http://blog.gmane.org/gmane.comp.lang.r.linguistics hourly 1 1901-01-01T00:00+00:00 http://gmane.org/img/gmane-25t.png http://gmane.org http://comments.gmane.org/gmane.comp.lang.r.linguistics/522 <pre>Dear R-langers, I'm trying to run a mixed effect model using the lmer() function and have run into some issues in interpreting the p-values generated by pvals.fnc(). The design is a between-subjects design, with two fixed effects (condition &amp; block; each with two levels), and one random effect (subject). Additionally, I have a set of weights that I want to include. When looking at the pvals.fnc() output,there appears to be a large discrepancy between the pMCMC values and the t-statistic p-values. Whereas one of the main effects and the interaction are far from significant judging by the pMCMC values, they are highly significant when looking at the t-statistic p-values (e.g. Condition: pMCMC = 0.2294; Pr(&gt;|t|) = 0.0000 &amp; Condition*Block: pMCMC = 0.3296; Pr(&gt;|t|) = 0.0000) . I have read that the t-statistic based p-values are less conservative, but the difference between these two values seems really extreme. Below some code that simulates the model and the data. The original data set has two precise characteristics that might influence the results, so I tried to simulate those characteristics in the mock data. That is: 1) there's fewer observations in block A than in block B; and 2) the weights for observations in block A generally are lower than those for block B. Running this code reproduces the original observation of conflicting pMCMC and p-T-test values. However, when excluding the weights argument from the lmer model, these values seem to converge, suggesting that the weights specification might be underlying these problems. In short, my question is whether anyone knows why these values diverge and what I could do to address this issue. Many thanks in advance! Tom block &lt;- as.factor(c(rep('a', times = 20), rep('b', times = 200))) condition &lt;- as.factor(c(rep(c('x', 'y'), each = 10), rep(c('x','y'), each = 100))) contrasts(block) &lt;- c(-0.5, 0.5) contrasts(condition) &lt;- c(-0.5, 0.5) subject &lt;- c(rep(1:4, each = 5), rep(1:4, each = 50)) intercept &lt;- 100block.me &lt;- 20condition.me &lt;- 30 err &lt;- rnorm(length(block), sd = 20) weights &lt;- c(rep(1, times = 20), rep(10, times = 200)) y &lt;- intercept + ifelse(block == 'a', block.me, 0) + ifelse(condition == 'x', condition.me, 0) + ifelse(block == 'a' &amp; condition == 'x', 30, 0) + (subject * 10) + err fm.1 &lt;- lmer(y ~ block * condition + (1 | as.factor(subject)), weights = weights, REML = FALSE) fm.1.mcmc &lt;- pvals.fnc(fm.1, addPlot=F) [[alternative HTML version deleted]] </pre> Tom Gijssels 2012-03-26T17:15:46 http://comments.gmane.org/gmane.comp.lang.r.linguistics/516 <pre>Hi all, this question is ultimately based on Florian's lecture1 slides here: http://hlplab.wordpress.com/2010/05/10/mini-womm/ I'm doing a mixed model logistic regression, with random intercepts for items and random slopes for items with respect to the fixed effect Indep2 (cf. slide 85): (a) glmer(formula = Dep ~ 1 + (1 | Item) + (0 + Indep2 | Item) + Indep1 + Indep2, data = my.data, family = binomial(link = "logit")) As per slide 88, I can also reduce the random effects to (1 + Indep2 | Item): (b) glmer(formula = Dep ~ 1 + (1 + Indep2 | Item) + Indep1 + Indep2, data = my.data, family = binomial(link = "logit")) It's not exactly clear to me what (1 + Indep2 | Item) does, since the output of both (a) and (b) includes random intercepts for items and random slopes for items by Indep2. At the same time, model (a) and (b) differ in their exact estimates. I would appreciate if someone could explain what the difference between model (a) and (b) is. Thanks Sverre </pre> Sverre Stausland 2012-03-19T16:37:15 http://comments.gmane.org/gmane.comp.lang.r.linguistics/511 <pre>Dear all, I have a few questions about how to report the results of mixed-effects analyses for publication. I have been perusing the Jaeger &amp; Kuperman presentation but a few questions remain. I have been asked by the reviewers to include a full regression table, which I take to comprise coefficient estimates, MCMC-based confidence intervals and MCMC-based p-value estimations. -Should the model that I use to report these values contain uncentered predictors, centered predictors, or centered and scaled predictors? -A few of my models involve random intercepts, and I believe that pvals.fnc() is not currently defined for models with random intercepts. Do you have any suggestions for how I should report these models? My models contain many control variables and only one or two variables that I am actually concerned with. As such, I have not worried about multicollinearity among the control variables. I suppose I should just state this somewhere to facilitate the interpretation of the regression tables? Lastly, is there any way to do a power analysis for mixed-effects models? One reviewer asked whether this was possible and noted that there may be rough approximations such as "the t-approximation to the coefficient-wise test". Thank you! Ariel </pre> Goldberg, Ariel M 2012-02-20T20:15:19 http://comments.gmane.org/gmane.comp.lang.r.linguistics/506 <pre>Hi, I am writing because I'm trying to run an lmer model and I keep getting negative deviances (and positive log-likelihoods, etc.). I've