This shows you the differences between two versions of the page.
Both sides previous revision Previous revision Next revision | Previous revision | ||
workshop:2016:questions [2016/03/11 15:04] 65.114.233.215 |
workshop:2016:questions [2016/03/11 16:31] 65.114.233.215 |
||
---|---|---|---|
Line 2: | Line 2: | ||
<note warning> | <note warning> | ||
- | [[http://goo.gl/forms/ | + | Ask questions at the [[http://openmx.psyc.virginia.edu/forums|OpenMx forums]] and the [[http:// |
</ | </ | ||
Line 24: | Line 24: | ||
??? 2. In her folder from Monday, Hermine Maes provided many different scripts, which is absolutely great! Is is possible to give some explanation to these scripts and in what they differ (except of ACE, ADE & SAT which is quite clear)? | ??? 2. In her folder from Monday, Hermine Maes provided many different scripts, which is absolutely great! Is is possible to give some explanation to these scripts and in what they differ (except of ACE, ADE & SAT which is quite clear)? | ||
!!! MN: Hermine' | !!! MN: Hermine' | ||
+ | |||
+ | Also, Hermine' | ||
+ | |||
+ | 1a- Analysis for 2 zygosity groups without any covariates and with an age covariate [a] | ||
+ | 1b- Analysis for 5 groups without any covariates and with age covariate [a] | ||
+ | |||
+ | 2- Each set includes a saturated model (SAT with sub models testing assumptions), | ||
+ | |||
+ | 3- Analyses for specific types of variables: continuous (c), binary (b), ordinal (o), mean/ | ||
+ | |||
+ | Two examples of the naming convention: | ||
+ | 1- The file called " | ||
+ | 2- The file called " | ||
??? 3. In classic Mx, we were able to include twin pairs with missing data. Can you clarify that we need to either drop twin pairs with any missing data or replace missing values with the mean? Are we not able to define a missing value and include that individual in the model? | ??? 3. In classic Mx, we were able to include twin pairs with missing data. Can you clarify that we need to either drop twin pairs with any missing data or replace missing values with the mean? Are we not able to define a missing value and include that individual in the model? | ||
Line 67: | Line 80: | ||
!!! For background, see [[http:// | !!! For background, see [[http:// | ||
- | Rob K. says: I would tentatively answer your second question with a " | + | Rob K. says: I would tentatively answer your second question with a " |
??? 11. If we get a code Mx status RED, what do we need to do/consider (in general)? E.g., are our model estimates still reliable? | ??? 11. If we get a code Mx status RED, what do we need to do/consider (in general)? E.g., are our model estimates still reliable? | ||
- | !!! Rob K. says: Status RED means the optimizer is not certain it has found a minimum of the fitfunction. | + | !!! Rob K. says: Status RED means the optimizer is not certain it has found a minimum of the fitfunction. So, no, your parameter estimates are probably not reliable, and the standard errors are even more suspect. Status RED with code 6 (first-order conditions not met) is worse than with code 5 (second-order conditions not met). You should always try to do // |
- | * If you're analyzing ordinal data, sometimes a status RED is unavoidable without changing some [[http:// | + | * If you're analyzing ordinal data, sometimes a status RED is unavoidable without changing some [[http:// |
* Use different start values. | * Use different start values. | ||
* Try a different optimizer. | * Try a different optimizer. | ||
* Reparameterize your MxModel. | * Reparameterize your MxModel. | ||
* Use [[http:// | * Use [[http:// | ||
+ | |||
+ | Trying different start values is the most important thing. | ||
??? 12. For Ben's presentation, | ??? 12. For Ben's presentation, | ||
