Increasing testing severity in multi-group designs

R
statistics
Author

Aaron R. Caldwell

Published

November 20, 2020

Introduction

A little over a year ago Patrick Forscher wrote a very nice blog post about a simple way to increase the severity of a hypothesis test in a multi-group experimental design (in this case one factor, 3-group design). He highlights the main problem as the following:

Multi-group designs are the workhorse of scientific psychology. Multi-group designs apply to any grouping of people, within-person states, situations, or stimuli; interest typically centers around the either the means or conditional means within each group. However, theory-testing using these means typically proceeds in just the way that Meehl criticized: scientists attempt to reject a null hypothesis of no group mean differences (which, due to Meehl’s “crud”, may be trivially false a priori), then follow up the rejected null hypothesis with tests intended to diagnose the pattern that produced the rejected null.

I agree that this is systemic problem. People sometimes default to just plugging numbers into a one-way ANOVA and usually conclude that their hypothesis is confirmed simply by a significant ANOVA-level effect. Patrick’s proposed solution to the problem is a priori orthogonal contrasts. As a quick aside, I am always shocked at how few people know about orthogonal contrasts. This was drilled into my head by my statistics professors at Arkansas (Thank you Ronna Turner and Sean Mulvenon!). For more details on orthogonal contrasts (and experimental design altogether), I highly recommend purchasing a copy of “Designing Experiments and Analyzing Data” by Maxwell, Delaney, and Kelley (2018). It covers this and numerous other topics; you won’t find a better value in a statistics textbook.

R Code

Now I will go through the code in R for how to implement Patrick’s approach using the afex and emmeans package. For more details on why you should use this approach please read Patrick’s blogpost!

Orthogonal Contrasts

There are many helper functions in R that make it easy to use orthogonal contrasts without having to write out the contrasts by hand. I will show these later but I quickly want to show a technique for checking your set of contrasts to ensure they are in fact orthogonal. Essentially, I have a function that takes the contrasts as matrix and calculates the cross-products of the matrix. If any off-diagonal element is not equal to zero then the contrasts are not orthogonal. Please note, that I made this function in ~2 minutes so I haven’t quite checked to ensure it works in all situations.

# build function
check_orthog = function(contrast_mat) {
  check_off = crossprod(contrast_mat)
  upper_check = sum(check_off*upper.tri(check_off))
  lower_check = sum(check_off*lower.tri(check_off))
  if (upper_check == 0 && lower_check == 0) {
    return(TRUE)
  } else{
    return(FALSE)
  }
}

# Test with Patrick's code -- should print TRUE

c1 <- c(-1/2,0,1/2)
c2 <- c(1/3,-2/3,1/3)

cons = cbind(c1,c2)

check_orthog(cons)
[1] TRUE
# Create non-orthogonal contrasts -- should print FALSE
cons2 = cbind(c(-1,1,0),
              c(0,-1,1),
              c(-1,0,1))

check_orthog(cons2)
[1] FALSE
# Another way of coding orthogonal polynomial contrasts. 
c1 <- c(-1, 0, 1)
c2 <- c(0.5, -1, 0.5)

cons3 <- cbind(c1,c2)

# Again should print TRUE
check_orthog(cons3)
[1] TRUE

I will use this at points in the blogpost just to check my own contrast coding (please email me if you notice a mistake!).

Generate Data

First, I will generate some data. Please note for this portion of the post I will be directly replicating/copying Patrick’s data and code.

library(tidyverse)
library(afex)
library(emmeans)
library(broom)

set.seed(432432) # For reproducibility

# For all examples, I'm using the below sampling error sd and n per cell
err <- 1
n_per_cell <- 200
# I'm also assuming the following:
# (1) our smallest effect of substantive interest is .4
# (2) all contrasts are Helment (successively compare one group to the average of the others)
# (3) all contrasts are unit-weighted

# This method should generalize to other orthogonal contrasts
# The parameter estimates for unit-weighted Helmert contrasts, however, have a 
# nice interpretation as a series of mean differences and are therefore easy to understand


#### THREE GROUP ####

# Ceofficients for this case. Modify as desired
# First is the intercept, second is the focal contrast, remainder are the residual contrasts
coefs <- c(0, .7, .2)