reinstalled R and updated all packages: platform x86_64-pc-mingw32 arch x86_64 os mingw32 system x86_64, mingw32 status major 2 minor 14.1 year 2011 month 12 day 22 svn rev 57956 language R version.string R version 2.14.1 (2011-12-22) lme4 version 0.999375-42 But the problem persists. It's not due to the data set. For example, I've rerun a simple model from Harald's languageR library (on the data set lexdec): library(languageR) data(lexdec) lmer(RT ~ Frequency + (1 | Subject), lexdec) Linear mixed model fit by REML Formula: RT ~ Frequency + (1 | Subject) Data: lexdec AIC BIC logLik deviance REMLdev -858.4 -836.8 433.2 -880.9 -866.4 [snip] Linear mixed model fit by REML Formula: RT ~ Frequency + Trial + (1 | Subject) Data: lexdec AIC BIC logLik deviance REMLdev -846.1 -819 428.1 -887.2 -856.1 [snip] The likelihood seems to develop in the expected (inverted) direction and so does the deviance estimate for the maximum REML model (REMLdev). Has anyone on this list been able to resolve this? I saw this behavior for version 2.13.1, but it persists after updating to 2.14.1. Do you get the same? I thought this had been resolved. Florian </pre> T. Florian Jaeger 2012-01-28T18:27:29 http://comments.gmane.org/gmane.comp.lang.r.linguistics/503 <pre>(I re send the message because it seems that it was not sent properly the last time) In my experiment, subjects were exposed to artificial languages with different word orders (two of them frequent among world languages: SOV, SVO and two of them infrequent: VSO, OSV). After training, subject had to classify new sentences as "correct" or incorrect, according to what they have learned. Sentences could either be correct, contain a syntax violation or a semantic violation (mismatch between a scene and the sentences describing it). Dependent variables were response latency and accuracy (right or wrong answer). I'm trying to analyze the accuracy (1 = right answer, 0 = wrong answer) data using a mixed logit model with "word order (OSV, SVO, SOV, VSO)" and "type of sentence" (correct, semantic violation, syntax violation) as fixed factors, and subject as a random factor. Word order is a between subjects variable, while type of sentences is a repeated measures factor.  My questions are: 1) In order to contrast each level of each factor with all the others, as well as their interactions: should I ran different models changing the reference category? Does this mean I should run 4 x 3 = 12 models? 2) Would it be correct to compare interaction levels with post hoc Tukey contrasts (for instance: OSV - correct vs. OSV semantic violation, SVO correct vs. OSV correct and so on?). 3) How do I interpret a significant interaction? For instance: ModeloAngel = lmer(respuest=="1" ~ 1 + grupo * tipoF + (1|sujeto), data=DatosAngel, family="binomial")  Fixed effects:                        Estimate Std. Error z value Pr(&gt;|z|)     (Intercept)             1.79585    0.19196   9.356  &lt; 2e-16 *** grupoOSV                0.25816    0.26740   0.965   0.3343     grupoSOV                0.70875    0.29315   2.418   0.0156 *   grupoSVO                0.59607    0.26769   2.227   0.0260 *   tipoFVsemanti          -1.01756    0.14765  -6.892 5.51e-12 *** tipoFVsintact          -1.46088    0.14566 -10.029  &lt; 2e-16 *** grupoOSV:tipoFVsemanti -0.29214    0.20841  -1.402   0.1610     grupoSOV:tipoFVsemanti -0.39714    0.23265  -1.707   0.0878 .   grupoSVO:tipoFVsemanti  0.03181    0.21459   0.148   0.8821     grupoOSV:tipoFVsintact  0.83284    0.21107   3.946 7.95e-05 *** grupoSOV:tipoFVsintact  0.42079    0.23408   1.798   0.0722 .   grupoSVO:tipoFVsintact  0.16667    0.21136   0.789   0.4304     If the reference levels are VSO and "correct": does this mean that performance of OSV in syntax violations trials is better than that of VSO in syntax violation trials. Or does this mean that OSV - syntax violations performance is better than VSO - "correct" performance? </pre> Angel Tabullo 2012-01-17T15:07:08 http://comments.gmane.org/gmane.comp.lang.r.linguistics/502 <pre>Dear colleagues, we would like to alert you to our upcoming workshop, which we believe may be interesting for many of the users of this list: *** BiMM 2012: Bielefeld Mixed Models Workshop *** Mixed-effects models are a powerful tool for the statistical analysis and modelling of psycho-/linguistic data and are increasingly used in experimental and corpus-oriented work. The Bielefeld Mixed Models Workshop (BiMM 2012) offers an opportunity to gain insight into the application of mixed-effects models. Focus will be placed on combining theoretical background with practical data analysis (using R). Participants can look forward to lively discussions with and between experts about how to use, report and interpret these methods. The workshop is targeted at students and researchers from psycho-/linguistics and related disciplines working or willing to work with mixed-effects models. No prior knowledge of these models is required. However, participants should be familiar with general statistics and R. Place and dates: Bielefeld University, Germany *** 22-24 February 2012 *** Topics: -Random effects models -Hierarchical Models -ANOVA vs. Mixed models -Model validation -Unbalanced designs -Outliers and data transformation, correlated predictors -Interpretation, visualisation and reporting Experts: -Hugo Quené (Utrecht University) -Dale Barr (Glasgow