- Short Report
- Open Access
The synergy factor: a statistic to measure interactions in complex diseases
- Mario Cortina-Borja^{1},
- A David Smith^{2},
- Onofre Combarros^{3, 4}Email author and
- Donald J Lehmann^{2}
https://doi.org/10.1186/1756-0500-2-105
© Combarros et al; licensee BioMed Central Ltd. 2009
- Received: 12 December 2008
- Accepted: 15 June 2009
- Published: 15 June 2009
Abstract
Background
One challenge in understanding complex diseases lies in revealing the interactions between susceptibility factors, such as genetic polymorphisms and environmental exposures. There is thus a need to examine such interactions explicitly. A corollary is the need for an accessible method of measuring both the size and the significance of interactions, which can be used by non-statisticians and with summarised, e.g. published data. The lack of such a readily available method has contributed to confusion in the field.
Findings
The synergy factor (SF) allows assessment of binary interactions in case-control studies. In this paper we describe its properties and its novel characteristics, e.g. in calculating the power to detect a synergistic effect and in its application to meta-analyses. We illustrate these functions with real examples in Alzheimer's disease, e.g. a meta-analysis of the potential interaction between a BACE1 polymorphism and APOE4: SF = 2.5, 95% confidence interval: 1.5–4.2; p = 0.0001.
Conclusion
Synergy factors are easy to use and clear to interpret. Calculations may be performed through the Excel programmes provided within this article. Unlike logistic regression analysis, the method can be applied to datasets of any size, however small. It can be applied to primary or summarised data, e.g. published data. It can be used with any type of susceptibility factor, provided the data are dichotomised. Novel features include power estimation and meta-analysis.
Keywords
- Binary Interaction
- Accessible Method
- Susceptibility Factor
- High Order Interaction
- Bootstrap Approximation
Background
The need
The remarkable progress made in the understanding of single-cause diseases has not yet been matched in the study of complex conditions. One problem is that susceptibility factors, e.g. genetic and environmental, all contribute risk that is to varying extents contingent on the presence of other factors [1–4]. Complex diseases cannot therefore be simply seen as due to the accumulation of many small independent effects. Rather, their very complexity lies in the interactions between contingent effects. Important effects may thus be missed if only single factors are independently examined (Discussion). The study of interactions between risk factors is thus central to the study of complex diseases.
Yet, unravelling interactions has proved confusing (Discussion). There is a need for a readily accessible method of measuring their strength, available to non-statisticians and applicable to summarised data and to datasets of any size. Methods are also needed to calculate the power to detect an interaction and to perform meta-analyses of interactions from published data; these two functions have not so far been readily available. There is a particular need for an accessible method for referees; untested claims of synergy are regularly published. Here we present a statistic, the synergy factor (SF), derived from logistic regression models, which aims to address these needs.
Modelling interactions in case-control studies
This paper is about statistical interactions; thus, drawing inferences about biological causality is beyond its scope. In general, a statistical interaction arises "when the effect of one explanatory variable depends on the particular level or value of another explanatory variable" [5]. Interactions may correspond to deviations from additive or multiplicative models for the joint effects of two risk factors. This has been thoroughly explored by Berrington de González and Cox [6, 7], with two procedures, one for each model.
Some epidemiologists, e.g. Rothman and Greenland [8], argue that assessment of interaction should be based on additive rate or risk models. These models are the norm in cohort studies. However, to assess interaction as departure from additive risks in case-control studies, three surrogate measurements of interaction based on the parameters of logistic regression models have been proposed [9, 10]: the relative excess risk due to interaction, the attributable proportion due to interaction and the synergy index. Skrondal has shown [11] that only the synergy index may be validly used for this purpose and only after fitting a linear odds model.
In case-control studies, the parameter which is both estimable and interpretable as a relative risk is the odds ratio (OR) [11]. In such studies, the predicted joint effect of two genetic or other factors may be defined as the product of the effects of each factor alone. We therefore propose a single statistic, the synergy factor (SF), which depends on a multiplicative definition of the null hypothesis.
Methods
A full description of the methodology for significance tests based on the SF appears in Additional file 1. We show there that ln(SF) is equivalent to the interaction term defined by two binary factors in a logistic regression model. We test the hypothesis of no interaction, using a Normal approximation for the statistic ln(SF)/stderr(ln(SF)), where the standard error of ln(SF) is easily obtained via the delta method [12]. This approximation is adequate even for relatively small sample sizes. We discuss a modification of the SF to cope with empty cells and propose two bootstrap approximations and a Bayesian inferential procedure that can be used as alternatives to the Normal approximation. We also propose methodology to calculate the power of significance tests and to perform meta-analyses based on the SF.
