Відео

Hi! This is a great question! Sorry for the delayed response; I needed to look into how `ggeffects()` implements interactions. You are correct that the plot here is additive; it's not an interaction. So, instead of the predicted probabilities for `scot` being based on `trust` and `age` set at their means, the predicted probabilities are when trust=1 and age=mean(age), trust=2 and age=mean(age), etc. If you do an interaction between `scot` and `trust` in the model, the predicted probability plot will look similar but not identical. If your interest is on capturing the moderating effect of a second variable, then an interaction in the model and then plotting the interaction for interpretation purposes would be better. You might want to check out the ggeffects interactions info: strengejacke.github.io/ggeffects/articles/ggeffects.html

​@@qssd Thanks a lot for explaining! What I ended up doing to visualize an interaction was using practically the same code you shared, but instead of setting an additive model in ggpredict(), I set it to a model with the interaction term in it.

Makes sense!

The simplest way is to add a factor version of the time variable as a predictor in the model (e.g., `as.factor(year)`). But, if you have proper time series data, you probably want to explore a time series count regression model (e.g., Poisson autoregression).

Hi - thanks! Do you mean like for a binary outcome variable, where you would normally use logit or probit? It doesn't appear it will work for that. I took a look at the package (`performance`) and function's code on Github, and it should throw an error if it's not a checking a count distribution. I wouldn't be surprised if there are R packages/functions that allow you to test if there are too many zeros in a binary outcome variable, but I don't know them off-hand. For binary and ordered outcomes, if I'm worried there might be too many 0s or 1s (or similar for ordered case), I'll run a model with that assumes an asymmetric error distribution (e.g., complementary log-log, skewed logit), compare AIC/BIC values with the logit/probit versions, and choose the model with the lowest AIC/BIC.

Thanks! If you are overfitting zeros using poisson regression, I would first use a negative binomial regression. If you are still overfitting zeros with NB regression, I would next use a hurdle model (probably the NB version) b/c of how it truncates the count data for only positive counts. That way you should get a better understanding of the relationship for 0 counts and positive counts.

Hi - sorry, I don't typically work with data used in a multivariate probit (i.e., correlated binary outcome variables) and I don't have an example queued up.

Hi - at present, there isn't a free version of the book; I don't know if there will be. Thanks

Hi - the Word file is just automatically generated by knitting the .Rmd in RStudio. So, there isn't anything special about the Word file; and also I'm not sure whether I still have it. Thanks

Hi - I'm not entirely sure, but I have a couple of ideas. #1, I noticed in your ggpredict() specification that you have 'termes' instead of 'terms'. I assume you are using the French for terms or this is error. Does ggpredict work with French? Have you tried "terms"? #2, have you tried adding `as_tibble()` on its own line between `ggpredict()` and `ggplot()` (with using the piping operator %>% to connect the lines)? #3, it looks like the error is saying it can't use an NA value. Does the variable "Period" have any values that only have NAs as observations? For example, all the observations for Period=2 are NA. It is also possible that NAs were created for in the ggpredict() function. For example, if the ggpredict() function didn't work correctly it may have produced NAs instead of numeric values. Have you checked the output when you run the ggpredict() function? #4, another thing to try is adding `filter(!is.na())` before `ggplot()`. So, and adding #2, ggpredict(Ado1, terms="Period") %>% as_tibble() %>% filter(!is.na()) %>% ggplot(...) (the ... is all the ggplot code). You can then see what gets plotted and see if it looks wrong. This might help you figure out the problem.

@@qssd Thank you so much. It finally worked

It's from the `performance` package

Hi - thanks for the question. Unfortunately, I'm not sure I have a great answer. If I understand correctly, if for each pond the numerator is number of fish with disease and the denominator is total number of fish in the pond (i.e., the population in the pond), can you just create a ratio or % and use something like OLS? Otherwise, if you have two related counts as outcomes, you can use the 'bivariate' versions of count models (e.g., Famoye 2010). I'm not very familiar with these models b/c I haven't needed to use them but you can find info online.