- | !!! MCK: I'll let Ben weigh in as well. But in essence, they should be two different ways of estimating the *same* parameter: SNP-heritability. I think that the differences in the literature are about what we'd expect given the SE' | + | !!! Ben here - Here's the original LD Score MS: http://www.nature.com/ |
- | Ben here - Here's the original LD Score MS: http://www.nature.com/ng/journal/v47/n3/full/ng | + | ???13. What is the difference in interpretation between GCTA h2 estimates and LD score regression h2 estimates? |
+ | !!!MCK: I'll let Ben weigh in as well. But in essence, they should be two different ways of estimating the *same* parameter: SNP-heritability. I think that the differences in the literature are about what we'd expect given the SE's on the estimates. Ben - are there any systematic differences between the two? | ||
+ | |||
+ | ??? 14. @Sarah where can we find the full syntax of this morning' | ||
+ | !!! | ||
+ | |||
+ | ??? 15. Where can researchers find publicly available twin data? | ||
+ | !!! One place a researcher can begin searching for publicly available data is the repository developed and managed by the Inter-university Consortium for Political and Social Research (ICPSR). | ||
+ | |||
+ | As an aside, another publicly available resource towards developing harmonized measures of biomedical phenotypes is the PhenX toolbox (https:// | ||
+ | |||
+ | |||
+ | MCK: Great question! Nick, Dorret, John?? Want to weight in? | ||
+ | Sadly (I think), twin research has not kept up with whole-genome research, where sharing of data is not only the norm, but also mandatory (for NIH funding at least). Would be scientifically useful if the same were true of twin research! | ||
+ | |||
+ | |||
+ | |||
+ | ====== Questions/ | ||
+ | |||
+ | ??? 1. Coming from using more of a path analysis software, I am having a hard time understanding where the R/OpenMx script indicates the relationships between latent ACE and the observed indicators (family, happy etc). Specifically this line of script | ||
+ | |||
+ | !!! There are several ways that expected means, thresholds and covariance matrices can be generated. | ||
+ | |||
+ | Note that it is also possible to specify models using mxPath() function calls, which may be more familiar if you are accustomed to specifying models that way in other software. | ||
+ | |||
+ | TCB: Good question! This is the code snippet you posted: | ||
+ | <code rsplus> | ||
+ | modelACE <- mxModel( " | ||
+ | </ | ||
+ | |||
+ | There is a lot of stuff bundled up in there. | ||
+ | |||
+ | Let's unpack and simplify... We'll make a " | ||
+ | |||
+ | Then we put " | ||
+ | |||
+ | <file rsplus ACE.r> | ||
+ | ntv = 2 | ||
+ | nv = ntv/2 | ||
+ | myACE = mxModel(" | ||
+ | mxModel(" | ||
+ | mxMatrix(name = " | ||
+ | mxMatrix(name = " | ||
+ | mxMatrix(name = " | ||
+ | mxMatrix(name = " | ||
+ | mxAlgebra(name = " | ||
+ | mxAlgebra(name = " | ||
+ | mxAlgebra(name = " | ||
+ | mxAlgebra(name = " | ||
+ | mxAlgebra(name = " | ||
+ | mxAlgebra(name = " | ||
+ | mxAlgebra(rbind(cbind(ACE, | ||
+ | cbind(AC , ACE)), name = " | ||
+ | mxAlgebra(rbind(cbind(ACE, | ||
+ | cbind(hAC, ACE)), name = " | ||
+ | ), | ||
+ | mxModel(" | ||
+ | mxExpectationNormal(" | ||
+ | mxFitFunctionML(), | ||
+ | mxData(mzData, | ||
+ | ), | ||
+ | mxModel(" | ||
+ | mxExpectationNormal(" | ||
+ | mxFitFunctionML(), | ||
+ | mxData(dzData, | ||
+ | ), | ||
+ | mxFitFunctionMultigroup(c(" | ||
+ | ) | ||
+ | |||
+ | </ | ||
+ | |||
+ | Now we can answer your question: "where does the script indicate the relationships between latent ACE and the observed indicators (family, happy etc)?" | ||
+ | |||
+ | The answer is: "in mxExpectationNormal and mxFitFunctionML" | ||
+ | |||
+ | The supermodel looks at how well each of the expectations (MZ and DZ) fit, and ensures the sum of both is as good (small) as possible. | ||
+ | |||
+ | Does that help? | ||
