Conducting post-hoc tests of compound coefficients using simple slopes for a categorical by categorical interaction Jane E. Miller, PhD The Chicago Guide to Writing about Multivariate Analysis, 2nd Edition. Overview • • • • Simple slopes defined Application of simple slopes to interactions Calculation of standard errors for simple slopes Charts to show conclusions of simple slope tests for interactions The Chicago Guide to Writing about Multivariate Analysis, 2nd Edition. Post-hoc probing • Post-hoc (“after the fact”) probing is a way to conduct formal statistical tests of differences that were not specified for a priori inferential testing, e.g., – Comparisons other than against the reference category • See podcast on testing statistical significance of differences across coefficients (s) – Those involving more than one , e.g., interactions • Can be conducted from output that is easily obtained from standard software packages • Does not require respecifying the model The Chicago Guide to Writing about Multivariate Analysis, 2nd Edition. Simple slopes calculation for compound coefficients • The simple slopes technique is a way to calculate the point estimate of effect size for a combination of two coefficients from the same model • The point estimate of effect size for each of those combinations is a compound coefficient – A linear combination of two is • Models involving main effects and interaction terms require summing more than one to obtain the overall effect of the variables involved in the interaction The Chicago Guide to Writing about Multivariate Analysis, 2nd Edition. Review: Calculation of overall effect of an interaction • The predicted value of the dependent variable from a model with a two-way interaction between predictors X1 and X2 can be written as a function of the estimated coefficients βi: Y = β0 + β1 X1 + β2X2 + β3 X1_X2 • Rearranging terms to group those that involve the focal predictor (X1 ) yields: (β0 + β2 X2) + (β1 +β3X2) X1 , simple intercept ώ0 simple slope ώ1 The Chicago Guide to Writing about Multivariate Analysis, 2nd Edition. Review: Focal and modifier variables in an interaction • In a statistical interaction between two independent variables X1 and X2 – X1 is sometimes referred to as the focal predictor – X2 is called the moderator or modifier variable because it alters the association between X1 and Y • Identifying which of the IVs in an interaction is the modifier depends on the research question – E.g., race/ethnicity is specified as the modifier because it is hypothesized to change the association between education and birth weight The Chicago Guide to Writing about Multivariate Analysis, 2nd Edition. Standard errors for simple slopes • Both the simple intercept and the simple slope are compound coefficients – Each is a linear combination of two estimated is • Simple slope: ώ0 = β0 + β2 X2 • Simple intercept : ώ1 = (β1 +β3X2) X1 • The inferential statistical tests for a compound coefficient require calculating the standard error of the simple slope using the estimated variances and covariances for those is s.e. (ώ1 | X2) = √ [var(β1) + (2 × X2× cov(β1,β3)) + (X2)2× var(β3)] The Chicago Guide to Writing about Multivariate Analysis, 2nd Edition. Calculating the standard error for a compound coefficient • s.e. (ώ1 | X2) is the standard error of the simple slope ώ1 s.e. (ώ1 | X2) = √ [var(β1) + (2 × X2× cov(β1β3)) + (X2)2× var(β3)] • Where var(j) and var(k) are the variances of j and k, respectively cov(j, k) is the covariance between j and k • The variance-covariance matrix can be requested as part of a regression command • Note that s.e. (ώ1 | X2) depends on the value of the moderator variable in the interaction, X2 The Chicago Guide to Writing about Multivariate Analysis, 2nd Edition. Formulas to calculate overall interaction pattern BW = f (NHB, MA, <HS, =HS, NHB_<HS, NHB_=HS, MA_<HS, MA_=HS) Subgroup βs involved in overall effect for that group* Non-Hispanic white, < HS = β<HS Non-Hispanic white, = HS = β=HS Non-Hispanic white, > HS NA (reference category) Non-Hispanic black, < HS = βNHB + β<HS + βNHB_<HS Non-Hispanic black, = HS = βNHB + β=HS + βNHB_=HS Non-Hispanic black, > HS = βNHB Mexican American, < HS = βMA + β<HS + βMA_<HS Mexican American, = HS = βMA + β=HS + βMA_=HS Mexican American, > HS = βMA • More than one β is involved in the calculations of overall effect for any subgroup that is not in the reference category for either IV in the interaction * Difference in birth weight