ANTIPERSPIRANT RESULTS 189 Table II (continued) Test standard deviations Slope Study BASLIN POSTRT CHGBAS ANCOVA estimate 802 0.103 0.127 0.103 0.099 0.78 803 0.092 0.099 0.096 0.087 0.52 804 0.123 0. ,143 0.147 0.130 0.52 805 0.101 0.143 0.157 0.142 0.26 806 0.131 0.225 0.205 0.205 0.75 807 0.137 0.158 0.161 0.145 0.49 809 0.159 0.228 0.198 0.198 0.74 811 0.128 0.201 0.183 0.183 0.69 812 0. 143 0.204 0.202 0. 192 0.52 813 0.096 0.171 0.164 0.155 0.77 901 0.085 0.088 0.096 0.082 0.40 902 0.105 0. 133 0.128 0.121 0.56 903 0.093 0.131 0.13 ! 0.124 0.49 904 0.116 0.10! 0.131 0.098 0.24 905 0.092 0.128 0.113 0.108 0.76 906 0.108 0.129 0. ! 13 0.107 0.68 907 0.071 0.104 0.089 0.093 0.69 908 0.099 0.123 0.109 0. 105 0.66 909 0.127 0.115 0.107 0.092 0.55 910 0.113 0.105 0.117 0.094 0.42 of 0.55. This demonstrates the extent of overcorrection resulting from the CHGBAS method, which has a fixed slope of unity. On average, the ANCOVA correction for baseline is slightly more than half the CHGBAS correction. Figures 13-15 show the differences in treatment-effect estimates over studies. For POSTRT vs CHGBAS, the middle 50% of the comparisons were about +0.020 or about +5% reduction. For ANCOVA vs POSTRT and ANCOVA vs CHGBAS, the middle 50% comparisons were about -+0.010 or about +2.5% reduction. The ANCOVA estimate tended to be the average of the POSTRT and CHGBAS estimates. Because of this wide variation in estimates, it is best to present to the client only the ANCOVA analysis. With,_. ratio 0 = O'b/O' p estimated at 0.76, the critical value of the correlation coefficient Pbp to attain variance reduction via simple change from baseline would be 0/2 = 0.38. The average correlation coefficient estimated from the data was 0.42, or slightly above 0.•0 1.02 1.14 1.26 1.•8 o lO 20 FREQUENCY FREQ. 1 22 22 3O Figure 10. Standard deviation (CHGBAS)/standard deviation (POSTRT).

190 JOURNAL OF THE SOCIETY OF COSMETIC CHEMISTS 1.40 o lO 2o FREQUENCY 26 15 1 1 Figure 11. Standard deviation (POSTRT)/standard deviation (ANCOVA). the critical value. Thus, in the five-day protocol, the correlation between baseline and post-treatment readings is such that little improvement in precision can be expected from the simple change from baseline adjustment. The validity of ANCOVA depends on the assumption of equality of slopes across groups. This was empirically supported by statistical tests for interaction of slopes and groups for 33 studies, where only one p-value was observed below 0.05. In conclusion, the three methods of analysis are mutually unbiased, with ANCOVA as the most precise. Hence ANCOVA is the method of choice for the analysis of antiper- spirant efficacy studies, requiring a protocol that provides for baseline information. MULTIPLE WITHIN-DAY SWEAT MEASUREMENTS Historically, two 20-minute sweat collections are taken in a single hot-room session following the 40-minute warmup period. This procedure generates two responses of sweat rate per axilla in a given test day. Recommendations as to the treatment of "repeated measures" data vary among investigators (4,6, 10). Some recommend that the two responses be analyzed separately, others discard one of the measurements, and still others average the two measurements and analyze the averages. Our practice is to average the two readings after log transformation, which would give the geometric mean upon conversion back to the original metric. Since the B and C sweat collections are measured close together in time, they will be ratio 0.9• 1.05 1.11 1.17 1.23 1.2• 1.35 1.41 FREQUENCY Figure 12. Standard deviation (CHGBAS)/standard deviation (ANCOVA).