+ | |||
+ | |||
+ | ??? 2. What to do with strongly skewed (non-normal) data? normalize or not? are log and square root transformations appropriate? | ||
+ | !!! The same answers hold here as elsewhere in stats: If your data are highly skewed, then it's likely that they are better expressed in a transform. | ||
+ | |||
+ | Reaction time for instance is skewed and long-tailed. The reciprocal or " | ||
+ | |||
+ | You should plot them, however. If the most common value is 0, then transforming won't normalize, and you might want an ordinal or tobit analysis instead. | ||
+ | |||
+ | - Another point of view on this (from Sarah)- | ||
+ | |||
+ | If you are going to transform, as a reviewer, I like to see the results of the untransformed data in an appendix/ | ||
+ | |||
+ | If after transformation your data does not have a single mode and the distribution is very asymmetric (ie symptom count type data in an unselected cohort) you should convert the raw data to an ordinal categorical variable. Remember that you do not really gain much power by adding extra categories if the prevalence of the extra categories is very low and you end up empty cells in your twin1/twin2 cross tab. | ||
+ | |||
+ | !!!Additional remarks (Conor): | ||
+ | 1) The distributional assumption in ML (as used here) is that the data are multivariate normally distributed. So if you are concerned about normality given the ML normality assumption, then you'll have to consider multivariate normality in addition to marginal (i.e., univariate) normality. | ||
+ | 2) Remember that the normality assumption concerns the distribution of the data conditional on fixed covariates (definition variables with a main effect on the phenotypes). So consider height in 500 males and 500 females (all 30 years old). The distribution of height in the complete sample (N=1000) may deviate from normality, while the distributions in the subsamples (males and females) may be perfectly normal. | ||
+ | |||
+ | !!!Rob K. remarks: If there is a generally accepted transformation for your kind of variable in your discipline, use that. If not, I advocate adjusting one's methods to suit one's data, not adjusting one's data to suit one's methods! | ||
+ | |||
+ | !!! Recode your data into ordinal categories, and use table(data$variable_twin1, | ||
+ | |||
+ | !!! Mike N: Try to understand //why// your data are non-normal. | ||
+ | |||
+ | ??? 3.How do epigenetic effects affect estimates of V(C), V(D)and V(A)? | ||
+ | !!! Conor: there are a number of interpretations. I think that such effects will render a (A path) potential random over individual. That is whereas we now think of a as a fixed parameter it may be a parameter that is ditributed, say, normally with mean mu(a) and standard deviation sd(a). Treating it as fixed, we get from the model mu(a). This is really just speculation. But for an elaborate presentation of this please see https:// | ||
+ | |||
+ | ??? 4. Could you please explain step by step how to solve for A and C given CVmz = A + C and CVdz = 1/2 A + C? | ||
+ | !!! Suppose that the MZ correlation is .8 and the DZ correlation is .5 | ||
+ | < | ||
+ | A + C = .8 | ||
+ | 1/2 A + C = .5 | ||
+ | </ | ||
+ | Then subtracting the DZ expression from the MZ, we have: | ||
+ | < | ||
+ | 1/2A = .3 | ||
+ | A = .6 | ||
+ | </ | ||
+ | substitute this back into either the MZ or the DZ equation so | ||
+ | < | ||
+ | .6 + C = .8 | ||
+ | C = .2 | ||
+ | </ | ||
+ | Given that these are correlations, | ||
+ | |||
+ | < | ||
+ | CVMZ = A + C | ||
+ | CVDZ = 1/2 A + C | ||
+ | |||
+ | So CVMZ-CVDZ = A + C - 1/2A - C = 1/2A. SO to get A: | ||
+ | A_hat = 2*(CVMZ-CVDZ). | ||
+ | |||
+ | Similarly to solve C, find a way to eliminate A and isolate C. E.g.: | ||
+ | 2CVDZ - CVMZ = A + 2C - A - C = C_hat | ||
+ | </ | ||
+ | |||
+ | |||
+ | |||
