003 8 Coggon et al, 2003 9 Acheson et al, 1984 10 Gardner et al, 1993 10 Gardner et al, 1993 11 Dell and Teta, 1995 11 Dell and Teta, 1995 12 Edling et al, 1987 (mortality) Edling et al, 1987 13 Hall et al, 1991 Harrington & Oakes, 15 Harrington & Shannon, 1975 15 Harrington & Shannon, 1975 16 Levine et al, 1984 17 Pinkerton et al, 2004 17 Pinkerton et al, 2004 18 Stayner et al, 1988 19 Stern et al, 1987 19 Stern et al, 1987 20 Stone et al, 2004 21 Stroup et al, 1986 22 Wong, 1983 22 Wong, 1983 22 Wong, 1983
1

Cohort
Iron foundries: Iron foundries: Iron foundries: Iron foundries: Pulp and paper NCI: 42 year NCI: 42 year NCI: 42 year NCI: 35 year NCI: 35 year NCI: 23 years NCI: 23 years Resin Chemical Chemical Chemical Chemical Chemical Plastics Plastics Abrasive Abrasive Pathologists: Pathologists: Pathologists Lab technicians Undertakers Garment Garment Garment Tannery Tannery Fiberglass Anatomists Chemical Chemical Chemical

All cause SMR All cancer SMR (95% CI) (95% CI)
0.93 (0.86-1.01) 0.91 (0.82-1.00) 0.95 (0.89-1.02) 1.01 (0.94-1.08) 0.92 d -1.02 (0.99-1.03) 0.90 (0.85-0.94) 0.95 (0.93-0.97) 0.77 (0.72-0.83) 0.95 0.97 0.92 (0.76-1.11) 1.15 (1.10-1.20) 1.04 (1.02-1.07) 0.87 (0.83-0.91) 1.03 (0.99-1.07) 0.95 (0.86-1.04) 0.95 (0.90-1.00) 0.70 (0.63-0.76) 0.99 (0.8-1.2) -0.44 (0.38-0.51) 0.6 0.6 0.67 0.99 0.85 0.92 (0.88-0.96) 0.74 (0.69-0.79) na 0.89 (0.85-0.94) 0.77 (0.72-0.82) 0.65 (0.60-0.70) 0.73 (0.61-0.87) 0.86 (0.50-1.37) 0.7 0.99 (0.82-1.17) 0.97 (0.79-1.19) 0.98 (0.84-1.14) 1.16 (0.99-1.34) 1.03 d -1.07 (1.03-1.11) 0.93 (0.84-1.03) 0.90 (0.86-0.94) 0.65 (0.56-0.75) 0.98 0.96 1.06 (0.76-1.43) 1.31 (1.21-1.42) 1.10 (1.04-1.16) 0.97 (0.88-1.07) 1.14 (1.06-1.22) 0.97 (0.81-1.15) 1.02 (0.92-1.14) 0.88 (0.73-1.05) 0.93 (0.5-1.5) 0.84 (0.5-1.3) 0.46 (0.35-0.60) 0.68 0.6 0.62 0.87 0.81 0.89 (0.82-0.97) 0.82 (0.73-0.93) 0.75 (0.59-0.94) 0.79 (0.70-0.89) 0.81 (0.71-0.91) 0.64 (0.53-0.76) 1.01 (0.70-1.41) 1.09 (0.22-3.18) 0.47

Healthy Worker Effect

AVERAGE SMRs

0.86

0.89

Adjusting for the Healthy Worker Effect

Industrial vs. Professional Workers

RR 95% CI X2/df P-hetero.

2.27 1.15-4.45 1.42 0.21

1.38 0.96-1.99

1.45 0.95-2.22

1.85 1.20-2.86 0.12 0.95

heterogeneity 1.61 1.26 0.13 0.28

COGGON ET AL.
14,014 men Employed six British factories after 1937 where formaldehyde produced or used (almost all formaldehyde production) Follow-up for mortality to 2000 All leukemia (ICD 204-208) No myeloid data High exposure group (estimated time weighted average > 2 ppm) (No IH data before 1970 so estimated from irritation symptoms and more recent IH measurements). Exposure estimates, no QC data, no data on peak exposure Observed cases = 8 Expected cases = 11.3 SMR = 0.71 (95% CI, 0.31 1.39) Chance?

