This vignette contains answers to questions received from users or posted on discussion boards like Cross Validated and Stack Overflow
Inf for the degrees of freedom?Estimated marginal means (EMMs), a.k.a. least-squares means, are predictions on a reference grid of predictor settings, or marginal averages thereof. See details in the “basics” vignette.
Perhaps your question has to do with interacting factors, and you want to do some kind of post hoc analysis comparing levels of one (or more) of the factors on the response. Some specific versions of this question…
My first answer is: plots almost always help. If you have factors A, B, and C, try something like emmip(model, A ~ B | C), which creates an interaction-style plot of the predictions against B, for each A, with separate panels for each C. This will help visualize what effects stand out in a practical way. This can guide you in what post-hoc tests would make sense. See the “interactions” vignette for more discussion and examples.
This is a situation where it may well be appropriate to compare the slopes of trend lines, rather than the EMMs. See the help page for emtrends() and the discussion of this topic in the “interactions” vignette
You need to be careful to define the reference grid consistently. For example, if you use covariates x and xsq (equal to x^2) to fit a quadratic curve, the default reference grid uses the mean of each covariate – and mean(xsq) is usually not the same as mean(x)^2. So you need to use at to ensure that the covariates are set consistently with respect to the model. See this subsection of the “basics” vignette for an example.
This will happen if you fitted a model where the treatments you want to compare were put in as a numeric predictor; for example dose, with values of 1, 2, and 3. If dose is modeled as numeric, you will be fitting a linear trend in those dose values, rather than a model that allows those doses to differ in arbitrary ways. Go back and fit a different model using factor(dose) instead; it will make all the difference. This is closely related to the next topic.
Equivalently, users ask how to get post hoc comparisons when we have covariates rather than factors. Yes, it does work, but you have to tell it the appropriate reference grid.
But before saying more, I have a question for you: Are you sure your model is meaningful?
sex (coded 1 for female, 2 for male), no problem. The model will produce the same predictions as you’d get if you’d used these as factors.brand (coded 1 for Acme, 2 for Ajax, and 3 for Al’s), this model is implying that the difference between Acme and Ajax is exactly equal to the difference between Ajax and Al’s, owing to the fact that a linear trend in brand has been fitted. If you had instead coded 1 for Ajax, 2 for Al’s, and 3 for Acme, the model would produce different fitted values. You need to fit another model using factor(brand) in place of brand.Assuming that issue is settled, you can do something like emmeans(model, "sex", at = list(sex = 1:2)), to get get separate EMMs for each sex rather than one EMM for the average numeric value of sex.
An alternative to the at list is to use cov.reduce = FALSE, which specifies that the unique values of covariates are to be used rather than reducing them to their means. However, the specification applies to all covariates, so if you have another one, say age, that has 43 different values in your data, you will have a mess on your hands.
See “altering the reference grid” in the “basics” vignette for more discussion.
Estimated marginal means summarize the model that you fitted to the data – not the data themselves. Many of the most common models rely on several simplifying assumptions – that certain effects are linear, that the error variance is constant, etc. – and those assumptions are passed forward into the emmeans() results. Doing separate analyses on subsets usually comprises departing from that overall model, so of course the results are different.
First step: Carefully read the annotations below the output. Do they say something like “results are on the log scale, not the response scale”? If so, that explains it. A Poisson or logistic model involves a link function, and by default, emmeans() produces its results on that same scale. You can add type = "response" to the emmeans() call and it will put the results of the scale you expect. But that is not always the best approach. The “transformations” vignette has examples and discussion.
Inf for the degrees of freedom?This is simply the way that emmeans labels asymptotic results (that is, estimates that are tested against the standard normal distribution – z tests – rather than the t distribution). Note that obtaining quantiles or probabilities from the t distribution with infinite degrees of freedom is the same as obtaining the corresponding values from the standard normal. For example:
qt(c(.9, .95, .975), df = Inf)## [1] 1.281552 1.644854 1.959964qnorm(c(.9, .95, .975))## [1] 1.281552 1.644854 1.959964so when you see infinite d.f., that just means its a z test or a z confidence interval.
Such a situation typically arises in observational data. What this means is that the model is not capable of producing estimates for all points in the reference grid. For example, you have data on combinations of factors A and B, but there are no data at certain combinations of these two factors. If the model includes the A:B interaction, then it can’t estimate those missing combinations. Here are a couple of possibilities:
A + B, without A:B, may be able to estimate the needed combinations.at to focus (at least mostly) on factor settings where estimates are possiblenesting argument. Perhaps the levels that B can have depend on which level of A is in force. Then B is nested in A and the model should specify A + A:B, with no main effect for B.The “messy-data” vignette has more examples and discussion.
As mentioned elsewhere, EMMs summarize a model, not the data. If your model does not include any interactions between the by variables and the factors for which you want EMMs, then by definition, the effects for the latter will be exactly the same regardless of the by variable settings. So of course the comparisons will all be the same. If you think they should be different, then you are saying that your model should include interactions between the factors of interest and the by factors.
This is a common misunderstanding of ANOVA. If F is significant, this implies only that some contrast among the means (or effects) is statistically significant (compared to a Scheffé critical value). That contrast may be very much unlike a pairwise comparison, especially when there are several means being compared. Another factor is that by default, P values for pairwise comparisons are adjusted using the Tukey method, and the adjusted P values can be quite a bit larger than the unadjusted ones. (But I definitely do not advocate using no adjustment to “repair” this problem.)
As is shown in my answer in a Cross Validated discussion, the unadjusted P value can be more than .15 when F is significant (remarkably, regardless of the significance level used for F!).
This probably happens because somehow the lsmeans package or its namespace has been loaded – possibly because some old objects from that package are still in your workspace. Try this:
emmeans:::convert_workspace()This non-exported function removes Last.ref.grid if it exists, converts every ref.grid or lsmobj object to class emmGrid, and unloads any vestige of the lsmeans package. [Note: You may get an error message if there are other packages loaded that depend on lsmeans. If so, detach(package:pkg, unload = TRUE) any that are in search() and use unloadNamespace("pkg") for others; then re-run emmeans:::convert_workspace().]