+ | ??? 5. Could you please explain why it is better to include age and sex as a covariate rather than using the residual in a model? | ||
+ | !!! One answer is that if the dependent variable is ordinal, you have no choice. | ||
+ | |||
+ | Another is that by incorporating the covariates in the model, you have the estimates in one place. | ||
+ | |||
+ | Rob K. says: Stated briefly, if you incorporate the covariates into the MxModel, you end up simultaneously // | ||
+ | |||
+ | |||
+ | ??? 6. Will we discuss methods that can be employed to identify/ | ||
+ | !!! Yes - Brad will be discussing this Friday. | ||
+ | |||
+ | ??? 7. Question about Power: How do you get from the individual family likelihood contribution to the power percent. We know it has to do with the chi-squared distribution but could you walk us through it? | ||
+ | !!! Not answered yet. | ||
+ | |||
+ | ??? 8. Is there a way of monitoring the progress of optimization while openMx is running? | ||
+ | !!! TCB: Kind of: if you set checkpointing, | ||
+ | |||
+ | You can do this with calls to options, or just use umx_set_checkpoint() | ||
+ | mxOption(NULL, | ||
+ | mxOption(NULL, | ||
+ | mxOption(NULL, | ||
+ | mxOption(NULL, | ||
+ | |||
+ | MN adds: Also, on a linux or OS X system you can use the command line function less to inspect the output. | ||
+ | |||
+ | Rob K. adds: if you happen to be using mxTryHard(), | ||
+ | ??? | ||
+ | |||
+ | ??? 9. In path-based (RAM) OpenMx, how do you include an algebra on a path? e.g. to allow a variable Z to moderate the impact of X on Y? | ||
+ | !!! Not answered yet. | ||
+ | |||
+ | ??? 10. If you do a study with low power (e.g. 30%) and you get a nominally significant p value (p < .05), can you trust your results? | ||
+ | !!! MCK: Two answer to this. A p-value of .04 with a large sample size (high power) and one of .04 with a small sample size (low power) both provide equal evidence for rejecting the null that the estimate is different from 0 (assuming that's your null). So one answer is, " | ||
+ | |||
+ | But another answer to this is " | ||
+ | |||
+ | JKH: Low power increases the likelihood that `significant' | ||
+ | |||
+ | |||
+ | ??? 11. In the Latent Dual Change | ||
+ | |||
+ | !!! Yes, in two main ways. An easy way is simply to treat the issue as a missing data problem. | ||
+ | |||
+ | Agreed. Treat as missing data problem. That advantage of full information maximum likelihood raw data analysis is that it's robust to missing data. | ||
+ | |||
+ | ??? 12. Can you correct for assortative mating without using the extended twin family design? (say, with data on how much people assortatively mate in the general pop) | ||
+ | !!! If you have an estimate of assortment, you could put it into the model by altering the value of DZr away from .5 gene sharing. | ||
+ | |||
+ | MN: A possibility would be to use a sensitivity analysis, depending on whether you think the marital similarity comes from assortment (selection of spouses) or from social environment similarity, or from marital interaction. | ||
+ | ====== Questions/ | ||
+ | ??? 1. What is the origin/ | ||
+ | !!! McArdle, J. J., & Goldsmith, H. H. (1990). [[http:// | ||
+ | |||
+ | ??? 2. Why can't we think of " | ||
+ | !!! One answer is, "you could, of course - that's just a different theory" | ||
+ | |||
+ | !!! I (Conor) consider General Health (GH) to be an index variable, that is a summary variable like the APGAR score (" | ||
+ | However, there are definitely situations in which GH may cause smoking. Consider an ex smoker, whose GH has declined due to, say, old age. The decline in GH may be a source of dispair and stress. In reaction the person starts smoking (to relieve stress). Here GH is causing smoking, but in a rather exceptional circumstance. So the causal status of GH may depend on the context? | ||