University) -Holger Mitterer (Max Planck Institute for Psycholinguistics, Nijmegen) -Shravan Vasishth (Potsdam University) -Marco van de Ven (Radboud Universität) Registration: To request a place, please send an email with a short description of your work and a motivation for participating in the workshop to BiMM2012-gM/Ye1E23mwN+BqQ9rBEUg&lt; at &gt;public.gmane.org by *** 22 January 2011 ***. The workshop fee (50 EUR) should be paid after receiving confirmation from the organisers. For further information please see http://www.spectrum.uni-bielefeld.de/BiMM2012/ With best regards, the organizers: Annett Jorschick Helene Kreysa Zofia Malisz Andreas Windmann Marcin Włodarczak ********************************************************************************************* Helene Kreysa Post-doctoral researcher Language and Cognition Group Cognitive Interaction Technology (CITEC) Room H1-132 Morgenbreede 39 33615 Bielefeld Germany tel: +49 (0)521 106 12248 (office) </pre> Helene Kreysa 2012-01-03T09:37:34 http://comments.gmane.org/gmane.comp.lang.r.linguistics/496 <pre>Dear r-lang users, I have a set of binary data from a 2 by 3 design study. I centered the two-level predictor ('LHcenter': local &amp; LD) but did not centered the 3 level predictor ('cond': A, B, &amp; C). As you can see below in the triangular matrix towards the end of the lmer output, there is significant collinearity; the absolute values of some of the correlations are above 0.6. ---8&lt;---------------------------------------------------------------------------------------------------------------------------- Generalized linear mixed model fit by the Laplace approximation Formula: true ~ cond * LHcenter + (1 | subject) + (1 | items) Data: offlineTarget AIC BIC logLik deviance 747.8 785.9 -365.9 731.8 Random effects: Groups Name Variance Std.Dev. items (Intercept) 0.13838 0.37199 subject (Intercept) 1.44652 1.20271 Number of obs: 864, groups: items, 30; subject, 29 Fixed effects: Estimate Std. Error z value Pr(&gt;|z|) (Intercept) -0.09775 0.30932 -0.316 0.75199 condB 0.62445 0.25534 2.446 0.01446 * condC 0.70552 0.27286 2.586 0.00972 ** LHcenter 4.86821 0.42482 11.459 &lt; 2e-16 *** condB:LHcenter -2.31392 0.51162 -4.523 6.1e-06 *** condC:LHcenter -1.00381 0.54643 -1.837 0.06621 . --- Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1 Correlation of Fixed Effects: (Intr) condB condC LHcntr cnB:LH condB -0.526 condC -0.491 0.607 LHcenter -0.085 0.130 0.118 cndB:LHcntr 0.070 -0.053 -0.093 -0.801 cndC:LHcntr 0.066 -0.093 0.033 -0.735 0.607 ---8&lt;---------------------------------------------------------------------------------------------------------------------------- I also tested whether the interaction is significant by comparing the full model with one without the interaction term: glmer(true ~ cond * LHcenter + (1 | subject) + (1 | items), data=offlineTarget, family=binomial) glmer(true ~ cond + LHcenter + (1 | subject) + (1 | items), data=offlineTarget, family=binomial) And I observed a significant difference, suggesting that there is significant interaction. So I proceeded to conduct planned comparisons. To do so, I created a new predictor in the table object by merging the two factors (a 3 level factor and a 2 level factor) together, resulting in a new factor named 'posthoc_cond' with 6 levels: Local_A, Local_B, Local_C, LD_A, LD_B, &amp; LD_C I conducted glmer WITHOUT centering the 6-level fixed predictor 'posthoc_cond'. posthoc_result = glmer(true~posthoc_cond + (1|subject) + (1|item), data=offlineTarget, family="binomial") and then use glht() from 'multcomp' to conduct paired comparisons. The problem with posthoc_result, again, is high collinearity (see below) ---8&lt;---------------------------------------------------------------------------------------------------------------------------- Correlation of Fixed Effects: (Intr) pst_A_ p_B_LD pst_B_ p_C_LD psthc_cndA_ -0.865 psthc_cB_LD -0.835 0.712 psthc_cndB_ -0.944 0.896 0.734 psthc_cC_LD -0.928 0.810 0.779 0.896 psthc_cndC_ -0.891 0.925 0.710 0.914 0.834 ---8&lt;---------------------------------------------------------------------------------------------------------------------------- So my question is, in my case, what can be done to reduce collinearity. It seems that centering those multi-level predictors is not applicable in my case since I am interested in whether different levels of the fixed predictors have different means. Centering these multilevel predictors would not allow me to test that. Thank you in advance for your help!! Best, Xiao </pre> Xiao He 2011-11-27T23:57:42 http://comments.gmane.org/gmane.comp.lang.r.linguistics/490 <pre>To whom it may concern: I would like to use a mixed-effect regression approach to examine the adaptation effect (i.e., learning) during sentence comprehension. Right now, I am trying to figure out the maximum random effects. One of the models produced the following summary Random effects: Groups Name Variance Std.Dev. Corr Subject (Intercept) 19867.54 140.952 cPrimeType 367.20 19.162 -1.000 cCondition 5558.49 74.555 -1.000 clogLength 8128.49 90.158 0.954 cLogPresOrder 3033.62 55.078 -0.216 cPrimeType:cCondition 17606.58 132.690 0.164 cPrimeType:cCondition:cPresOrder 351.92 18.759 0.142 Item (Intercept) 2870.65 53.578 cCondition 3391.10 58.233 0.668 cCondition:cPrimeType 1224.53 34.993 -0.652 Residual 31654.56 177.917 Baayen, Davidson, and