# Create the data
# The last line creates the outcome using the coefficients above and the desired sampling error, err
# %*% is matrix multiplication
dat_three <- data.frame(group = rep(paste("group", 1:3), n_per_cell))
dat_three <- mutate(
  dat_three,
  c1 = case_when(group == "group 1" ~ 2 / 3, TRUE ~ -1 /
                   3),
  c2 = case_when(group == "group 1" ~ 0, group == "group 2" ~ 1 /
                   2, TRUE ~ -1 / 2),
  outcome = as.vector(cbind(1, c1, c2) %*% coefs + rnorm(nrow(dat_three), sd =
                                                           err))
) %>%
  mutate(group = factor(group,
                        levels = c("group 1",
                                   "group 2",
                                   "group 3"),
                        ordered = TRUE)) # creates and ordered & labeled factor for group


knitr::kable(head(dat_three),
             caption = "3-group data")
3-group data
group c1 c2 outcome
group 1 0.6666667 0.0 0.1940993
group 2 -0.3333333 0.5 -0.5684096
group 3 -0.3333333 -0.5 -0.3451584
group 1 0.6666667 0.0 1.4218414
group 2 -0.3333333 0.5 0.2550905
group 3 -0.3333333 -0.5 0.3534099

Now, we have some data to work with. However, rather than simply using the lm function to make the contrast comparisons I will use afex to build the model and then emmeans to apply the specific contrasts.

Three Group Example

In Patrick’s post he used the lm function to test the contrasts. I think this fine but may be difficult to understand for many beginners. Plus any addition comparisons will have to be made by creating a new model. This is why I prefer using afex: it has an easier to use interface and the model can be passed onto emmeans for specific additional comparisons. If Patrick’s approach works for you that is great, but I want to make sure people know of alternative approaches.

# Patrick's Example
m_three <- lm(outcome ~ c1 + c2, data=dat_three)
knitr::kable(tidy(m_three),
             caption = "3-group Contrasts using lm")
3-group Contrasts using lm
term estimate std.error statistic p.value
(Intercept) 0.0437823 0.0417829 1.047853 0.2951303
c1 0.6570966 0.0886348 7.413527 0.0000000
c2 0.1873883 0.1023467 1.830918 0.0676111

Now, let us go through the same process with afex and emmeans. We will need to create an id column so that afex knows that these are all between-subject comparisons. I also like to have the partial eta squared (\(\eta^2\)) for the default effect size output so I am also going to set this using the afex_options function.

# Add subject id
dat_three = rowid_to_column(dat_three, var = "id")
afex_options(es_aov = "pes")
# Build model using afex; note we must have an error term "(1|id)"
afex_three = afex::aov_4(outcome ~ group + (1|id),
                         data = dat_three)
Contrasts set to contr.sum for the following variables: group
# Output tyical ANOVA table with type-3 SS
knitr::kable(nice(afex_three))
Effect df MSE F pes p.value
group 2, 597 1.05 29.16 *** .089 <.001

Most people reading this should be familiar with the table above. It is a simple one-way ANOVA output.

Now I can make the same comparisons Patrick made with lm using emmeans. However, I can also make my tests directionl using the side argument in the contrast function.

# Now get emmeans
emm_three = emmeans(afex_three, ~group)

# Estimated marginal means
knitr::kable(tidy(emm_three) %>%
               select(-df,-statistic,-p.value))
group estimate std.error
group 1 0.4818467 0.07237
group 2 -0.0815557 0.07237
group 3 -0.2689441 0.07237
# Create orthogonal contrasts
con_three = list(c1 = c(1,-.5,-.5),
                 c2 = c(0,1,-1))
check_orthog(cbind(con_three$c1,
                   con_three$c2))
[1] TRUE
# Make comparisons with contrast function
knitr::kable(contrast(emm_three,con_three),
             caption = "3-group emmeans contrasts")
3-group emmeans contrasts
contrast estimate SE df t.ratio p.value
c1 0.6570966 0.0886348 597 7.413527 0.0000000
c2 0.1873883 0.1023467 597 1.830918 0.0676111
knitr::kable(contrast(emm_three,con_three,
                      side = ">"),
             caption = "3-group emmeans w/ directional contrasts")
3-group emmeans w/ directional contrasts
contrast estimate SE df t.ratio p.value
c1 0.6570966 0.0886348 597 7.413527 0.0000000
c2 0.1873883 0.1023467 597 1.830918 0.0338055

Patrick also mentions that we would want to perform equivalence tests to rule out that the differences are within our equivalence bounds (answers the question “Are these contrast differences, if they exist, smaller than what we consider meaningful?”). Remember, c1 is “group 1 - group 2 & group 3” and c2 is “group 2 - group 3”.