+ | |||
+ | ??? 3. Will the factor analysis of an 84-item autism scale be satisfactory (real) if the threshold is .3% (autistic spectrum disorder) and 99.7% of the subjects in the FA sample are normal? | ||
+ | I would argue that this threshold is not arbitrary and most children are not slightly autistic, they either are or not. So can I use 9,970 normal subjects scoring very low on an autism scale and 30 autistic children score very high on the scale in a factor analysis. Will the results of the FA validly reflect the underlying factor structure of the 84-item autism scale used in the study? | ||
+ | |||
+ | !!! Conor answers: 84 items is alot! Suppose that the response format is dichotomous (yes, no), and the probability of endorsement is low (say on average 2% of testees respond | ||
+ | |||
+ | ??? 4. What are the different optimizers for OpenMX, and why would you use one versus another? | ||
+ | !!! Legal values (from mxAvailableOptimizers() ) are:' | ||
+ | |||
+ | NPSOL is proprietary, | ||
+ | |||
+ | CSOLNP is our own open-source implementation of solnp (which has an R implementation in package [[https:// | ||
+ | |||
+ | SLSQP is one of a suite of optimizers implemented in [[http:// | ||
+ | |||
+ | There are also Newton-Raphson and EM optimizers (for advanced users). | ||
+ | |||
+ | FYI: If you are using OpenMx you can set the optimizer easily using the following: | ||
+ | < | ||
+ | You can also see the currently set optimizer with | ||
+ | < | ||
+ | If you are using umx you can see and set the optimizer easily with < | ||
+ | |||
+ | Concerning why you would use one versus another: NPSOL and SLSQP are similar in many respects. We're using NPSOL in this workshop largely for historical reasons–NPSOL was the ONLY available optimizer in OpenMx version 1, and in classic Mx before that. It's therefore what the workshop faculty are most familiar with. Savvy users interested in technical details may wish to consult the help pages for [[http:// | ||
+ | ??? | ||
+ | |||
+ | ??? 5. Can you please provide some guidance about which type of matrix to select (i.e., when to select " | ||
+ | !!! This is dependent on what you want the matrix to contain/do. A full matrix allows all cells to take any value. A symmetric matrix makes the upper triangle and lower triangle equal (as you might see in a correlation matrix), a lower matrix only stores the lower triangle (as you might want in a Cholesky where you are going to %*% the matrix with its transform). In general, there is no general answer to this question :-). | ||
+ | |||
+ | You might also check out the [[http:// | ||
+ | |||
+ | |||
+ | ??? 6. Can OpenMx accommodate sampling weights for the observations? | ||
+ | !!! Yes. What you do is run your model with a row-likelihood (likelihood for each subject), then embed that in a model which weights each row and tries to minimise the sum of those weighted likelihoods. | ||
+ | |||
+ | |||
+ | ??? 7. If you drop the C path in the AE model why doesn' | ||
+ | !!! When ' | ||
+ | |||
+ | ??? 8. This question will likely bring some face-palms, nonetheless…..I’m having trouble understanding the “starting value” component of the modeling. It seems like such an arbitrary method of beginning the analyses. | ||
+ | !!! One way to see that it is not arbitrary is to try putting in impossible values: the model will usually code red with non-positive definite matrices, or other problems. So you need to start things in the legal space. The further away from the bet vlue that you start things, the longer the model will take to run, as it finds its way to the best solution. | ||
+ | |||
+ | Another way to think about this is that if you knew where to start things exactly, you wouldn' | ||
+ | |||
+ | ??? | ||
+ | |||