Bates (2008) noted that "the high correlation of the intercept and slope for the subject random effects (-1.00) indicates that the model has been overparameterized" (p. 395). So, I inspected the correlations and identified three high correlations from the table. Therefore, I simplified the model by removing the by-subject adjustments to the slopes of cPrimeType, cCondition, and clogLength. The simplified model produced the following summary. Random effects: Groups Name Variance Std.Dev. Corr Subject (Intercept) 19443.07 139.438 cLogPresOrder 2832.97 53.226 -0.341 cPrimeType:cCondition 1043.35 32.301 -0.133 cPrimeType:cCondition:cPresOrder 204.57 14.303 0.364 Item (Intercept) 2935.82 54.183 cCondition 3545.04 59.540 0.740 cCondition:cPrimeType 1878.62 43.343 -0.656 Residual 36990.34 192.329 Now, the correlation values seem to be in the range of acceptable values (i.e., not too high). Finally, in order to verify that the simpler model is justified, I carried out a likelihood ratio test. Df AIC BIC logLik Chisq Chi Df Pr(&gt;Chisq) riming.lmer 27 22598 22744 -11272 rev_priming.lmer 45 22474 22718 -11192 159.35 18 &lt; 2.2e-16 *** --- Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1 My understanding is that this likelihood ratio test indicates that the removal of the by-subject adjustments to the slopes is NOT justified. My question is which model should I choose to report my results? priming.lmer or rev_priming.lmer? Thank you very much for your suggestions on this matter in advance! Best, Sunfa </pre> Sunfa Kim 2011-11-19T11:14:18 http://comments.gmane.org/gmane.comp.lang.r.linguistics/486 <pre>Hello everyone, I'm trying to set up a model with a three-way contrast: between linguistic, non-linguistic, and control. I'd like to compare all three to each other, but for the moment, I'll focus on one set of Helmert contrasts: [,1] [,2] l -1 -1 n 1 -1 c 0 2 This should compare 1) linguistic to non-linguistic, and 2) 2 times control to the average of linguistic and non-linguistic. So far so good. My "baseline" model includes some control predictors, and more importantly, random intercepts and slopes for subject and item. The slopes are by "condition," and in this case I've used the contrasts, as follows: baseline&lt;-lmer(logRT~(1+contr1+contr2|subject)+(1+contr1+contr2|item)+controlpredictors) This, in order to specify the "maximal" structure allowed/justified by the data. So far still so good (right?). Then I add the fixed effect contrasts: ofInterest&lt;-update(baseline,.~.+contr1+contr2) The fixed effect output of this type of model indicates that, for one set of RT's, the t-value for both contrasts "should be" significant. The output for an identical model for a different set of RT's indicates that one (but not the other) "should be" significant. To get a usable significance, I have to use log-likelihood tests, because pvals.fnc() won't work with random slopes (or is it just covariances? Regardless...). But because my model comparison includes both contr1 and contr2, the test isn't going to spit out individual significance values for each, t-values or not. Is there some way to tell when contr1 is significant, independent of the significance of contr2, in a model like the one I've described? Can I rely on those t-values at all? I'm happy to provide more information if it would clarify anything I've muddied. Jason </pre> Jason Kahn 2011-11-10T21:49:07 http://comments.gmane.org/gmane.comp.lang.r.linguistics/485 <pre>Hello all, I wonder if anyone can help me with a question about pvals.fnc and lmer. I'm running some re-analyses on a dataset analysed some years ago (2007), working with the languageR package. In the earlier analyses, to obtain values for reporting, we used the then current version of pvals.fnc, with the following code: x=pvals.fnc(model.lmer) x$summary x$anova x$summary gave outputs like: # Estimate Std.Error DF t.value pvals ci950 ci990 ci999 #(Intercept) 42.4514 2.1493 973 19.751 0.00000 TRUE TRUE TRUE #typePsPr -21.2240 1.7903 973 -11.855 0.00000 TRUE TRUE TRUE #stressPN -5.0640 0.8967 973 -5.648 0.00000 TRUE TRUE TRUE #MorDMis 2.5627 1.7903 973 1.431 0.15275 FALSE FALSE FALSE #typePsPr:stressPN 5.8140 1.0298 973 5.646 0.00000 TRUE TRUE TRUE #stressPN:MorDMis 1.6569 1.0296 973 1.609 0.10794 FALSE FALSE FALSE while x$anova gave e.g. # Df SumSq MeanSq Denom F pvals #type 1 7379.4 7379.4 973 113.83044 0.00000 #stress 1 398.6 398.6 973 6.14858 0.01332 #MorD 1 256.3 256.3 973 3.95354 0.04705 #type:stress 1 2069.7 2069.7 973 31.92602 0.00000 #stress:MorD 1 167.9 167.9 973 2.58993 0.10787 These days, based on a more recent version of languageR, my usual practice when running lmer is to do: pvals.fnc(model.lmer)$fixed This gives an output broadly similar to that of x$summary above, but with two sets of p values, one based on Monte Carlo Markov chains. What I can't seem to obtain is an output like that of x$anova above, i.e. one that includes SumSq, MeanSq, df and F. Does anyone know if this is possible? If not, what can I report as the equivalent of (in ANOVA) the df, F and p for a main effect or interaction term? Any advice very gratefully received. Best wishes, Rachel </pre> Rachel Smith 2011-11-09T00:34:33 http://comments.gmane.org/gmane.comp.lang.r.linguistics/475 <pre>Dear R users, I’m using mixed effects models (lmer) to predict a binary dependent variable as a function of 1.a categorical