compared to reference category The Chicago Guide to Writing about Multivariate Analysis, 2nd Edition. Estimated coefficients from an OLS model of birth weight in grams β t-statistic Main effects terms Race (ref. = non-Hispanic white) Non-Hispanic Black (NHB) Mexican American (MA) Mother’s ed. (ref. = > HS) Less than high school (<HS) High school graduate (=HS) Interactions: race & education NHB_<HS MA_<HS NHB_=HS MA_=HS F-statistic Degrees of freedom (df) ** denotes p < 0.01 –168.1** –104.2** –5.66 –2.16 –54.2** –62.0** –2.35 –3.77 –38.5 99.4 18.4 93.7 65.59 13 –0.88 1.72 0.47 1.49 The Chicago Guide to Writing about Multivariate Analysis, 2nd Edition. Example calculation of a simple slope ώ1 = (β1 +β3X2) × X1 • Where X1 = NHB and X2 = <HS ώ1 = (β<HS +β<HS_NHB) × NHB • From the model results on the previous slide β<HS = –54.2 β<HS_NHB = –38.5 • For non-Hispanic black infants, NHB = 1 • Substituting those values into the equation: ώ1 = (–54.2 + (–38.5)) × 1 = –92.7 The Chicago Guide to Writing about Multivariate Analysis, 2nd Edition. Example: Calculation of the standard error of the simple slope for one race_education category Variance-covariance matrix for the estimated coefficients (βs) βNHB β<HS β<HS_NHB βNHB 882.10 β<HS 522.94 β <HS_NHB –461.22 1,918.81 • In this example, <HS is X1, coded 1 = non-Hispanic black 0 = other racial/ethnic groups NHB is X2, coded 1 = non-Hispanic black 0 = other racial/ethnic groups s.e. (ώ1 | NHB) = √ [var(β<HS)+(2×NHB × cov(β<HSβ<HS_NHB))+(NHB)2 × var(β<HS_NHB)] = √ [522.94 + 2 × 1× (–461.22) + (1)2× 1,918.81] = 39.0 The Chicago Guide to Writing about Multivariate Analysis, 2nd Edition. Predicted birth weight with 95% CI, by race and educational attainment <HS =HS >HS Difference in birth weight (grams) 50 0 -50 -100 Non-Hispanic white -150 Mexican-American -200 Non-Hispanic black -250 -300 -350 -400 The Chicago Guide to Writing about Multivariate Analysis, 2nd Edition. Which contrasts are possible with non-Hispanic white, > HS as the reference category Predicted birth weight (grams) by mother’s education and race/ethnicity, United States, 1988–1994 NHANES III Mother’s educational attainment Race/ethnicity < HS = HS > HS Non-Hispanic white 3,044 3,044 3,112 Mexican American 3,027 3,031 3,003 Non-Hispanic black 2,820 2,883 2,937 • The circled cell is the reference category (Non-Hispanic white, mother’s education > HS) • Yellow-shaded cells can be compared to the reference category based on the standard errors of the associated main effects terms alone • Green-shaded cells can be compared to the reference category using standard errors calculated using the simple slope for a compound coefficient The Chicago Guide to Writing about Multivariate Analysis, 2nd Edition. Ballpark assessment of other contrasts • If there is substantial overlap between confidence intervals for two values, you can usually safely conclude that they are not statistically significantly different from one another – It is very unlikely that taking the covariance between the two βs into account when computing the standard error of their simple slopes would alter that conclusion • If the confidence intervals do not overlap or overlap only slightly, formally test the statistical significance of that difference by respecifying the model with a different reference category – See the podcast on that topic The Chicago Guide to Writing about Multivariate Analysis, 2nd Edition. Visual assessment of other contrasts based on confidence intervals from simple slope <HS Difference in birth weight (grams) • The confidence intervals for non-Hispanic white and Mexican American infants overlap considerably in the < HS and = HS groups • Even if these birth weight differences were statistically significant, they are so small that they would be of trivial substantive interest =HS >HS 50 0 -50 -100 -150 -200 -250 -300 -350 -400 Non-Hispanic white Mexican-American Non-Hispanic black The Chicago Guide to Writing about Multivariate Analysis, 2nd Edition. Another contrast to test with a different model specification • Re-specify the model with non-Hispanic black as the reference category <HS Difference in birth weight (grams) • The confidence intervals for non-Hispanic black infants don’t overlap very much in the < HS and > HS groups • The birth weight difference between these groups is >100 grams, so it is worth testing whether that difference is statistically significant =HS >HS 50 0 -50 -100 -150 -200 -250 -300 -350 -400 Non-Hispanic white Mexican-American