+ | ??? 9. I notice that when we are testing submodels we opt to drop parameters as opposed to constraining or un-constraining them. What is the benefit to these different approaches? Which is the better approach? | ||
+ | !!! In general | ||
+ | |||
+ | ??? 10. How can you tell if something is behaving as a moderator vs. a mediator? | ||
+ | !!! By building and running | ||
+ | |||
+ | ??? 11. Is there a reference (say to give to reviewers) to say why univariate ACE models are/are not of value (in addition to one's multivariate results? | ||
+ | !!! They are probably right, in that the univariates can be a guide to how to model the multivariate model, and a good opportunity to test assumptions like sex-limitation and variance equation in a simpler context. | ||
+ | |||
+ | ??? 12. Is undertaking sex-limitation and other assumption tests at a univariate level acceptable (if one, say rejects sex-lmitation, | ||
+ | !!! The multivariate model is (somewhat) more powerful. It would likely allow you higher power to detect sex-limitation. | ||
+ | |||
+ | ??? 13. We mentioned running a saturated model for bivariate model, however the code provided does not test the twin order, means, ext. assumptions, | ||
+ | !!! (Scripts for bivariate saturated are in Hermine’s folder). Indeed, assumptions have to be tested in the multivariate case as well. | ||
+ | |||
+ | ??? 14. What to do with strongly skewed (non-normal) data? normalize or not? are log and square root transformations appropriate? | ||
+ | !!! | ||
+ | |||
+ | ====== Questions/ | ||
+ | |||
+ | ??? 1. In Sarah' | ||
+ | !!! (An) Answer: | ||
+ | |||
+ | Regarding fitGofs and fitEsts. | ||
+ | |||
+ | ??? 2. Where can I get classic reference papers? | ||
+ | !!! The QIMR website for classic papers [[https:// | ||
+ | |||
+ | Lindon mentioned [[http:// | ||
+ | ??? | ||
+ | |||
+ | ??? 3. When not using the projector screen on the left for code, please put it in the mode where it mirrors the presenter' | ||
+ | !!! Great idea! It's not perfect because it's mirroring the Powerpoint notes view. | ||
+ | |||
+ | ??? 4. How do I access the faculty folder when I can't access the workshop wifi? | ||
+ | !!! The faculty folder can only be accessed while on the Workshop network. It cannot be accessed from the hotel' | ||
+ | |||
+ | When the Workshop is over, all of the faculty files will be collected and placed on the web in a location that is accessible from everywhere. Once this is done an email will be sent out notifying you that it is available. | ||
+ | ??? | ||
+ | |||
+ | |||
+ | ??? 5. Could we be provided with code for a categorical moderator (Sex? or levels of SES) to compare the syntax? | ||
+ | !!! For 2 or 3 levels, you can " | ||
+ | |||
+ | BUT | ||
+ | |||
+ | If you want to run a Purcell type moderation script there are no changes required. The script is the same regardless of whether the moderator is continuous or categorical. Just remember to make your lowest category 0 so that you can interpret the output easily. | ||
+ | |||
+ | |||
+ | ??? 6. Why do we talk about opposite sex DZ twins having any common environment? | ||
+ | !!! Because the model is asking a question: If I posit something shared, how much impact does it appear to have? The answer might be (as your question implies) "not at all - they' | ||
+ | |||
+ | ??? 7. Where can I learn more about " | ||
+ | !!! [[https:// | ||
+ | !!! (Another) Another nice reference on tetrachoric correlations can be found at [[http:// | ||
+ | ??? | ||
+ | |||
+ | ??? 8. Why would you use **mxRun** over **mxTryHard**? | ||
+ | !!! **mxRun** just runs your model. You can manually rerun it by just entering the model you get back from mxRun into mxRun again. mxTryHard tries to be intelligent in moving parameter start values, and estimation to solve intransigent models. Basically use `mxRun` unless things don't work, then use **mxTryHard**. nb: mxTryHard is not a magic bullet - you might need to think about your start values and might need to fix your model. | ||