predictor (A)with 2 levels (A1 and A2) , 2. another categorical predictor (B) with three levels (B1, B2 and B3) and 3. The interaction between these two predictors. I have tried two models but they return different results and I’m not sure which one is correct. I’m interested in the main effect of B and the interaction between A and B (because A alone has a significant effect in both models). My problem is that there seem to be two sensible ways of examining the main effect of B: 1. to helmert code and 2. to center. But these two methods produce opposite results! I don’t know which one I should use. Here are the two models with some details and their outputs: Model 1: ‘A’ is centered. ‘B’ is helmert coded (‘B1’(baseline)=2, ‘B2’=-1, ‘B3’=-1) so that I can get a main effect of B by checking to see whether baseline condition in B differs from the mean of B1 and B2 . The lmer output returns a significant effect of B and no significant AxB interaction. However, as is highlighted below (in pink), the correlation between B and the ‘AxB’ interaction is high (-54%). Generalized linear mixed model fit by the Laplace approximation Formula: response ~ A * B+ (A + 1 | sub) + (1 | item) Data: mydata AIC BIC logLik deviance 783 822.6 -383.5 767 Random effects: Groups Name Variance Std.Dev. Corr item (Intercept) 0.7293 0.85399 sub (Intercept) 2.0871 1.44468 A 1.3812 1.17524 0.562 Number of obs: 1038, groups: item, 42; sub, 36 Fixed effects: Estimate Std. Error z value Pr(&gt;|z|) (Intercept) 1.05261 0.30283 3.476 0.000509 *** A -3.91080 0.32239 -12.131 &lt; 2e-16 *** B 0.36128 0.09751 3.705 0.000211 *** A:B -0.29638 0.18681 -1.586 0.112626 --- Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1 Correlation of Fixed Effects: (Intr) A B A 0.155 B 0.160 -0.278 A:B -0.156 0.238 -0.540 Model 2: ‘A’ and ‘B’ are both centered. The lmer output returns no significant effect of B but the A:B interaction is significant. The correlations between predictors are generally lower and the correlation between B and A:B is reduced to -26%. Generalized linear mixed model fit by the Laplace approximation Formula: resonse ~ A * B + (A + 1 | sub) + (1 | item) Data: mydata AIC BIC logLik deviance 756.1 795.7 -370.1 740.1 Random effects: Groups Name Variance Std.Dev. Corr item (Intercept) 0.87028 0.93289 sub (Intercept) 2.41707 1.55469 A 1.23669 1.11206 0.533 Number of obs: 1038, groups: item, 42; sub, 36 Fixed effects: Estimate Std. Error z value Pr(&gt;|z|) (Intercept) 1.1004 0.3239 3.398 0.000679 *** A -4.0941 0.3248 -12.605 &lt; 2e-16 *** B -0.1461 0.1400 -1.043 0.296851 A:B 1.7923 0.2818 6.360 2.01e-10 *** --- Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1 Correlation of Fixed Effects: (Intr) A B A 0.138 B -0.148 0.185 A:B 0.106 -0.292 -0.265 I personally think Model 2 is better but the thing is that I have centered a categorical predictor with *three* levels. In my searches in the web, I have never seen a three-level predictor to be centered; they were all two-level categorical predictors. I have used the scale() function to center the predictors (I first converted them to numeric variables and then used the scale () function to center them). As I mentioned, my problem is that I don’t know how to get a main effect of B as well as a *main* A:B interaction. On the one hand, it seems logical to compare ‘B1’ (baseline) with the mean of the other two B conditions to see if the B manipulation has a general effect. On the other hand, I hear that one needs to center variables to get a main effect. I would be grateful of you could please help me Regards, Hossein </pre> hossein karimi 2011-10-10T14:05:10 http://comments.gmane.org/gmane.comp.lang.r.linguistics/472 <pre>I have a text file encoded in CSV format that I plan to run a series of linguistic analysis upon using R. My first thought is to use the read.csv function to get the data into R. But suspect that this might be naive as the file itself contains over 250,000 records and occupies over 1.5Gbytes of disk space in its raw form. Each record has a large text field that extends of multiple lines of the file; in some records this is equivalent to a few paragraphs in other the text would run to pages. There is a second multiline text field but this is much shorter --- maybe 20 different phrases. There appear to escaped quotes throughout; though with a file this size verifying this is difficult and some text editor crash trying to read it all in. I am therefore dubious that read.csv is the right mechanism and seeking a better method. I am open to suggestions for input method and even segmentation if necessary, provided segmentation does not prevent analysis of the entire data. Would access to the records by R be any quicker if I pre-processed the file into some relational/object database prior to analysis? I will note in passing I did not collect the original data neither did I decide to use CSV. This is the format in which the file was given to me for the purposes of my study. CSV is the worst possible format for the originator to have used for such a large collection of text. Regards, Trevor. &lt;&gt;&lt; Re: deemed! </pre> Trevor Jenkins 2011-09-08T12:34:53 http://comments.gmane.org/gmane.comp.lang.r.linguistics/467 <pre>Dear R users, We created a logistic regression model to investigate the influence of different factors on Dutch word order variation. To avoid collinearity and to