# Create contrast
con_three_eq = contrast(emm_three,con_three) 
# Perform equivalence test (takes absolute difference)
test_three_eq = test(con_three_eq,
                     delta = .4, # eq bound
                     side = "equivalence")
knitr::kable(test_three_eq,
             caption = "Equivalence Tests for 3-group example")
Equivalence Tests for 3-group example
contrast estimate SE df t.ratio p.value
c1 0.6570966 0.0886348 597 2.900628 0.9980694
c2 0.1873883 0.1023467 597 -2.077368 0.0190973

Four Group Example

Again, we will need to generate the same data that Patrick did in his post.

#### FOUR GROUP ####

# Ceofficients for this case. Modify as desired
# First is the intercept, second is the focal contrast, remainder are the residual contrasts
coefs <- c(0, .7, .2, .1)

# Create the data
# The last line creates the outcome using the coefficients above and the desired sampling error, err
# %*% is matrix multiplication
dat_four <- data.frame(group = rep(paste("group", 1:4), 
                                   n_per_cell))
dat_four <- mutate(
  dat_four,
  c1 = case_when(group == "group 1" ~ -3 / 4, TRUE ~ 1 /
                   4),
  c2 = case_when(group == "group 1" ~ 0, 
                 group == "group 2" ~ 2 /
                   3, TRUE ~ -1 / 3),
  c3 = case_when(
    group %in% c("group 1", "group 2") ~ 0,
    group == "group 3" ~ 1 / 2,
    TRUE ~ -1 / 2
  ),
  outcome = as.vector(cbind(1, 
                            c1,
                            c2, 
                            c3) %*% coefs + rnorm(nrow(dat_four), 
                                                  sd = err))
) %>%
  mutate(group = factor(
    group,
    levels = c("group 1",
               "group 2",
               "group 3",
               "group 4"),
    ordered = TRUE
  )) %>% # creates and ordered & labeled factor for group
rowid_to_column(var = "id")

knitr::kable(head(dat_four),
             caption = "4-Groups Data")
4-Groups Data
id group c1 c2 c3 outcome
1 group 1 -0.75 0.0000000 0.0 -0.6004404
2 group 2 0.25 0.6666667 0.0 1.4770844
3 group 3 0.25 -0.3333333 0.5 2.2156623
4 group 4 0.25 -0.3333333 -0.5 -0.1881092
5 group 1 -0.75 0.0000000 0.0 0.4617127
6 group 2 0.25 0.6666667 0.0 0.0611009

Now, we replicate the process to build our base model.

# Build model using afex; note we must have an error term "(1|id)"
afex_four = afex::aov_4(outcome ~ group + (1|id),
                         data = dat_four)
Contrasts set to contr.sum for the following variables: group
# Output tyical ANOVA table with type-3 SS
knitr::kable(nice(afex_four),
             caption = "ANOVA: Four Groups")
ANOVA: Four Groups
Effect df MSE F pes p.value
group 3, 796 1.02 25.52 *** .088 <.001

And, we can then pass on this model to the emmeans function to make our specific contrasts.

# Now get emmeans
emm_four = emmeans(afex_four, ~group,
                   adjust = "none")

# Estimated marginal means
knitr::kable(tidy(emm_four) %>%
               select(-df,-statistic,-p.value))
group estimate std.error
group 1 -0.5275826 0.0714616
group 2 0.3089254 0.0714616
group 3 0.1323786 0.0714616
group 4 0.0253503 0.0714616
# Create orthogonal contrasts
con_four = list(c1 = c(1,-1/3,-1/3,-1/3),
                c2 = c(0,1,-0.5,-0.5),
                c3 = c(0,0,.5,-.5))
check_orthog(cbind(con_four$c1,
                   con_four$c2,
                   con_four$c3))
[1] TRUE
# Perform joint test
knitr::kable(test(contrast(emm_four,
                           con_four[2:3]),
                  joint = TRUE),
             caption = "Joint Test of Residual Contrasts")
Joint Test of Residual Contrasts
df1 df2 F.ratio p.value
2 796 4.016 0.0183995

We start by performing “joint test” of the residual hypotheses (Note: c1 is the focal hypothesis so we only include c2 and c3). Now, we observed that our residual contrasts actually account for a significant portion of the variance.