Non-Hispanic black The Chicago Guide to Writing about Multivariate Analysis, 2nd Edition. Online calculator for simple slopes • Preacher has created an online calculator to compute simple intercepts, simple slopes, and their standard errors from regression output – Coefficients (βs) – Variances and covariances of the βs – Values of the moderator variable for which to calculate standard errors of the simple slope – Degrees of freedom for the model • Can also graph the shape of the interaction with confidence intervals, given values of the focal predictor and moderator variables • See http://www.quantpsy.org/interact/index.html The Chicago Guide to Writing about Multivariate Analysis, 2nd Edition. Presenting results of post-hoc tests • In methods section, mention use of simple slopes technique to calculate standard errors for compound coefficients. – Provide citation to that method • Create a table to present the is, standard errors, and goodness-of-fit statistics from the multivariate model • Conduct post-hoc tests for your substantive hypotheses behind the scenes – Calculate the simple slope – Calculate the standard error of the simple slope • Create a table or chart to report predicted values and confidence intervals calculated from is and standard errors – Use symbols or prose to convey which contrasts other that against the reference category are statistically significantly different from one another The Chicago Guide to Writing about Multivariate Analysis, 2nd Edition. Summary • Overall pattern of an interaction involves calculation of compound coefficients • Standard errors for compound coefficients can be calculated using the simple slope technique based on variance-covariance matrix from regression output • Comparison of predicted values of the dependent variable are against the reference category, with all continuous independent variables set to 0 – To conduct contrasts for other values of IVs in the interaction • Respecify the model with different reference categories • Use the all-interaction-dummies approach Separate podcasts • Conduct calculations behind the scenes, present the conclusions of those tests in the text The Chicago Guide to Writing about Multivariate Analysis, 2nd Edition. Suggested resources • Cohen, Jacob, Patricia Cohen, Stephen G. West, and Leona S. Aiken. 2003. Applied Multiple Regression/Correlation Analysis for the Behavioral Sciences, 3rd Edition. Florence, KY: Routledge, chapters 7, 8, and 9. • Figueiras, Adolfo, Jose Maria Domenech-Massons, and Carmen Cadarso. 1998. Regression Models: Calculating the Confidence Interval of Effects in the Presence of Interactions. Statistics in Medicine 17: 2099–2105. • Preacher, Kristopher J., Patrick J. Curran, and Daniel J. Bauer. 2006. “Computational Tools for Probing Interaction Effects in Multiple Linear Regression, Multilevel Modeling, and Latent Curve Analysis.” Journal of Educational and Behavioral Statistics.31: 437–448. • Miller, J. E. 2013. The Chicago Guide to Writing about Multivariate Analysis, 2nd Edition. University of Chicago Press, chapter 16. The Chicago Guide to Writing about Multivariate Analysis, 2nd Edition. Suggested online resources • Preacher, Kristopher J. 2011. “Probing Interactions in Multiple Linear Regression, Latent Curve Analysis, and Hierarchical Linear Modeling: Interactive Calculation Tools for Establishing Simple Intercepts, Simple Slopes, and Regions of Significance.” • Available online at http://www.quantpsy.org/interact/index.html The Chicago Guide to Writing about Multivariate Analysis, 2nd Edition. Suggested online resources, cont. • Podcasts on – Calculating the overall shape of an interaction from OLS coefficients – Approaches to testing statistical significance of interactions – Using alternative reference categories to test statistical significance of interactions The Chicago Guide to Writing about Multivariate Analysis, 2nd Edition. Suggested practice exercises • Study guide to The Chicago Guide to Writing about Multivariate Analysis, 2nd Edition. – Question #5 in the problem set for chapter 16 – Suggested course extensions for chapter 16 • “Reviewing” exercise #2 • “Applying statistics and writing” exercise #2 The Chicago Guide to Writing about Multivariate Analysis, 2nd Edition. Contact information Jane E. Miller, PhD jmiller@ifh.rutgers.edu Online materials available at http://press.uchicago.edu/books/miller/multivariate/index.html The Chicago Guide to Writing about Multivariate Analysis, 2nd Edition.