Results
The synergy factor (SF)
Let us assume we wish to estimate from a case-control study whether there is an interaction between any two (binary) factors, x_{1} and x_{2}, in the risk of a certain (binary) condition. Taking subjects with neither factor as reference, we first estimate the ORs for factor x_{1} alone (OR_{1}), factor x_{2} alone (OR_{2}) and both factors combined (OR_{12}). The SF is then defined as: SF = OR_{12}/(OR_{1} × OR_{2}) and is the ratio of the observed OR for both factors combined, to the predicted OR assuming independent effects of each factor. Susceptibility factors may be associated with increased or reduced risk, i.e. risk or protective factors, respectively (we make no assumptions about causality). In either case, interactions may be positive (synergy) or negative (antagonism). Thus, if SF > 1 (< 1), then there is a positive (negative) interaction between two risk factors. The opposite applies to protective factors.
Odds ratios of Alzheimer's disease, taking subjects with the BACE1 rs638405 C allele and without APOE4 as reference
BACE1 | APOE4 | Controls | Cases | OR |
---|---|---|---|---|
C+ | - | 125 | 80 | Reference |
GG | - | 80 | 38 | 0.742 |
C+ | + | 48 | 74 | 2.409 |
GG | + | 19 | 60 | 4.934 |
Totals | 272 | 252 |
Synergy between risk factors
Let us take the potential interaction in risk of Alzheimer's disease (AD) between the ε4 allele of apolipoprotein E (APOE4) and the GG genotype of the C/G polymorphism (rs638405) in exon 5 of the β-site APP-cleaving enzyme (BACE1) [13] (Table 1). Taking subjects with neither BACE1 GG nor APOE4 as reference, the OR for BACE1 GG alone was 0.742 and that for APOE4 alone was 2.409. That gave a predicted OR of 1.788 (= 0.742 × 2.409) for the combination, compared with an observed OR of 4.934. Hence: SF = 2.76 (= 4.934/1.788), 95% confidence interval (CI): 1.25–6.09, ln(SF) = 1.015, stderr(ln(SF)) = 0.404, Z = 2.25 and p = 0.012. Thus the null hypothesis of no interaction was rejected and significant synergy was found. The observed joint effect of the two variants was nearly three times greater than the predicted joint effect.
The above example is of synergy between risk factors. Examples of antagonism and of protective factors are given in Additional file 2. SF calculations may be performed using the Excel programme in Additional file 3; an R function is available (from MCB) to compute the bootstrap approximation.
Power
Meta-analyses
Data for an SF meta-analysis of the interaction between BACE1 rs638405 GG and APOE4
Study | APOE4-positive, BACE1GG | APOE4-positive, BACE1C+ | APOE4-negative, BACE1GG | APOE4-negative, BACE1C+ | ||||
---|---|---|---|---|---|---|---|---|
Controls | Cases | Controls | Cases | Controls | Cases | Controls | Cases | |
Nowotny et al 2001 [13] | 19 | 60 | 48 | 74 | 80 | 38 | 125 | 80 |
Gold et al 2003 [15] | 3 | 14 | 16 | 16 | 41 | 16 | 90 | 46 |
Clarimon et al 2003 [14] | 4 | 40 | 10 | 40 | 21 | 18 | 52 | 38 |
Kirschling et al 2003 [16] | 22 | 48 | 40 | 62 | 63 | 22 | 112 | 50 |
Discussion
The need
The real examples in Tables 1 and 2 and Table S1–S3 [Additional file 2] show the dangers of neglecting interactions. In all these examples, the effects of one or both variants were completely masked by the interacting factor. For instance, in the meta-analysis of four BACE1 studies (Table 2 and Figure 3), the effect of the BACE1 exon 5 GG was hidden in the absence of APOE4 [pooled OR = 0.8 (95% CI: 0.6–1.1; p = 0.17), random effects model [17]], but revealed in its presence [1.9 (1.3–2.9; 0.0015)]. Tables S1–S3 [Additional file 2] give further examples of such masking.
