Hello,
I am working with maize drought tolerance. I am analyzing data under drought stress and normal condition. While doing data analyses I have found that traits like Anthesis silking interval, thousand kernel weight has a heritability of zero under drought stress condition. But when I combine data of both drought stress and normal condition heritability of those two traits is not zero. Is there any problem with my data? or is it explainable in other ways?
Looking forward to answers..
It is not strange to have an h2 of zero. This is common in some traits, particularly when there is no genetic control. However, the other reason is that you have too much residual variances, which is common in drought studies, as they tend to be quite noisy. Also, an h2 of zero, when is not expected can be a sign of serious issues in your data. I have seen that when people have made errors in measurements (or in sorting data).
Now, the reason why you get a heritability when you combine the data with the normal condition is because you are effectively averaging both heritabilities.: the one for dorught and the one for normal, and because the latter is non-zero you have a non-zero average. This is just an indication that you can not just blind report heritability.
So would you please suggest what can I do to avoid that? Because I do not expect a heritability of zero for an important trait like anthesis silking interval. How can I reduce residual error variance? Also I get heritability of around 0.04-0.1 for yield and other traits. So, how can I minimize this problem? Please suggest.
First, tell me about your experiment...there are always ways to improve your analysis by better statistical models...
I have three locations and 5 environment. In location 1 and 2 I have drought and irrigated condition and in location 3 I have only drought. So I have analyses each environment separately. Each location has two replications. Then I took the mean values from each single analysis to perform a combine anova. This is basically a two stage analysis. There are 48 lines and 2 checks. For the single analysis I used rep+genotype+rep*genotype model. Moreover in location 1 the design was alpha lattice. And in other two it was rcbd design. Now in the analysis of lattice design for each environment I have this heritability of quite low for drought condition and that zero heritability as well. In the combined anova I used the model loc+env+geno+loc*env+geno*loc+geno*env+loc*env*geno. I have good heritability and correlation in this analysis. Again I performed combined anova only for drought condition. There I also have found the heritability for ASI and thousand kernel weight Zero. So this is the situation. So what can be done now?
I am doing the analysis in plabstat. After the analysis it is showing a considerable number of outliers. That is also an issue
Very low heritability is entirely possible, and when that is the case point estimates can easily be low or even negative, as a result of random estimation errors. Apart from resulting from such errors, zero or negative estimates can result from data properties that include various types of departure from normality. One would want to characterize the distributional properties of the data and, better still, explain them. While appropriate data transformations may help, as may specifying heterogeneous variance structures in the fitted model, there is no guarantee that such measures may help. Assuming inappropriate coefficients of relationship may bias heritability estimates, but would not be expected to generate zero or negative estimates, except perhaps with some complex pedigree structures. Violation of the assumption of independence of variation at the lowest level of the hierarchy in the classification can generate negative estimates. With forest trees, using multi-tree plots (often no longer recommended), competitive interactions for stem diameter, with larger trees suppressing growth in smaller ones, can generate spurious negative heritability estimates through inflating tree-to-tree environmental variance, but you may not be getting any such effect in your case. Incidentally, even when one gets a competitive interaction variance of a sort that may generate bias, it might still be possible to estimate it and correct for it, but the algothms entailed would not be straightforward.
I can add one more thought. If you have a negative genetic correlation between performance in two environments (as might occur with droght versus non-drought) one could well get effectively a zero heritability for average performance over the pair of environments. Note that between two environments the true genetic correlation has bounds of +1 and -1, whereas the genetic correlation across multiple environments has bounds of +1 and zero
another point is if the genetic correlation is very low, zero, or negative, you can not average the two values as the environments are different. also as it was indicated above the masking variance is very high in drought conditions
You can analyse your data using Meta-R software, there you can get heritability location wise as well as combined across the location. Also you can get correlation among the locations and correlation of traits across the locations.
Look for the variations among the replications in each locations.
Since the traits are quantitative in nature, you can expect the low heritability but not as low as zero.
Data quality check will be important as suggested by many: influential data points (high residual error), outliers... Check the yield reduction under stress and you may find that some data points are higher under stress environment compared to their counterparts under a well-watered environment (this is unlikely unless there is an issue in the well-watered environment). You may also try fit spatial models in your analyses to remove row, column and local variation effects.
- Badges
- Science topic
- Similar topics
- Philosophy
- Esthetics
- Interpretation