Hi - good question. Honestly, I usually don't worry about confidence intervals of odds ratio as odds ratio is just a way to interpret the coefficients; where confidence intervals matter for the coefficients. I don't know an easy way per se. Since we use `exp(\beta)` to get the odds ratio value, my best guess is that we can use the `confint()` function to get the 95% confidence intervals of the coefficients and then use the `exp()` function. For example, `exp(confint(model.mlogit, level=.95))` will give you the odds ratio values of the 95% CIs of the coefficients.

Hi - I'm not sure why the difference, but I imagine it's something small w/ how they are calculated (e.g., a control variable is held at mode in one and mean in another). But, check out these new videos that I have on this (much easier code): ua-cam.com/video/qKchFtTuaBE/v-deo.htmlsi=7V2QQJC6OT15luG4 ua-cam.com/video/0-kSeGPHMFk/v-deo.htmlsi=Qob26xkMM_7Bh25m

Hmmm. Generically, the error means you are trying to do something that requires numeric vectors (etc) but you are using non-numeric vectors (etc.). I don't think that matters here. My guess is that the patchwork package is not loaded (`library(patchwork)`). I got the same error as you when I didn't first run `library(patchwork)`. So, double check that.

You are very welcome

Yeah, that's a different kettle of fish. I don't have direct experience running that in R; besides from a two-wave panel survey where the time aspect is essentially irrelevant. But, the ordered probit part is likely the easiest part. The harder part is correctly specifying the errors of the panel overtime, i.e., fixed vs. random effects, etc. I'm sure you can find an R package that does this -- just google it. Sorry can't be more helpful.

There is a link in the video description --- the line starting 'April 2023 Update:'

You're very welcome!

Thanks for the comment. The `glimpse()` function is used in this chapter in the 2nd edition. The function is from the `dplyr` library, but simpler to use `library(tidyverse)` to load the core tidyverse packages at once.

Awesome! Glad they were useful.

Terrific, thanks so much!

There is a bit of tricky nuance. I use this example - Think of the odds as an event occurring vs. not occurring, and the odds ratio as the odds an event occurs for one group vs. another group(s). In the regression context, we are comparing different groups (our predictors) for an event occurring. Extending beyond 'groups' (i.e., nominal predictors), Long (1997) shows that for ordered predictors exp(\beta_i) is the one-unit change in odds, which is the odds ratio and not odds. I prefer to reference logit coefficients as providing 'logits' and not 'log of odds' to reduce confusion.

@@qssd Thanks.

Mainly, if you want to use odds ratio for interpretation you have to use logit. Otherwise, it is personal preference --- only difference between logit and probit is assumption about error distribution, and almost always get same results.

@@qssd Thanks.

Hi - The datasets, as part of the digital resources, should be available on the book's website sometime this month. When they are available, I'll include the links on the channel. Sorry they're not available yet.

@@qssd Thanks! I've just received my Kindle version so looking forward to tackling these exercises! :)

The data and codebooks are now up: study.sagepub.com/Fogarty2e

@@qssd Thank you!!!

awesome!

Thanks. You're correct about the data location. It appears that the site has been overhauled. The data used here is the 'microdata teaching file' from the 2011 Scottish Census: www.scotlandscensus.gov.uk/documents/microdata-teaching-file-and-user-guide/ Some of the variables in the dataset used in the video were recoded from the original variables in the downloadable data.

@@qssd Thanks for your quick reply.

That's right - the '2' refers to the 2 values of gender (which are the lines to plot). The answer of what to do when you have two factor variables comes down to what you want to plot? For this type of plot, I believe the x-axis variable must be numeric and you can only have one variable as a factor that will generate the different lines in the plot. So the number to put in will be how many levels the one factor variable has. If your second factor is just in the plot as a 'control', like marital_dummy, you might need to convert it to a numeric or you could try setting to one of the levels (like the modal value); but I haven't tried it. An alternative way to plot these is using the `ggpredict()` function from the `ggeffects` package. The `ggpredict()` function essentially does all of the code prior to the ggplot() code and then you can specify all the ggplot() code and link the two using the pipe operator %>%

@@qssd Thanks! I converted the gender variable to a numeric one because that is not the one I wanted to plot but the three-level factor variable is. Works perfectly now!

Thanks so much!

Wow, thanks!

You're so welcome!