increase interpretability of the effects, we decided to center the predictors. For one predictor with two levels and one ordinal (5 point scale) predictor, we did this by subtracting the mean; for a third predictor with three levels, we used contrast coding. The levels of the latter predictor (called PP_TYPE_3) were coded as follows: [,1] [,2] abs 1 0 loc 0 1 temp -1 -1 The model summary gives us the following: Coef S.E. Wald Z P Intercept -0.26760 0.10303 -2.60 0.0094 cDEF3S -0.27503 0.06216 -4.42 0.0000 cANIM2_S -0.17869 0.18324 -0.98 0.3295 PP_TYPE_3=loc 0.60699 0.13471 4.51 0.0000 PP_TYPE_3=temp 0.08269 0.12773 0.65 0.5174 cDEF3S * cANIM2_S -0.38080 0.12298 -3.10 0.0020 The names suggest that the model gives the estimates for the levels 'loc' and 'temp', taking 'abs' as reference level. But this is not what we want, is it? Shouldn't the model have taken the columns of the contrast matrix as the recoded factors? We are also a bit confused about the correct interpretation of the estimates of the three-level predictor. Do these represent the difference between one level and the intercept (=the grand mean), or between one level and the mean of the other levels of that predictor? As a side question: is it correct to center the ordinal predictor the way we did? Thanks for your help! Best, Jorrig Vogels Geertje van Bergen </pre> J. Vogels 2011-08-30T12:57:21 http://comments.gmane.org/gmane.comp.lang.r.linguistics/463 <pre>Hi, I have a question about making inferences when power might be an issue. I'm examining whether a variable has a significant effect in different parts of the syllable. To do this, I have 2 different data sets, Onset and Coda, which I'm using to determine if the variable has effects in the syllable onset and coda, respectively. The variable is significant (very small p-value) in the onset but is marginally significant in the coda (p= .055 in the full model, and model comparison with a baseline model that does not contain this variable gives a p-value of .07). While it's always difficult to know how to interpret a marginally significant effect, one issue that complicates the matter is that the Coda dataset has fewer items and trials than the Onset dataset. One thing that I'd like to do is determine whether the marginal effect could simply be due to a lack of power. My idea was to take a random sample of the Onset dataset so that it matches the size of the coda dataset and see if the variable of interest remains significant even in this reduced dataset. I figure that I would need to do this sampling many times (e.g., 10,000 times) to make sure that the effect is robust. Is this a sensible approach? Am I going to run into a Type I/II error situation by doing 10,000 model comparisons? Thank you, Ariel </pre> Ariel M. Goldberg 2011-08-03T21:26:13 http://comments.gmane.org/gmane.comp.lang.r.linguistics/456 <pre>Dear list, I have had a problem with model comparison for several months, so now I finally worked up my courage to ask for your help and hope that you can settle the question. I have frequently encountered positive logLik values and now heard that this might be due to bug in the lmer function. However, I also recently found Douglas Bates stating that "a positive log-likelihood is acceptable in a model for a continuous response" in an S-list. Positive logLiks appear in Baayen's 2008 introductory book, always together with negative AIC and BIC. He does not seem to treat them as erroneous. Instead, if I understood correctly, he chooses the model with more negative AIC/BIC (smaller value) and more positive logLik (larger value) as the better model in these comparisons. So did I get it right and is this the way to go or is there a bug that inverts the polarity of the numbers? As second question: Is there a general rule of thumb for cases when AIC and BIC point into different directions? Does it depend on the data set? Or is it a matter of taste how much one wants to avoid overfitting? Should one trust the value that agrees with the logLik? Many thanks in advance Anja </pre> Anja Arnhold 2011-08-01T08:29:40 http://comments.gmane.org/gmane.comp.lang.r.linguistics/439 <pre>Dear R-users, I have been wondering about something with the pvals.fnc function. As we know, the pvals function gives two p-values, one based on the posterior distribution (pMCMC) and one based on the t-distribution. In my experience most of the time the two values are very similar. However, I have recently come across situations where they are wildly different. I have been particularly surprised to see t-values above 2 that have associated pMCMC values that are not even close to significance, while at the same time the t-distribution based p-value is significant. For example, a recent model I worked with looked something like this: model1 = lmer(RT~x*y+(1+x|Subject)+(1|Item) and gave me a t-value of 2.07 for the interaction, with a pMCMC p-value of 0.4756 and a t-distribution p-value of 0.0381. Obviously I like one of these better than the other! I know that the latter p-value is anticonservative, but the magnitude of the discrepancy is nonetheless surprising to me, given the t-value. I'd be very grateful for any advice on how to proceed in cases like this. I'm using lme4 version 0.99875-6. Many thanks, Jakke </pre> Jakke Tamminen 2011-07-29T19:58:25 http://comments.gmane.org/gmane.comp.lang.r.linguistics/437 <pre>Dear ling-r-lang