week 11

Learning Objectives

  • Define interaction and effect modification.
  • Discuss two strategies to evaluate interaction:
    • Heterogeneity of effects.
    • Comparing observed vs. expected joint effects.
  • Be aware of issues when evaluating interaction.
  • Differentiate between interaction, confounding, and mediation.

Definition of Interaction

  • Interaction is when two or more risk factors modify the effect of each other with regard to the occurrence or level of a given outcome.
  • For dichotomous variables: the effect of the exposure on the outcome differs depending on whether another variable (the effect modifier) is present.

Synergistic and Antagonistic Interaction

  • Synergistic (positive interaction): The effect modifier strengthens the effect of the exposure of interest.
  • Antagonistic (negative interaction): The effect modifier diminishes or eliminates the effect of the exposure of interest.

Strategies to Evaluate Interaction

  • Two equivalent approaches:
    1. Heterogeneity of effects: Interaction occurs when the effect of a risk factor A on the risk of an outcome Y is not homogeneous in strata formed by a third variable Z. Variable Z is the effect modifier.
    2. Comparing observed vs. expected joint effects: Interaction occurs when the observed joint effect of A and Z differs from that expected on the basis of their independent effects.
  • Each can be used to assess interaction on the multiplicative and on the additive scale.

Approach #1: Heterogeneity of Effects

  • Assess if the association between a risk factor X and disease Y is of similar magnitude in subgroups of the population.
  • If NO, interaction may be present.
  • If YES, no interaction.
  • Consider if random variability, confounding, or bias are present.

Example of Interaction

  • In a study in Eastern Finland, the association between alcohol consumption and dementia was modified by a gene (APOE-4).
    • Gene=YES: Positive effect of alcohol consumption and dementia.
    • Gene=NO: Null effect between alcohol consumption and dementia.
  • The gene is considered the effect modifier rather than alcohol use because the effect of the gene on dementia depends on alcohol consumption.

Heterogeneity of Effects

  • The association is stratified according to the suspected effect modifier.
  • The effect of the risk factor is examined within each strata.
  • Compare the stratified estimates to each other and the crude estimate.
  • This is similar to how confounding is assessed.

Confounding vs. Effect Modification

  • Confounders are treated as a nuisance.
  • Effect modification is an interesting interaction to understand.

Heterogeneity of Effects: Multiplicative Interaction

  • Multiplicative interaction is present when the relative ratio in the risk of an outcome Y between subjects exposed and those not exposed to a putative risk factor A differs as a function of a third variable Z.
  • Examples of relative measures of disease association:
    • Relative risk
    • Relative rate
    • Odds ratio
    • Hazard ratio

Example: Oropharyngeal (OP) Cancer and Alcohol/Asbestos

  • The relative risk of getting oropharyngeal cancer among people who did and did not use alcohol varies depending on whether they were exposed to asbestos.
  • Assess by calculating the relative risk for alcohol and OP cancer among people who were exposed to asbestos, and people who were not exposed to asbestos.

Example 1: Multiplicative Interaction

  • Example demonstrates how to calculate relative risk (RR) for alcohol and OP cancer with and without asbestos exposure.
  • If the RR is the same in each level of asbestos exposure, interaction is NOT present.

Example 2: Multiplicative Interaction

  • Example demonstrates how to calculate relative risk (RR) for alcohol and OP cancer with and without asbestos exposure.
  • If the RR differs by level of asbestos exposure, interaction IS present on the multiplicative scale.

Heterogeneity of Effects: Additive Interaction

  • Additive interaction is present when the attributable risk of disease in those exposed to factor A varies as a function of a third variable Z.
  • How to assess interaction on the additive scale:
    • Calculate the attributable risks/rates for those exposed to risk factor A for each stratum defined by levels of the potential effect modifier Z.
  • Examples of additive measures of disease association:
    • Risk/rate difference
    • Risk difference proportion
    • Population attributable risk

Example 3: Additive Interaction

  • If the risk difference (RD) differs by level of Z, interaction IS present on the additive scale.

Example 4: Additive Interaction

  • In each level of Z, the AR is the same, interaction is NOT present on the additive scale.

Heterogeneity of Effect: Interpretation

  • Once you stratify, you cannot estimate the independent effect of the interaction variable (Z).
  • Can estimate:
    • The overall (average) effect of E (alcohol) on D (OP cancer).
    • The overall (average) effect of E on D among Z+ (people with asbestos exposure).
    • The overall (average) effect of E on D among Z- (people without asbestos exposure).

Statistical Test for Interaction

  • Breslow-Day test of homogeneity.
    • H0H_0: the strength of association is homogeneous across all strata (no interaction).
    • HAH_A: the strength of association differs across strata (there is interaction).
  • Interpretation: p < 0.05 - statistically there is interaction.