So, we can perform equivalence testing for these contrasts. We find that c2 and c3 are within our equivalence bounds.

# Make comparisons with contrast function
knitr::kable(test(contrast(emm_four,
                           con_four),
                  delta = .4,
                  side = "equivalence"),
             caption = "4-group equivalence contrasts")
4-group equivalence contrasts
contrast estimate SE df t.ratio p.value
c1 -0.6831340 0.0825167 796 3.431233 0.9996841
c2 0.2300610 0.0875222 796 -1.941668 0.0262649
c3 0.0535141 0.0505310 796 -6.856903 0.0000000

Then, we can finally test focal hypothesis c1.

knitr::kable(tidy(contrast(emm_four,con_four[1])),
             caption = "4-group: c1 contrast")
4-group: c1 contrast
term contrast null.value estimate std.error df statistic p.value
group c1 0 -0.683134 0.0825167 796 -8.278736 0

From these results Patrick concludes the following (and I agree).

In this case, although the deviation from the focal contrast is significantly different from zero, we can say that the amount of deviation is smaller than our threshold for what constitutes a meaningful effect. If we think that any deviation from the focal contrast is reason for concern, we might think about amending our theory. If, on the other hand, we believe that the null hypothesis of no relationship is basically always false and that many of these non-zero relationships are “crud”, we might advance the claim that our theory has been corroborated. In either case our testing procedure is more severe than a mere test of the null hypothesis of no differences between group means.

Another way of saying this is the following: despite the data being incompatibility with the focal contrast (evidenced by the joint test), the differences observed in these other contrasts is small or within the established equivalence bounds. Meanwhile, we can reject the null hypothesis that our focal contrast is zero.

Extension to Multilevel/Hierarchical Models

In many cases, the designs that Patrick laid out are as simple or clean as a 3-group one-way ANOVA. Very often we have multiple levels of variance we would like to take into account. The most common example, and one I will repeat here, is in education wherein we have students inside classes within schools. In addition, we often have covariates that (like gender, socioecononimc status, or age) and we need to include those in our models.

So, I will apply Patrick’s approach to a study where we have test scores on students within classes within schools. We also have informative covariates such as age and gender on the test scores. Let us say we are testing the hypothesis that our intended treatments (groups 2 and 3) will have a positive effect, and there will be an additional benefit for treatment group 3. Therefore, we will have similar contrasts to the original 3-group example except we will reverse the coding and only have one-sided tests because I specified directional hypotheses.

library(simstudy)
# taken from https://kgoldfeld.github.io/simstudy/articles/clustered.html
gen.school <- defData(
  varname = "s0",
  dist = "normal",
  formula = 0,
  variance = 3,
  id = "idSchool"
)
gen.school <- defData(gen.school,
                      varname = "nClasses",
                      dist = "noZeroPoisson",
                      formula = 3)

set.seed(282721)

dtSchool <- genData(8, gen.school)
dtSchool <- trtAssign(dtSchool, nTrt = 3)

gen.class <-
  defDataAdd(
    varname = "c0",
    dist = "normal",
    formula = 0,
    variance = 2
  )
gen.class <-
  defDataAdd(gen.class,
             varname = "nStudents",
             dist = "noZeroPoisson",
             formula = 20)

dtClass <-
  genCluster(dtSchool,
             "idSchool",
             numIndsVar = "nClasses",
             level1ID = "idClass")
dtClass <- addColumns(gen.class, dtClass) %>%
  mutate(t2 = ifelse(trtGrp == 2, 1, 0),
         t3 = ifelse(trtGrp == 3, 1, 0))
  

