Evidence Factors For Matched Triples With Two Control Groups
In an observational complete block design, with bocks of size three, each containing a treated individual and one control from each of two control groups, ef2C performs the evidence factor analysis suggested in Rosenbaum (2023). One factor compares the treated group to the first control group in a matched pairs analysis. The other factor pools the treated group and the first control group and compares it to the second control group.
ef2C(y, gamma = 1, upsilon = 1, alternative = "greater", trunc = 0.2)
y: With I blocks and 3 individuals in each block, y is and I x 3 matrix or dataframe containing the outcomes. The first column is the response of the treated individual. The second response is the response of the control from the first control group. The third response is the response of the control from the second control group.gamma: A real number >=1 giving the value of the sensitivity parameter for the comparison of the treated group and the first control group. gamma=1 yields a randomization test.upsilon: A real number >=1 giving the value of the sensitivity parameter for the comparison of the second control group and the combination of the treated group plus the first control group. upsilon=1 yields a randomization test.alternative: Use alternative=greater if the treatment is expect to cause an increase in the response in y. Use alternative=less if the treatment is expect to cause an decrease in the response in y. In this context, a two-sided test is best viewed as two one-sided tests with a Bonferroni correction, e.g., testing in both tails at level 0.025 to ensure overall level of 0.05; see Cox (1977). For more information, see the notes.trunc: The two P-values from the two factors are combined using the trucated product of P-values due to Zaykin et al. (2002): it is the P-value derived from the product of those P-values that are less than trunc. For more information, see the notes.This is the recommended analysis in Rosenbaum (2023). The example below reproduces some results from the example in that paper. That paper considered 40 test statistics in terms of the Bahadur efficiency of a sensitivity -- all of these analyses can be reproduced by the more flexible but more complicated dwgtRank function.
The comparison of the treated group and the first control group is equivalent to dwgtRank(y[,1:2],gamma=gamma,m=8,m1=7,m2=8,range=TRUE,alternative="greater"), and these settings are motivated by results in Rosenbaum (2011, 2015). Notice that y[,1:2] uses the first two columns of y.
The comparison of the second control group and the merger of the treated group with the first control group is equivalent to dwgtRank(y[,3:1], gamma=upsilon, m=8,m1=8,m2=8, range=FALSE, alternative="less", scores=c(1,2,5)), and these settings are motivated by results in Rosenbaum (2023). Notice that y[,3:1] compares the third column to the pooled group consisting of columns 1 and 2.
pvals: Upper bounds on the one-sided P-values for the two factors and their combination.
detail: A matrix with some details of the computations that produced the P-values.
The two P-values from the two factors are combined using the trucated product of P-values due to Zaykin et al. (2002): it is the P-value derived from the product of those P-values that are less than trunc. Taking trunc=1 yields Fisher's method for combining independent P-values. Fisher's method is not ideal when combining P-value bounds produced by sensitivity analyses; see Hsu et al. (2013). Reasonable values are trunc=.1, truc=.15 and trunc=.2. As illustrated in the example below, lower truncation values produce smaller combined P-values when the P-values are below the truncation point, but a P-value that barely exceeds the truncation point is effectively discarded. Hsu et al. (2013) compare truncation values when used in a sensitivity analysis. For discussion of combining sensitivity analyses as independent, see the required conditions in Rosenbaum (2011b, 2021). These conditions hold for the comparison performed by ef2C.
The setting alternative = "less" simply replaces y by -y before testing in the upper tail.
For a deeper understanding, see the documentation of dwgtRank. That function is more general, but it requires more attention to detail by the user. The documentation for dwgtRank also produces additional analyses from Rosenbaum (2023).
Cox, D. R. (1977). The role of significance tests [with discussion and reply]. Scandinavian Journal of Statistics, 4, 49-70.
Hsu, J. Y., Small, D. S., & Rosenbaum, P. R. (2013). doi:10.1080/01621459.2012.742018 Effect modification and design sensitivity in observational studies. Journal of the American Statistical Association, 108(501), 135-148.
Rosenbaum, P. R. (1987). doi:10.1214/ss/1177013232 The role of a second control group in an observational study. Statistical Science, 2, 292-306.
Rosenbaum, P. R. (2011a). doi:10.1111/j.1541-0420.2010.01535.x A new U‐Statistic with superior design sensitivity in matched observational studies. Biometrics, 67(3), 1017-1027.
Rosenbaum, P. R. (2011). doi:10.1198/jasa.2011.tm10422 Some approximate evidence factors in observational studies. Journal of the American Statistical Association, 106(493), 285-295.
Rosenbaum, P. R. (2015). doi:10.1080/01621459.2014.960968 Bahadur efficiency of sensitivity analyses in observational studies. Journal of the American Statistical Association, 110(509), 205-217.
Rosenbaum, P. R. (2021) doi:10.1201/9781003039648 Replication and Evidence Factors in Observational Studies. Chapman and Hall/CRC.
Rosenbaum, P. R. (2023) A second evidence factor for a second control group. Manuscript. This paper develops a new statistic that performs well when there are two control groups and the second control group forms a second evidence factor. See also Rosenbaum (2011b, 2021).
Zaykin, D. V., Zhivotovsky, L. A., Westfall, P. H. and Weir, B. S. (2002) doi:10.1002/gepi.0042 Truncated product method of combining P-values. Genetic Epidemiology, 22, 170-185.
Paul R. Rosenbaum
# The calculation below reproduce analyses from Rosenbaum (2023). data(aBP) attach(aBP) yD<-t(matrix(bpDiastolic,3,207)) yS<-t(matrix(bpSystolic,3,207)) vS<-c(yS[,1]-yS[,2],yS[,1]-yS[,3],yS[,2]-yS[,3]) vD<-c(yD[,1]-yD[,2],yD[,1]-yD[,3],yD[,2]-yD[,3]) y<-(yD/median(abs(vD)))+(yS/median(abs(vS))) # EVIDENCE FACTOR ANALYSIS, COMBINING TWO FACTORS # The evidence factor analysis compares treated to the first control group, # then compares the second control group to the pooled group consisting of # treated and first control, then combines the two analyses using meta-analysis. # Treated/first-control matched pairs are compared using the method in # Rosenbaum (2011). ef2C(y,gamma=2.3,upsilon=1.45) amplify(2.3,4) amplify(1.45,2.5) # THE COMBINED ANALYSIS IS INSENSITIVE TO LARGER BIASES # The combined analysis is insensitive to larger biases # than are its components ef2C(y,gamma=2.6,upsilon=1.7) amplify(2.6, 5) amplify(1.7,c(2.7,3)) # The calculations above are also produced in the # example for dwgtRank, where alternative # analyses from Rosenbaum (2023) are compared. #################################################### # Comparing trucation points to understand trunc: ef2C(y,gamma=2.6,upsilon=1.7,trunc=.2) # Default ef2C(y,gamma=2.6,upsilon=1.7,trunc=1) # Fisher's method ef2C(y,gamma=2.6,upsilon=1.7,trunc=.1) ef2C(y,gamma=2.5,upsilon=1.6,trunc=.2) ef2C(y,gamma=2.5,upsilon=1.6,trunc=.1) # See Hsu et al. (2013) for discussion of the # truncation point for a sensitivity analysis.
Useful links