DMAR (version 1.0.0)

test_market: Controlled Test-Market Experiment (Bryant & Bruvold, 1980)

Description

The controlled test-market experiment of Bryant and Bruvold (1980), used to illustrate multiple-comparison procedures in the analysis of covariance (ANCOVA) when the covariate is random. A company compared \(k = 6\) marketing strategies (“panels”) for a brand, randomly assigning them to retail outlets within \(s = 4\) blocks of outlets that were homogeneous in size, locality, and ownership (a randomized complete block design, one outlet per panel-by-block cell). During the experiment a concomitant variable, the remaining category movement in each outlet, becomes available; it cannot be controlled by the experimenter and is best modeled as a random covariate. Adjusting brand movement for this covariate sharply reduces unexplained error, permitting far finer comparison of the panels than the raw outcome allows.

Usage

test_market

Format

A data frame with 24 observations (6 panels \(\times\) 4 blocks) on 4 variables.

panel

Factor with levels 1--6: the marketing strategy (treatment) randomly assigned to the outlet. Different panels entail different methods of packaging, displaying, or pricing.

block

Factor with levels 1--4: the block of retail outlets, grouped to be homogeneous in size, locality, ownership, and other considerations that influence brand movement.

brand_movement

Test-brand movement during the test period, in hundreds of statistical cases. The dependent variable (\(y\) in the source).

category_movement

Remaining category movement, the random concomitant variable (covariate; \(x\) in the source). It is not identically distributed across blocks, which is precisely the setting Bryant and Bruvold's grouped-covariate extension was designed for.

Author

Ken Kelley

Details

Why it is a benchmark for ANCOVA multiple comparisons. The model fitted by Bryant and Bruvold (their Eq. 3.1) is a randomized-block ANCOVA, $$y_{ij} = \theta_i + \beta_j + (x_{ij} - \delta_j)\, u + e_{ij},$$ with \(\theta_i\) the \(i\)th panel (adjusted) mean, \(\beta_j\) the \(j\)th block effect, and \(u\) the within-cell covariate slope. The point of the example is that the studentized range of the adjusted panel means does not follow the ordinary Tukey distribution, because the covariate is random and its adjustment must be estimated; the correct reference distribution is the Bryant--Paulson generalized studentized range (bryant_paulson).

Reproducible quantities. Fitting lm(brand_movement ~ panel + block + category_movement) gives a covariate slope of \(0.4079\) and an error mean square of \(0.01326\) on \(\nu = 14\) degrees of freedom, with adjusted panel means \(3.595, 3.619, 4.102, 4.515, 4.618, 4.876\), exactly the values reported in the paper. With \(q_{.05;\,1,6,14} = 4.83\) (qbryant_paulson), every pairwise simultaneous 95% interval is a difference of adjusted panel means plus or minus \(0.278\), so two panels differ at the simultaneous 95% level exactly when their adjusted means are more than \(0.278\) apart. Had the covariate not been measured, the error mean square would have been \(0.2368\), roughly eighteen times larger, and the intervals about four times wider. See data-raw/test_market.R for the construction script and its verification checks.

References

Bryant, J. L., & Paulson, A. S. (1976). An extension of Tukey's method of multiple comparisons to experimental designs with random concomitant variables. Biometrika, 63, 631--638.

Maxwell, S. E., Delaney, H. D., & Kelley, K. (2027). Designing experiments and analyzing data: A model comparison perspective (4th ed.). Routledge. (See Chapter 9.)

See Also

ci_c_ancova_bp for the simultaneous intervals this data set illustrates, bryant_paulson for the critical values, and ancova for an ANCOVA fit.

Examples

data(test_market)
str(test_market)

# Reproduce the published ANCOVA (slope 0.4079, error MS 0.01326, df 14).
fit <- lm(brand_movement ~ panel + block + category_movement,
          data = test_market)
coef(fit)["category_movement"]
sum(residuals(fit)^2) / fit$df.residual

# Adjusted panel means at the covariate grand mean.
xbar <- mean(test_market$category_movement)
adj <- vapply(levels(test_market$panel), function(p) {
  nd <- data.frame(panel = factor(p, levels = levels(test_market$panel)),
                   block = factor(1:4, levels = levels(test_market$block)),
                   category_movement = xbar)
  mean(predict(fit, nd))
}, numeric(1))
adj  # 3.595 3.619 4.102 4.515 4.618 4.876

# Bryant-Paulson simultaneous 95% intervals (s = 4 blocks => n = 4,
# df 14), every one of them the difference plus or minus 0.278.
ci_c_ancova_bp(adj_means = adj, s_ancova = sqrt(0.01326),
               n = 4, num_covariates = 1, df = 14)