Approach #2: Joint Effects

  • Calculate the expected joint effects of the 2 risk factors in question on the outcome.
  • Examine whether the observed joint effect of the two risk factors is different from that expected from their independent effects.

Joint Effects

  • Evaluating interaction based on comparing expected and observed joint effects relies on an alternative definition of interaction:
    • The observed joint (combined) effect of A and Z on Y differs from the expected joint effect.

Joint Effects: Example

  • The RR of poor diet on the risk of having an MI is 3.0.
  • The RR of lack of exercise on the risk of having an MI is 2.0.
  • Then the expected joint effect = 6.0 (2.0 x 3.0).
  • If a combined effect of both diet and exercise on the risk of MI is different than 6, then there may be interaction.

Synergism and Antagonism

  • Synergism: the observed joint effect is greater than that expected from the individual effects.
  • Antagonism: the observed joint effect is smaller than that expected from the individual effects.

Expected vs Observed Joint Effects

  • When there is NO interaction, the observed joint effect of risk factors A and Z equals the sum of their independent effects.

Example of No Interaction

  • Examples given to illustrate when observed = expected, indicating no interaction.

Synergism (positive interaction)

  • Observed joint effect is greater than the expected effect (further from 1).

Examples of Synergism (positive interaction)

  • Examples given to illustrate when observed > expected, indicating synergism.

Antagonism (negative interaction)

  • Observed joint effect is less than the expected (closer to 1).

Examples of Antagonism (negative interaction)

  • Examples given to illustrate when observed < expected, indicating antagonism.

Evaluating Interaction using Expected vs. Observed Method

  • Common equivalent table set ups shown.

E vs. O: Multiplicative Scale

  • Example: Calculate the Expected joint RR and Observed joint RR.
  • If the joint effects of X and Z were no more than expected, then multiplicative interaction is NOT present.

E vs. O: Multiplicative Scale

  • Example: Calculate the Expected joint RR and Observed joint RR.
  • If the joint effects of X and Z were MORE than expected, then multiplicative interaction IS present.

E vs. O: Additive Scale

  • Example: Calculate the expected joint AR and the observed joint AR.
  • If the joint effects of X and Z were NO more than what we expected, then additive interaction is NOT present.

E vs. O: Additive Scale

  • Example: Calculate the expected joint AR and the observed joint AR.
  • If the joint effects of X and Z were more than expected, then additive interaction IS present.

Interchangeability of the Definitions of Interaction

  • Two equivalent approaches:
    1. Heterogeneity of effects
    2. Comparing observed vs. expected joint effects
  • The two definitions of interaction are interchangeable.

Strategies to Evaluate Interaction in Case-Control Studies

  • The assessment of interaction discussed has relied on absolute and relative measures of risk from cohort studies.
  • Special circumstances to be aware of for case-control studies.

Interaction in Case Control Studies

  • Assessing interaction using the heterogeneity of effect method:
    • Absolute measures of disease occurrence are not generally available in a case-control study.
    • Can only directly assess multiplicative interaction.
    • Use same method as for cohort studies.
  • Assessing interaction using the observed vs. expected joint effects method: use special equations.

Joint Effects in a Case Control Study

  • Detection of additive interaction:
    • Expected OR<em>11=obs OR</em>01+obs OR10–1Expected \ OR<em>{11} = obs \ OR</em>{01} + obs \ OR_{10} – 1
  • Detection of multiplicative interaction:
    • OR<em>11=obs. OR</em>10X OR01OR<em>{11} = obs. \ OR</em>{10} X \ OR_{01}

Interaction Fallacy

  • Can happen when assessing heterogeneity of odds ratios.
  • Odds ratio can overestimate the risk ratio when the disease is not rare (>10%).
  • When you stratify by a third factor, the disease may be very prevalent in one stratum and can overestimate the odds ratio in that stratum and give the appearance of interaction.

Heterogeneity Observed Due to Bias

  • Observed heterogeneity may be due to a bias present in only one stratum.
  • Example given.

Subgroup (Sensitivity) Analysis

  • Evaluation of presence of heterogeneity of effect modification (interaction) should be hypothesized a priori (as part of your analysis plan, before you start the analysis).
  • Avoids a fishing expedition to find an effect in some group.
  • Sometimes call ‘sub-group’ analyses, when associations are examined in one stratum only.

Quantitative vs. Qualitative Interaction

  • Quantitative interaction:
    • The associations within each stratum of the 3rd variable is in the same direction in relation to the null – but the strength of the association varies across strata.
  • Qualitative interaction – 2 ways:
    1. The effects of E on the outcome D are in opposite directions of the null according to the presence of Z.
    2. There is a significant association in one of the strata (by the third variable) but not in the other.

What to Report?

  • If there is evidence of interaction then the effect of E on D depends on Z.
  • Report the stratum specific measures of association (or joint effect measures).
  • The crude measure of association is not a biased estimate of the exposure disease association, so report it as well (it's an unbiased measure of the total population, or average effect).