+ | |||
+ | ??? 9. Does V(D) in the twin model account for all non-additive genetic effects (e.g. [[https://en.wikipedia.org/wiki/ | ||
+ | !!! The dominance component will be able to account for epistatic inheritance. However, the issue of where variance components go when not modeled correctly is covered in more depth in Mat Keller' | ||
+ | |||
+ | ??? 10. Is there such a thing as a univariate threshold model, or does a threshold model only apply when there are 2+ traits of interest? | ||
+ | !!! You can very much have a threshold model for one phenotype: In fact it is quite common. e.g. ACE model of thresholded depression status. | ||
+ | !!! It sounds like you are confusing the number of traits with the number of thresholds/ | ||
+ | |||
+ | ??? 11. I see the workshop sessions are being recorded - is there somewhere we can see them online? | ||
+ | !!! There will be but not yet. | ||
+ | |||
+ | ??? 12. Question regarding Sanja' | ||
+ | !!! In this particular plot, Va <- (a + am*ses)^2 was the expected variance conditional on the level of the moderator, and was represented by the blue line. Here the x-axis was SES (given by ses <- sort(unique(c(mzData$ses))), | ||
+ | |||
+ | ??? 13. Is it true that the effects of age can mimic the effects of C in twins? | ||
+ | !!! Twins share their age, which, if behavior is linked to age (which it often is) makes pairs more similar just for that reason. | ||
+ | |||
+ | This is why we either residualize for age (and sex) (McGue1984), | ||
+ | |||
+ | McGue, M., & Bouchard, T. J., Jr. (1984). Adjustment of twin data for the effects of age and sex. Behavior Genetics, 14(4), 325-343. | ||
+ | |||
+ | Neale, M. C., & Martin, N. G. (1989). The effects of age, sex, and genotype on self-report drunkenness following a challenge dose of alcohol. Behavior Genetics, 19(1), 63-78. doi: | ||
+ | |||
+ | Schwabe, I., Boomsma, D. I., Zeeuw, E. L., & Berg, S. M. (2015). A New Approach to Handle Missing Covariate Data in Twin Research : With an Application to Educational Achievement Data. Behavior Genetics. doi: | ||
+ | ??? | ||
+ | |||
+ | ??? 14. Is it true that the effects of assortative mating can mimic the effects of C in twins? | ||
+ | !!! Assortative mating will drive up genetic similarity in DZ twins. As such it would be modelled by setting rDZ > .5 | ||
+ | !!! Assortative mating increases genetic variance, which increases the phenotypic variance, the MZ twin covariance AND the DZ covariance all by the same amount. | ||
+ | |||
+ | When not modelled explicitly, then yes - the increased DZ similarity will look like C (shared environment). | ||
+ | ??? | ||
+ | |||
+ | ??? 15. I am studying risk-taking in twins. I find the following: rMZF = 0.60, rMZM = 0.60, rDZF = 0.30, rDZM = 0.30, rDOS = 0.0. What could possibly account for these correlations? | ||
+ | !!! The trait is heritable, but behavioral expression of these genes is completely sex-specific, | ||
+ | |||
+ | ??? | ||
+ | !!! | ||
+ | |||
+ | ====== Questions/ | ||
+ | |||
+ | ??? 1. I'm still having trouble getting files out of Faculty folder (says Read Only + can't save to separate location) | ||
+ | |||
+ | !!! This should now be fixed. If you still have the problem, follow the directions below, and the problem should not return. | ||
+ | |||
+ | The " | ||
+ | ??? | ||
+ | |||
+ | ??? 2. Are there plug ins for drawing path diagrams in either R or OpenMx? | ||
+ | |||
+ | !!! (An) Answer: The free [[http:// | ||
+ | |||
+ | ??? 3. What is meant by Expectation Objects in the oneSATc script? | ||
+ | !!!(An) Answer: The expectation objects organize the predictions (i.e., expectations) of the model. | ||
+ | |||