gen.student <- defDataAdd(varname = "Male",
                          dist = "binary",
                          formula = 0.5)
gen.student <-
  defDataAdd(gen.student,
             varname = "age",
             dist = "uniform",
             formula = "9.5; 10.5")
gen.student <-
  defDataAdd(
    gen.student,
    varname = "test",
    dist = "normal",
    formula = "50 - 2*Male + s0 + c0 + 4 * t2 + 12 * t3",
    variance = 2
  )
dtStudent <-
  genCluster(
    dtClass,
    cLevelVar = "idClass",
    numIndsVar = "nStudents",
    level1ID = "idChild"
  )

con_three = list(c1 = c(-1,.5,.5),
                 c2 = c(0,1,-1))
check_orthog(cbind(con_three$c1,
                   con_three$c2))
[1] TRUE
dat_mlm <- addColumns(gen.student, dtStudent) %>%
  select(-s0,-c0,-t2,-t3) %>%
  mutate(trtGrp = factor(trtGrp,
                         levels = c(1,2,3),
                         ordered = TRUE))

Now we can utilize the lme4 package to build the mlm and use emmeans plot function to take a glimpse at the difference between groups.

test_mlm = lme4::lmer(test ~ Male + trtGrp + (1|idClass:idSchool) + (1|idSchool),
                      data = dat_mlm)
knitr::kable(broom.mixed::tidy(test_mlm),
             caption = "Summary Table of MLM")
Summary Table of MLM
effect group term estimate std.error statistic
fixed NA (Intercept) 55.9993351 0.5561662 100.688138
fixed NA Male -1.7990154 0.1491026 -12.065621
fixed NA trtGrp.L 7.6402443 0.7468733 10.229640
fixed NA trtGrp.Q 2.0859008 1.1211914 1.860432
ran_pars idClass:idSchool sd__(Intercept) 1.4617520 NA NA
ran_pars idSchool sd__(Intercept) 0.9791225 NA NA
ran_pars Residual sd__Observation 1.5290580 NA NA
# Plot estimated differences between treatments
emm_mlm = emmeans(test_mlm, ~ trtGrp)
plot(emm_mlm)

Finally, we can apply emmeans to look at the our specific contrasts. For these multi-level models emmeans defaults to Kenword-Roger degrees of freedom.

con_mlm = list(c1 = c(-1,.5,.5),
                 c2 = c(0,-1,1))
check_orthog(cbind(con_mlm$c1,
                   con_mlm$c2))
[1] TRUE
knitr::kable(test(contrast(emm_mlm, con_mlm),
                  side = ">"))
contrast estimate SE df t.ratio p.value
c1 6.826355 1.030733 5.934995 6.622817 0.0002986
c2 7.957165 1.489940 10.030437 5.340594 0.0001623

We see that our data is incompatible with the null hypotheses. We may also want to include “conditional equivalence tests” to ensure that our effects are not within a range we deem practically equivalent. Therefore, we can use almost the same code as the chunk above but add a delta argument as well as change the side argument to “equivalent”.

knitr::kable(test(contrast(emm_mlm, con_mlm),
                  delta = 2,
                  side = "equivalent"))
contrast estimate SE df t.ratio p.value
c1 6.826355 1.030733 5.934995 4.682450 0.9982574
c2 7.957165 1.489940 10.030437 3.998258 0.9987449

Conclusion

As you can see the process of creating specific contrasts is fairly straightforward in R and the hypothesis testing procedures are simplified by using the emmeans package. I find contrast coding to be a refreshing alternative to the typical inspection of “ANOVA-level” effects that is often followed up pairwise comparisions between the levels of a factor where there is a significant effect. Instead, contrast coding demands that the user be specific in their hypotheses. As Patrick notes in his blog post, specifying these contrasts a priori in many cases may result in a more severe tests of your hypotheses which arguably increases the strength of your claims if your experiments support your hypotheses. In my opinion there is added advantage in how you can go about describign your results (no need for the mundane langauge about “significant main effects”). Also, orthogonal contrasts, by nature, do not require adjustments for multiplicity and therefore may be more statistical powerful than default ANOVA tests. I’ll admit that orthogonal contrasts are not a silver bullet (there is no free lunch in statistics), but I do believe there many experiments that would benefit from this type of analysis.

Miscellaneous Notes

There are other ways to code contrasts. This post by Rose Maier notes a way of performing contrasts with the summary function. https://rstudio-pubs-static.s3.amazonaws.com/65059_586f394d8eb84f84b1baaf56ffb6b47f.html