users, I'm writing to get some advice on the use of logit mixed model for eyetracking data anlaysis. My experiment has three IVs with two levels for each - language (English vs.Korean),stress pattern (trochaic vs. iambic), phonation type (aspirated vs.lax), and one continuous IV, time. And the DV is binary, either 0 or 1. What I want in the test is where in the time course of word recognition, the trochaic and iambic words are different in their activation of the target words, and how they are interacting with the phonation types in the two language groups. For this, I first tried logit mixed effect model as below. lmer(gaze~stress*lg*phonation*time+(1|subj)+(1|item), data ,family=="binomial") The issue that I had with this model is that it doesn't show interactions between specific levels of factors. For example, I couldn't test whether English speakers' behavior for aspiraed trochaic words (default level) is different from the one for aspirated iambic ones. So I have made a dummy combinatorial variable column, "int", which combines the levels of lg, stress pattern, and phontion type (e.g., "eiasp" as a combination of English spk's respose for aspirated iambic words) and ran the model as below: lmer(gaze~int*time+(1|subj)+(1|item), data, family=="binomial") My question is whether having such a dummy combinatorial variable is a legitimate for the mixed effect model. If it's not legitimate, I'd like to know what's the way to examine the interactions between specified levels of different factors of interest. My another question is how can we test where in the timecourse the two levels of interest are sigificantly different from each other (in terms of slope change). For this , I have segmented time into every 100ms window and treated the window as a factor. It looks the outcome supports slope change in plot (in logit) and compares two levels of interest in each time window. But again, I'm not sure whether this is a right way to examine the time effect. If not, what model or approach do I have to make? My questions might have been arisen by my misunderstanding of the model, so it would be greatly appreaciated if you would be able to give me your valuable advice. Thank you! Best regards, Jeonghwa Shin </pre> Jeonghwa Shin 2011-07-28T17:07:56 http://comments.gmane.org/gmane.comp.lang.r.linguistics/425 <pre>Hi all, I'm sure I'm not the only Windows user out there trying to embed IPA symbols in R graphics and then make a pdf out of it. If anyone knows how to do that, I would appreciate some help. Because I have not succeeded. Here is what I have tried. I'm creating a plot in R with IPA phonetic symbols, using the standard font Doulos SIL (http://scripts.sil.org/cms/scripts/page.php?item_id=DoulosSILfont&amp;_sc=1) When attempting to create a pdf of this plot in R using the pdf() function, I will get the following error messages: Error in text.default(my.x, my.y, : Invalid font type In addition: Warning messages: 1: In text.default(my.x, my.y, : font family not found in PostScript font database I've been advised that I can resolve this problem using the Cairo package for R. But I have not succeeded. Here is my call: It produces a pdf file, but all the IPA symbols come out as boxes. Please note that R has no difficulties producing this plot in its graphic window with IPA symbols. It's only the pdf output I'm having difficulties with. Any help would be appreciated! Further details follow below. R version 2.13.1 (2011-07-08) Platform: i386-pc-mingw32/i386 (32-bit) locale: [1] LC_COLLATE=English_United States.1252 [2] LC_CTYPE=English_United States.1252 [3] LC_MONETARY=English_United States.1252 [4] LC_NUMERIC=C [5] LC_TIME=English_United States.1252 attached base packages: [1] stats graphics grDevices utils datasets methods base other attached packages: [1] Cairo_1.4-9 loaded via a namespace (and not attached): [1] tools_2.13.1 [1] "LC_COLLATE=English_United States.1252;LC_CTYPE=English_United States.1252;LC_MONETARY=English_United States.1252;LC_NUMERIC=C;LC_TIME=English_United States.1252" </pre> Sverre Stausland 2011-07-20T19:33:49 http://comments.gmane.org/gmane.comp.lang.r.linguistics/412 <pre>Dear R users, I have a logit mixed model with two categorical predictors (two types of salience measures) and a categorical dependent variable (pronoun used Y/N). One predictor has 2 levels, and the other has 3. I centered the 2-level predictor, and transformed the 3-level predictor into two binary predictors using contrast (sum) coding. I determined the random-effects structure by starting from a full model, and eliminating step by step all terms without a significant contribution to the model. In the final model, I end up with random intercepts for subjects and items, and a by-subject random slope for my 2-level predictor. In this model, I get significant interactions between the fixed factors, which I had not expected to be significant by just looking at the data. Removing the random slope from the model completely eliminates these interactions, but model comparison suggests the random slope should be included. I have attached the two model summaries below. Now my question is: is it normal to find such a large influence of random effects on the fixed effects structure? How do I know the interaction effects are not spurious? And what exactly do these findings mean? Participants varied greatly in their reaction to