You could make them explicitly ordered categorical or numeric. Here, you will get the same regression coefficients with either.

@@qssd strictly speaking, you would have to recode them into dummies to include them in a regression model as independent variables, or not? You are treating them as numeric, which I don't get- is this accepted as a quantitative social science practice`?

@@kasberge7164 You would only dummy them if they are nominal-level variables. Edu and scot id are ordinal-level variables, so the simplest option is treating them as numeric. You would only recode them as dummies if wanted to create a new version(s) of the variable -- for example, scotID = 7 (so, '1' in a dummy) vs. scotID=1-6 (so, '0' in a dummy).

Good question. Although odds are related to probability, we often don't think about them the same as probability (or predicted probability). Generally, we can think of odds as the "chance" or "likelihood" of an event occurring. Here, the greater the odds, the more likely an event will occur.

Glad it helped!

Wow, thanks!

The error is likely referring to a mismatch in what the factor levels are currently and what you are trying are referencing in the code. So, it might be they are 1 and 2, but you asking for levels 0 and 1. Try changing the code to '1:2' or recode the factor levels prior to the plotting code. (Sorry for the late response, you might have figured this out already).

Sorry for the late reply. For this type of plot, the x-axis needs to be numeric in some way (so, a variable that is ordered with at minimum 3 values). You would have to use a different plotting technique to plot just factor variables (that is, nominal-level variables).

@@qssd thank you !

I don't think you can with the coefplot() function. It also would defeat the purpose of this plot as the dashed line at 0 provides info on statistical significance; most of the other coefficient plotting functions I've seen have some version of the 0 line. The coefplot() manual says you could get rid of the line with zeroType=0 as an option. Otherwise, you could try to do this plotting using ggplot2.

Hi - if your outcome variable is continuous then you should use linear regression and not logit regression. If you mean a predictor variable that is continuous, then the odds ratio interpretation is done in the same way as for 'general_health' in the video. The one possible difference is that if the continuous predictor's values are meaningful then you can use them instead of the generic 'unit'. For example, if the predictor's values are raw dollars, you can still say 'for a one-unit increase', but you can also say 'for a one-dollar increase'.... You just need to make sure you know the unit of measurement (e.g., raw dollars, millions of dollars, etc.), otherwise you might mess up the interpretation.

Thank you! Cheers!

I also would like to know, how to interpret the coefficients. Can I just use the "(exp(model.1$coefficients[-1])" command to get the odds-ratio? is there an opportunity how I could get the marginal effects? And is it also valid to use the "scale()" command to standardize the coefficients? Thank you very much for your videos, they helped a lot!

Yeah, sorry, I haven't had time to create new videos on the interpretations. To get polr to run with a factor outcome variable, you need to classify it as ordered (e.g., as.ordered(as.factor()) ). The interpretations, though, should be the same for a numeric or ordered factor outcome variables.

Yes, you can use the same code for the odds ratio. The difference for interpretation from the binary logit is that the odds need to be discussed as 'more', 'increased, 'greater', etc. instead of a specific outcome. This is b/c ordered logit uses cumulative odds ratio and so the odds value is the cumulative odds of a lower to higher outcome. For example, 'for a one-unit increase in education, the odds of *greater/increased/more* trust increases by a factor of..... I don't usually work with marginal effects, but there is a lot you can do with predicted probabilities. You should be able to use the scale() function, but I haven't thought through the impact on the interpretations... Thanks.

@@qssd Thank you so much for the fast answer, it helped a lot!

Welcome!

Agreed, it is burdensome. You can use tidyverse functions mutate() and recode() to label the factor levels then pipe it (%>%) to the ggplot2 code, but it still requires you to write-out the labels at least once.

Glad it was helpful!

Thanks!

Yes, you are right. Pseudo R2 is based on log-likelihood values.

You're welcome

Thanks! Sorry for the delay. If the dependent variable is ordinal, then either ordered logit or probit work equally well. There is no real difference. The choice is usually what you prefer like the choice between binary logit or probit. The one caveat is that, obviously, if you use ordered probit then you can't use odds ratios for interpretations.

Thanks! Sorry for a delayed response. I haven't had the time to create more of the videos, unfortunately.