In order to cause important confounding a factor must meet each of the following criteria
1. Associated with the exposure 2. Must be a risk factor for the disease 3. Can not be in the causal chain between the exposure and the disease 4. Must be fairly strongly associated with the exposure and the outcome to cause important confounding (Axelson, 1978, Scan J Work Environ Health, 4:85-89)

Unadjusted

Unadjusted

Unadjusted

Axelson, 1978, Scan J Work Environ Health, 4:85-89

Possible confounding factors Benzene Radiation Viruses Others Unknown mystery confounder What is the likelihood these are strongly enough associated with formaldehyde exposure and leukemia to cause a RR 0f 2.47?

Increased diagnosis in professional workers?

PUBLICATION BIAS

log RR In publication bias, usually smaller studies with null or negative findings tend not to be published.
Larger studies

Standard error of logRR

Smaller studies

Eggers p = 0.97; Beggs p = 0.96

Current Meta-analysis
RR = 1.53 (1.11-2.11)
Author, year Andjelkovich Hall et al, 199169 Harrington & Shannon Harrington & Shannon Levine et al, 198470 Stellman et al, 199852 Beane Freeman et al Coggon et al, 200349 Dell & Teta, 199547 Hauptmann et al Pinkerton et al, 200471 Stern et al, 198772 Stroup et al, 198648 Wong, 198373 Blair et al, 200125d Partanen et al 199326 Marsh et al, 200417 Matanoski, 199112 Robinson et al, 198713 Cohort Iron foundry Pathologists Pathologists Lab technicians Undertakers US population Ten US industries Chemical plants Plastics workers Embalmers Garment workers Tannery workers Anatomists Chemical workers Iowa & Minnesota Wood industry Ten US industries Pathologists Plywood mill Exposed group Exposed Total cohort Total cohort Total cohort Total cohort Exposed Peak 4+ ppm Average 2+ ppm R and D Peak 9.3+ ppm Duration 10+ yrs Duration 10+ yrs Total cohort High exposure Not used Not used Not used Not used Leukemia All All All All All All Myeloid All All Myeloid Myeloid All Myeloidc Myeloid RR 0.43 1.52 0.63 0.45 1.60 0.96 1.78 0.71 2.65 2.90 2.19 1.70 8.80 1.35 0.68 N 2 4 1 1 4 12 19 8 8 11 8 6 3 2 1 Samea Samea Samea Samea Samea Samea All exposed Total cohort Not used Not used Total cohort Total cohort Samea Total cohort All exposed All exposed All exposed Total cohort Total cohort

Bachand et al. 2010

RR = 1.05 (93-1.20)*
Exposed group Leukemia Samea Samea Samea Samea Samea Samea All Samea RR 0.43 1.52 0.63 0.45 1.60 0.96 1.02 0.91 N 2 4 1 1 4 12 116 31

All Samea All Samea All All All All All

1.09 0.75b 1.50 1.18 0.98 1.40 0.79 1.35 0.59

24 10 10 2 64 2 69 31 1

Employed < 1961 All

Diffn exp. assessment Used Beane-Freeman Government report Government report

*for cohort studies; 0.99 (95% CI: 0.71, 1.37) for case-control studies

Current Meta-analysis
RR = 1.53 (1.11-2.11)
Author, year Beane Freeman et al Hauptmann et al Stern et al, 198772 Cohort Ten US industries Embalmers Tannery workers Exposed group Peak 4+ ppm Peak 9.3+ ppm Leukemia Myeloid Myeloid RR N 1.78 19 2.90 11 2.19 1.70 8 6

Bachand et al. 2010 RR = 1.05 (93-1.20)*

Exposed group Leukemia RR All exposed Not used Total cohort Total cohort All Samea 1.09 0.75b 24 10 All 1.02 N 116

Pinkerton et al, 200471 Garment workers

Duration 10+ yrs Myeloid Duration 10+ yrs All

Summary of Results
Overall summary RR: 1.53 (95% CI, 1.11-2.11) Unlikely due to chance: p = 0.005 Higher in myeloid leukemia: RR = 2.47 (95% CI, 1.42-4.27) Including other subtypes could bias RR toward the null: RR = 1.42 (1.21-1.67) Higher RR with higher exposures: RR = 1.55 (95% CI, 1.04-2.31) Assessing an all exposed category might miss effects: RR = 1.07 (95% CI, 0.86-1.32) Higher RRs with adjustment for the healthy worker effect Excess RRs about 10-25% higher Higher RRs in professional workers compared to industrial workers Related to one study Coggon et al.

CO-INVESTIGATORS Erika Schwilk MD, MPH: UCSF, Kaiser Luoping Zhang PhD: Associate Professor, University of California Berkeley Martyn Smith PhD: Professor, University of California Berkeley Allan H Smith MD, PhD: Professor, University of California Berkeley Northern California Center for Occupational and Environmental Health (COEH)