predictor B, but when this variation is accounted for, predictor B affects pronoun use, but differently for each level of predictor A? Jorrig Vogels PhD candidate Tilburg Univ., Netherlands ================================================================ Model with random slope: Generalized linear mixed model fit by the Laplace approximation Formula: PRO ~ cAGTOP * cAGVIS + (1 + cAGVIS | SUBJ) + (1 | ITEM) Data: vislingag AIC BIC logLik deviance 318.4 361.4 -149.2 298.4 Random effects: Groups Name Variance Std.Dev. Corr SUBJ (Intercept) 49.6457 7.0460 cAGVIS 21.8342 4.6727 0.663 ITEM (Intercept) 1.3205 1.1491 Number of obs: 544, groups: SUBJ, 48; ITEM, 12 Fixed effects: Estimate Std. Error z value Pr(&gt;|z|) (Intercept) -2.578 1.217 -2.117 0.03422 * cAGTOP1 -6.627 0.913 -7.259 3.90e-13 *** cAGTOP2 9.868 1.502 6.569 5.05e-11 *** cAGVIS -1.699 1.008 -1.685 0.09207 . cAGTOP1:cAGVIS -3.223 1.170 -2.755 0.00587 ** cAGTOP2:cAGVIS 3.120 1.371 2.275 0.02289 * --- Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1 Correlation of Fixed Effects: (Intr) cAGTOP1 cAGTOP2 cAGVIS cAGTOP1: cAGTOP1 0.075 cAGTOP2 -0.041 -0.867 cAGVIS 0.535 0.108 -0.059 cAGTOP1:AGV 0.074 0.562 -0.346 0.128 cAGTOP2:AGV -0.049 -0.480 0.528 -0.054 -0.668 Model without random slope: Generalized linear mixed model fit by the Laplace approximation Formula: PRO ~ cAGTOP * cAGVIS + (1 | SUBJ) + (1 | ITEM) Data: vislingag AIC BIC logLik deviance 324.3 358.7 -154.2 308.3 Random effects: Groups Name Variance Std.Dev. SUBJ (Intercept) 21.63217 4.65104 ITEM (Intercept) 0.61539 0.78447 Number of obs: 544, groups: SUBJ, 48; ITEM, 12 Fixed effects: Estimate Std. Error z value Pr(&gt;|z|) (Intercept) -1.41142 0.77639 -1.818 0.0691 . cAGTOP1 -4.59707 0.52139 -8.817 &lt;2e-16 *** cAGTOP2 7.13115 0.84489 8.440 &lt;2e-16 *** cAGVIS -0.35538 0.40416 -0.879 0.3792 cAGTOP1:cAGVIS -0.59940 0.58255 -1.029 0.3035 cAGTOP2:cAGVIS -0.08268 0.56682 -0.146 0.8840 --- Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1 Correlation of Fixed Effects: (Intr) cAGTOP1 cAGTOP2 cAGVIS cAGTOP1: cAGTOP1 0.070 cAGTOP2 -0.040 -0.867 cAGVIS 0.000 0.061 -0.036 cAGTOP1:AGV 0.038 0.082 -0.012 0.102 cAGTOP2:AGV -0.020 0.008 -0.037 0.037 -0.575</pre> Jorrig Vogels 2011-05-11T09:50:22 http://comments.gmane.org/gmane.comp.lang.r.linguistics/411 <pre>Dear R-lang-ers, I am currently trying to do some modelling on corpus extracted data with my students using a random intercept model. Our model tries to predict the position of French attributive adjectives wrt to their head noun given several variables. Our setting is a logistic regression with a random variable (random intercept) set on an adjective lemmata variable. Using lmer(...) we have that : (1) the distribution of the conditional modes is all but normal. (2) lmer does not converge properly : (the deviance function is getting flat as the algorithm progresses towards the solution, hence lmer iterates way too many times and yields an overfitted model) We tried to solve issue 2 : on convergence, we have been able to get a better convergence with the library lme4a. However problem (1) remains : the words lemmatas being Zipf distributed, most random intercepts have estimates close to 0 (the grand mean). These random intercepts are also mostly those for the words with low frequency in the data (hapax or quasi hapax words) for which the standard error of estimation is largest. Words with a higher frequency have better estimates, and their intercepts apparently follow a normal pattern. But the overall distribution looks like a mixture of (1) a peak around 0 made mostly (but not only) of poorly estimated ranefs and (2) a flat normal pattern of better estimated ranefs whose apparent mean is clearly much greater than 0. We tried to fix this, - by replacing in the data low frequency words forms (below some given threshold) by a unique word form, say 'hapax' aiming at reducing the above mentioned estimation problem for low frequency words. - by adding and removing as fixed effect an other word frequency variable in the model (with which the random effect could interact) yet that does not really help. The distribution remains highly skewed... This question seems to be a general one, since I suspect people using corpus distributed data should experience similar problems. I wonder whether someone has already run into similar problems, and which kind of solution he might have found... any hint would help... many thanks, Benoit </pre> Benoit Crabbé 2011-03-03T13:23:56 http://comments.gmane.org/gmane.comp.lang.r.linguistics/410 <pre>Hi dear r-lang users, I have a question about ANOVA type of main effects. From what I've read, one way to obtain the type of main effect one would get with ANOVA is to do model comparisons, like below: mod1&lt;-lmer(dv ~ iv1 + iv2 + (1|subject), data) mod2 &lt;-lmer(dv ~iv1 + (1|subject), data) anova(mod1, mod2). A significant difference indicates that the factor iv2 make significant contribution to the model. However, I wonder if there are other ways to obtain the same information. Specifically, I have a 2X2 design, I wonder if, after I center the two 2-level factors, the results would be equivalent to ANOVA type of main effects as opposed to simple effects. Thank you in advance! Xiao </pre> Xiao He 2011-02-23T19:42:43 Search the mailing list at Gmane query http://search.gmane.org/?group=$group=gmane.comp.lang.r.linguistics