+ | ??? 4. When testing if means are equal across twin order/zygosity, what level of significance of p do you use? p < .05 or p<.01? | ||
+ | |||
+ | !!!(An) Answer: Up to you. As there are few causes for means for T1 and T2, they are typically not significantly different. | ||
+ | !!!(Another) Answer: It is also important to consider the effect size. Because sample sizes for twin studies are often quite large, in the 100's or 1, | ||
+ | |||
+ | ??? 5. When reporting MZ and DZ correlations (e.g. in a paper), should we give the values of the saturated model, or those we get once we check the equality of means, variances / thresholds? | ||
+ | |||
+ | !!! (An) Answer: The model produces expected, not obtained covariances. So If you are reporting correlations, | ||
+ | !!!(Another) Answer: It can be better to report the model predicted covariances than those calculated from the sample. | ||
+ | |||
+ | ??? 6. Re inclusion of covariates, should we check their effect before or after exploring the equality of means/ | ||
+ | |||
+ | !!!(An) Answer: | ||
+ | |||
+ | ??? 7. Please provide a brief review of why dominance makes DZ twins less phenotypically similar. | ||
+ | |||
+ | !!!(An) Answer: Dominant genes don't make DZ less similar, like anything shared, it makes them more similar. Just much less so than do additive genes. | ||
+ | |||
+ | The example of you and your DZ twins sharing two alleles of an additive gene for height, and two alleles of a dominant-pattern gene for eye color. | ||
+ | |||
+ | Even if you share only one allele for the height genes, the copy you share will have the same effect in both of you, and make your heights similar. | ||
+ | |||
+ | However if you differ in your blue/brown allele status, the phenotype can be the same in both of you even though you differ genetically. That masked-case lowers the phenotypic correlation between DZ twins below where it would be for an additive model (.25 rather than .5) | ||
+ | |||
+ | ??? | ||
+ | |||
+ | ??? 8. is it possible to get example code to allow different means to be estimate in univariate models | ||
+ | |||
+ | !!!(An) Answer: Yes :-) | ||
+ | |||
+ | ??? 9. What is the benefit of using umx? | ||
+ | |||
+ | !!!(An) Answer: Reliability and mental and workflow efficiency. Because the functions operate at a high level, they encapsulate a lot of sub-tasks, and help you think at that level about what you are doing. Many of the helper functions make tasks that you do infrequently easier to learn or remember, Because the scripts are used by many people, they are less likely to contain errors. They also contain a lot of error-checking code, so can guide you to solve problems. | ||
+ | |||
+ | Reference: | ||
+ | ??? | ||
+ | |||
+ | ??? 10. What to do if your assumptions of twin order / zygosity are not met (e.g. means of twin 1 are significantly higher than means of twin 2) | ||
+ | |||
+ | !!!(An) Answer: Good question - should be discussed. In practice people often assign such events to multiple testing inflation and proceed. However you can simply leave (e.g.) the variances between sexes not equated, they will consume a degree of freedom but, if they were significantly different, will improve model fit. | ||
+ | |||
+ | ??? 11. How do you decide on the start value for lower bounds? | ||
+ | |||
+ | !!!(An) Answer: Start values and lower bounds are distinct ideas. | ||
+ | |||
+ | Start values are initial guesses, which hopefully allow the optimizer to find a global optimum for the model parameters. Lower (and upper) bounds are hard barriers which the parameter will never be allowed to cross. | ||
+ | |||
+ | So if you know that negative values are illegal, for instance, you might set lbound = 0 | ||
+ | |||
+ | If you want to divide variance among A, C, and E, you might start those parameters at around 1/3 of the variance each. | ||
+ | ??? |