There is a common view that interactions, e.g. between genes (epistasis), should only be examined between risk factors that have already shown a significant main effect. But in many cases, such as most of the above, the association would be missed by the traditional single-factor approach [1–3]. Indeed, this was so in most of the examples of significant epistasis uncovered in our recent survey of sporadic AD [18]. Out of 36 such examples, 34 with SFs ≥ 2, the main effects of the gene variants other than APOE4 were generally very weak. The ORs were ≤ 1.2 in 20 out of 36 cases and were only significant in 5 cases. Thus, preliminary screening for main effects will miss many, possibly most cases of epistasis.
On the other hand, synergy can be too easily claimed. A common misconception is that a high combined OR necessarily implies synergy. A single OR by itself says nothing about synergy; it is the relation between the three relevant ORs that matters. For instance, let us assume that two risk factors are associated with ORs of 3 and 5 alone and of 15 when combined. Although the combined value is impressive, there is no synergy: SF = 15/(3 × 5) = 1. Claims of synergy are frequently published on the basis of such invalid evidence. Indeed, we have noted at least 20 claims of interactions, in the field of AD genetics alone, that were published in leading journals in recent years, but which may be clearly refuted by SF analysis. There is thus a need for a readily accessible method of testing such claims.
Limitations of the SF method
We suggest that SF analysis, being based on logistic regression analysis, is best used for assessing binary interactions [2]. Various methods have been devised to examine higher order interactions [19, 20]. However, some have only limited value for purposes of interpretation. Moreover, nearly all case-control sample-sets currently used for association studies lack the power for the proper study of higher order interactions [18]. Where a third interacting factor is suspected and a sufficiently large dataset is available, SF analysis may be performed twice, after stratification by the third factor, e.g. gender.
Where the relevant data are available, logistic regression analysis is the appropriate method for adjusting for covariates, while SF analysis should be the preferred method for stratification by covariates. Stratification can produce very small subsets, even of zero, which logistic regression analysis cannot handle. In contrast, SF analysis produces a realistic p value in each subgroup, if one adds 0.5 to each cell in any 4 × 2 table with at least one zero cell [21, 22].
Advantages of the SF method
SF analysis is simple to perform, through the Excel programmes in Additional files 3 and 4. It is a matter of a few minutes to perform the analysis, e.g. to check a claim of synergy in a published paper. The value of the method may be seen in the study of Combarros et al 2008 [18], in which SF analysis was used to examine each of the 89 studies of interactions cited in that review. The method measures both the size and significance of a binary interaction, using either primary or summarised data. Unlike logistic regression analysis, it can be applied to datasets of any size, however small, even with zero cells (above). The method can be used with all types of susceptibility factors, both risk and protective, for instance, age, gender, diet, medication or genetic polymorphisms, provided the data are dichotomised, e.g. age ± 75 years. It can be applied both to synergistic and to antagonistic interactions. Novel features include power estimation (through an R function available from MCB) and meta-analysis, an increasingly important application (through the Excel programme in Additional file 4). Neither function has been readily available before.
Declarations
Acknowledgements
We are most grateful to Dr Jonathan Marchini for his detailed reading of a previous version of the manuscript and to Dr Kirsty Little for her advice on implementing the procedure for doing meta-analyses in Excel. Most of this work was undertaken at GOSH/UCL Institute of Child Health, which received a proportion of funding from the Department of Health's NIHR Biomedical Research Centres funding scheme. The Centre for Paediatric Epidemiology and Biostatistics also benefits from funding support from the Medical Research Council in its capacity as the MRC Centre of Epidemiology for Child Health (G0400546). OPTIMA is supported by major grants from the Charles Wolfson Charitable Trust and Merck Inc. and is a centre in the Alzheimer's Research Trust Network.
