(Eqn. 1)
Where L E and L I is leaf area at t E - and t I,respectively.
The difference t Et Ireflects the 84 days between the start of the treatment phase (t I) and the day of the final harvest (t E).
Drought tolerance was defined as the capacity of a genotype to maintain its growth under drought stress (restricted-water) and was computed per genotype as the ratio of the trait mean in restricted-water to the trait mean in ample-water across blocks.
Data analysis
Linear mixed effect models were applied to test the effects of water treatment on selected growth traits across (i) cultivation status, (ii) genetic group or (iii) location. Linear mixed-effect models were used because mixed models account for unbalanced, nested designs (such as varying numbers of genotypes by cultivation status, genetic groups and location) that occurred in our data (Bates et al. , 2015) . To estimate the impact of water shortage on plant traits across cultivation status, genetic groups and locations, genotypes were considered a random effect both in terms of the intercept: i.e ., the absolute trait value in ample-water, and the slope: i.e ., the response to drought (difference between the trait value in ample and restricted-water conditions). To account for the heterogeneity of variance in the observations, variances in the traits were dependent on the cultivation status, genetic group or location (Zuur, Ieno and Elphick, 2010) .
The model with cultivation status had 12 parameters: three levels of cultivation status (CS) and two water treatments (making six parameters), three parameters of the random effect to model differences across genotypes: (i) a parameter to model the variation of traits in ample-water conditions (intercept), (ii) a parameter related to the variation in the treatment effect (slope), and (iii) a parameter that models the correlation between the intercept and the slope, and three parameters to account for a different residual variance per cultivation status (see Model 1 in Appendix Data Box A.1.). The model with the genetic group had 18 parameters: two for each genetic group (making 10) and three parameters of the random effect to model differences across genotypes: (i) a parameter to model the variation of traits in ample-water conditions (intercept), (ii) a parameter related to the variation in the effect of the treatment effect (slope), and (iii) a parameter that models the correlation between the intercept and the slope, and five parameters to account for a different residual variance per genetic group (see Model 2 in Appendix Data Box A.1.). Note that while testing the genetic group effect, all genotypes that were misclassified and/or hybrids were not considered.
Factor location analysis tests for differences in terms of the environment but also for genetic basis, and therefore, indirectly for putative local adaptation. Therefore, the model with location had in total 24 parameters, two for each location (14) and three parameters of the random effect to model differences across genotypes: a parameter to model the variation of traits in control (intercept), a parameter related to the variation in the treatment effect (slope), a parameter that models the correlation between the intercept and the slope, and seven parameters to account for a different residual variance per location (see Model 3 in Appendix Data Box A.1.).
Post-hoc Tukey tests were performed to determine whether: (i) drought had significant effects on performance (RGRA) of genotypes across cultivation status, genetic groups and location; (ii) genotypes of different cultivation status, genetic groups or location responded significantly differently to drought and (iii) absolute performance of genotypes in ample-water and restricted-water conditions differed across cultivation status, genetic groups and locations. Tukey adjustment to p-values was done in case of multiple comparisons. The analyses were performed in R version 3.5.0 Statistical Software using “me”, “emmeans” and “ggplot2” packages. For all the analyses, any effect with p < 0.05 was considered statistically significant and non-significant at p > 0.05.

Multivariate analysis of growth-related traits

To explore the multivariate dependency between the measured traits, a principal component analysis was performed on the genotypic means. Only genotypes were included for which there were at least two replicates available. All variables were centred to a mean of zero and scaled to unit variance before the analysis contained both treatments. Next, to test whether location and cultivation significantly affected the suite of traits, a multivariate analysis was performed using a PERMANOVA. This PERMANOVA tests, similar to a classical multivariate ANOVA, whether the dissimilarities between genotypes from the same location, status and treatment are smaller than the dissimilarities between genotypes across locations, status and treatment (Anderson, 2001). We used Euclidean distances between the centred and scaled observations, and 999 permutations.
Drought tolerance performance trade-off
Type (II) major axis regression was performed to determine the relationship between the genotypic average growth trait in ample water versus restricted water. Type II regression was used to account for both measurement errors in the independent and the dependent variable (David and Neville, 2002) and to test whether the slope and intercept were different from each other. The analysis was performed in R version 3.5.0 Statistical Software using the packages “smatr”.
Drought tolerance climate relationship
In addition, a weighted linear regression analysis was performed to determine the relationship between drought tolerance based on RGRA, TNL and TLDW and wetness index (WI). The analysis was performed in R version 3.5.0 Statistical Software. Because the number of replicates varied across genotypes in locations, we introduced weights for replicates in the analysis. In this weighted linear regression analysis, we excluded genotypes from Kawanda and Kituza because the genotypes in these collections were sourced from different origins and assembled asex-situ collections at NARO institutes, and therefore, we could not retrieve the WI of these genotypes. The probability of rejecting the null hypothesis that there is no relationship between performance in low-water conditions and wetness index (WI) or no relationship between drought tolerance and wetness index (WI) was set at p-value > 0.05. The weighted linear regression models were fitted with lm () functions in R version 3.5.0 Statistical Software.
RESULTS
The study results are presented in hierarchical order starting with: (i) the effect of experimental treatments on the grand mean of growth response traits (i.e., lumping genotypes together), (ii) the main effects of factors, i.e. cultivation status, genetic groups and location on growth response traits, (iii) the detailed synthesis of the effect of drought on RGRA as our proxy trait for plant performance, (iv) the relationship between performance under ample and restricted-water conditions, (v) the relationship between performance under restricted-water conditions and wetness index of the locations and (vi) the relationship between drought tolerance and wetness index of the locations.