Authors’ Affiliations
References
- Culverhouse R, Suarez BK, Lin J, Reich T: A perspective on epistasis: limits of models displaying no main effect. Am J Hum Genet. 2002, 70: 461-471. 10.1086/338759.PubMed CentralView ArticlePubMedGoogle Scholar
- Moore JH, Williams SM: New strategies for identifying gene-gene interactions in hypertension. Ann Med. 2002, 34: 88-95. 10.1080/07853890252953473.View ArticlePubMedGoogle Scholar
- Moore JH: The ubiquitous nature of epistasis in determining susceptibility to common human diseases. Hum Hered. 2003, 56: 73-82. 10.1159/000073735.View ArticlePubMedGoogle Scholar
- Pembrey M: The Avon Longitudinal Study of Parents and Children (ALSPAC): a resource for genetic epidemiology. The ALSPAC Study Team. Eur J Endocrinol. 2004, 151: U125-U129. 10.1530/eje.0.151U125.View ArticlePubMedGoogle Scholar
- Fitzmaurice G: The meaning and interpretation of interaction. Nutrition. 2000, 16: 313-314. 10.1016/S0899-9007(99)00293-2.View ArticlePubMedGoogle Scholar
- Berrington de González A, Cox DR: Additive and multiplicative models for the joint effect of two risk factors. Biostatistics. 2005, 6: 1-9. 10.1093/biostatistics/kxh024.View ArticlePubMedGoogle Scholar
- Berrington de González A, Cox DR: Interpretation of interaction: a review. Ann Appl Stat. 2007, 1: 371-385. 10.1214/07-AOAS124.View ArticleGoogle Scholar
- Rothman KJ, Greenland S: Modern Epidemiology. 1998, Philadelphia: Lippincott-Raven, 2Google Scholar
- Rothman KJ: Synergy and antagonism in cause-effect relationships. Am J Epidemiol. 1974, 99: 385-388.PubMedGoogle Scholar
- Rothman KJ: The estimation of synergy or antagonism. Am J Epidemiol. 1976, 103: 506-511.PubMedGoogle Scholar
- Skrondal A: Interaction as departure from additivity in case-control studies: a cautionary note. Am J Epidemiol. 2003, 158: 251-258. 10.1093/aje/kwg113.View ArticlePubMedGoogle Scholar
- Tanner M: Tools for statistical inference. 1990, Berlin: Springer-VerlagGoogle Scholar
- Nowotny P, Kwon JM, Chakraverty S, Nowotny V, Morris JC, Goate AM: Association studies using novel polymorphisms in BACE1 and BACE2. Neuroreport. 2001, 12: 1799-1802. 10.1097/00001756-200107030-00008.View ArticlePubMedGoogle Scholar
- Clarimón J, Bertranpetit J, Calafell F, Boada M, Tàrraga L, Comas D: Association study between Alzheimer's disease and genes involved in Aβ biosynthesis, aggregation and degradation: suggestive results with BACE1. J Neurol. 2003, 250: 956-961. 10.1007/s00415-003-1127-8.View ArticlePubMedGoogle Scholar
- Gold G, Blouin JL, Herrmann FR, Michon A, Mulligan R, Duriaux Saïl G, Bouras C, Giannakopoulos P, Antonarakis SE: Specific BACE1 genotypes provide additional risk for late-onset Alzheimer disease in APOE ε 4 carriers. Am J Med Genet. 2003, 119B: 44-47. 10.1002/ajmg.b.10010.View ArticlePubMedGoogle Scholar
- Kirschling CM, Kölsch H, Frahnert C, Rao ML, Maier W, Heun R: Polymorphism in the BACE gene influences the risk for Alzheimer's disease. Neuroreport. 2003, 14: 1243-1246. 10.1097/00001756-200307010-00011.View ArticlePubMedGoogle Scholar
- Der Simonian R, Laird N: Meta-analysis in clinical trials. Control Clin Trials. 1986, 7: 177-188. 10.1016/0197-2456(86)90046-2.View ArticleGoogle Scholar
- Combarros O, Cortina-Borja M, Smith AD, Lehmann DJ: Epistasis in sporadic Alzheimer's disease. Neurobiol Aging. 2008,Google Scholar
- Musani SK, Shriner D, Liu N, Feng R, Coffey CS, Yi N, Tiwari HK, Allison DB: Detection of gene × gene interactions in genome-wide association studies of human population data. Hum Hered. 2007, 63: 67-84. 10.1159/000099179.View ArticlePubMedGoogle Scholar
- Thornton-Wells TA, Moore JH, Haines JL: Genetics, statistics and human disease: analytical retooling for complexity. Trends Genet. 2004, 20: 640-647. 10.1016/j.tig.2004.09.007.View ArticlePubMedGoogle Scholar
- Anscombe FJ: On estimating binomial response relations. Biometrika. 1956, 43: 461-464.View ArticleGoogle Scholar
- Breslow N: Odds ratio estimators when the data are sparse. Biometrika. 1981, 68: 73-84. 10.1093/biomet/68.1.73.View ArticleGoogle Scholar
Copyright
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.