Видео

They are more prestigious. Check the residual on the y-axis. Anything above zero is respected more than what the model predicts.

@@mronkkoSo it is below the y=x line just means that theoretically on that percentile we would expect the residual to be even higher? What does the theoretical percentile mean? Is it just based on rank

Robust regression is a technique that automatically drops or down-weights outliers. vce option defines how standard errors are calculated. Vce(robust) calculates heteroskedasticity robust standard errors. I personally never use robust regression because I think that I am better at handling outliers than the computer. Outliers can also result from non-linear functional form, in which case regression + diagnostic plots will lead to better models than any automated approach. See my video on outliers. Here is an explanation by Claude: "To explain the difference between Robust Regression (rreg) and vce(robust) regression in Stata, let me break it down for you: 1. Robust Regression (rreg): Robust regression, implemented with the 'rreg' command in Stata, is a method designed to be less sensitive to outliers compared to ordinary least squares (OLS) regression. It uses an iterative process to down-weight observations that have large residuals. The key points are: - It's an alternative estimation method, not just a way to calculate standard errors. - It's particularly useful when your data contains influential outliers. - The coefficients produced by rreg can be quite different from OLS, especially if there are significant outliers. - It uses Huber weights and biweights in its iterations to reduce the influence of outliers. 2. vce(robust) Regression: The 'vce(robust)' option in Stata, which can be added to various regression commands like 'regress', doesn't change the coefficient estimates. Instead, it changes how the standard errors are calculated. Key points: - It uses the Huber-White sandwich estimators for variance. - The coefficient estimates remain the same as in OLS. - It's robust to heteroskedasticity (unequal variance of errors across observations). - It doesn't specifically address the issue of outliers in the way that rreg does. Main differences: 1. Estimation method: 'rreg' changes how coefficients are estimated, while 'vce(robust)' only changes standard error calculations. 2. Outlier handling: 'rreg' is specifically designed to handle outliers by down-weighting them, while 'vce(robust)' doesn't address outliers directly. 3. Heteroskedasticity: 'vce(robust)' is primarily used to address heteroskedasticity, while 'rreg' is more focused on outlier resistance. 4. Interpretation: Coefficients from 'rreg' might be interpreted differently due to the weighting process, while 'vce(robust)' allows standard OLS interpretation of coefficients. In practice, the choice between these methods depends on your specific data issues and research questions. If outliers are a major concern, 'rreg' might be more appropriate. If you're mainly worried about heteroskedasticity and want to keep your OLS estimates, 'vce(robust)' could be the better choice. Would you like me to elaborate on any specific aspect of these methods?"

There is a 50 minute version of that talk on the channel under name "normality assumption" ;)

I am not aware of any research that would show that to be the case. My experience is that PAF gives you a solution more often than ML when the factor model fit is really bad, but I have not seen a connection to normality.

Glad it was helpful!

Please do so. I havent done courses outside Finland after the pandemic, but I did a few before it.

You are welcome!

I have quite a few postdocs who have taken my methods course after their doctorate who have felt the same ;)

@@mronkko Are we supposed to go with the log-link if it works for more than 50% of variables? If we have 7 variables and 4 need logged, we should then use log-link? Do we just log-transform the values if only 3/7 variables needed it? I read your paper but don't see a multivariate situation like this.

@@George70220 So you are saying that it looks like y is increasing linearly with x1 but exponentially with x2?

@@mronkko yes

Sorry for a delay. If x1 has a linear effect and x2 has an exponential effect, you need some kind of interaction of the two in the model. I talk about this kind of models on page 67 of journals.sagepub.com/doi/10.1177/1094428121991907 But as a practical matter, unless you have a good theoretical reason to combine a linear and an exponential effect, I would probably go for a linear model for both effects.

Right. A sample distribution is what is observed and cannot conform to any probability distribution. I am could have mentioned this in the talk, but cut a bit of corners to make it as simple as possible.

​@@mronkkoI think people often mix the following two questions: 1. Is the population mean an appropriate choice for the analysis? 2. Given we have chosen the population mean, can we reliably estimate it? Thank you for helping to disentangle these

@@justintondt5245 Right. I would further add 3) how can we test if the mean is different from another mean or a fixed value.

@@mronkko Yes, good point

You are welcome!

You are welcome. I guess this video met a demand ;)

Short answer: Yes. Long answer: I am not a big fan of winsorization for two reasons: 1) I have never seen it discussed in a really good method book and I do not know the origin of the idea. (I have not looked very hard, though). 2) The procedure may produce erronneous outcomes. Let's assume that we do a survey of people's heights. A short 150cm person accidentally types their height as 1500 centimeters. Winsorization would make the person e.g. 205cm tall in the data. It is still incorrect. If you have outliers, rather than winsorizing the data, check if the outliers are potential data errors and fix where possible. Otherwise, consider deleting observations if they are truly non-representative.

You are welcome.

I was asked to do this topic in Finnish.

Yes, you can use mediation analysis in that case. If your X variables are unordered categories, I would recommend modeling X-M relationships with linear regression. Do you mean your model is X-->M1->M2->Y, or do the M1 and M2 work in parallel representing two different causal mechanisms? If you can justify linear functional forms between the variables, specifying a mediation model is simple. (note that count variables can be modeled with linear regression contrary to common belief, see journals.sagepub.com/doi/full/10.1177/1094428121991907).

@@mronkko Thank you very much for your reply. I will try modelling linear relationship between X and M. My model is X-->M1->M2->Y. Thanks for the references indeed!

@@goncamert7375 If you can model linear between M1-M2, then things are simple. If not, the approach that I explain should still work easily because of linear X->M1

You are welcome!

Do you mean all missing data slides? I can easily provide you slides for a specific video and my goal is to link the slides to every video. However, I do not have a single set that would contain them all in one.

@@dipmalyaroy987 nextcloud.jyu.fi/index.php/s/qXECY3a5GiJKM5B It contains everything that I have on the channel and more. Ultimately, I will have slides up at osf.io/c8qt6/ and linked to the video descriptions, but I have not had time to do it for all the videos yet.

Thanks. I have tried to learn how to pronounce Poisson (French name) correctly. Now I need to add Arellano to the list. On my defence, most non-Finns pronounce my last name incorrectly ;)

You are welcome!

Model non-convergence can occur for many different reasons and it is impossible to say without having access to the model and the data. I recommend going through this playlist to see if you can figure out the problem ruclips.net/video/4_ip9F8EpVs/видео.html&pp=gAQBiAQB

i cannot come up with any specific examples. However, if you search for "parallel trends" "Difference-in-differences" in google scholar, you should find lots of examples.

Good point. Unfortunately RUclips does not allow editing published videos, but I will adjust this on my slideset so that colors will be better if I re-record this ever.

Happy to hear that!

You are welcome. 1) If your research design indicates that the variables should be included, then they need to be included. If they are highly correlated with the IV of interest, this becomes even more important. Otherwise you may be in violation of uncorrelated error term. 2) Yes, linear multilevel models are interpreted like regression. The intercept is typically not interpreted.

You are welcome!

Thanks for the feedback. Can you be more specific: what exactly did you find confusing about it? I will run a course next academic year where I will use this video and I might record it if I know what the pain points are.

You are welcome!

You are welcome. I do not do much software demos because I teach multiple software and because, particularly with Stata, the documentation explains the software well. Understanding the conceptual side is more challenging, and once you know that, specifying the command is not very challenging.

You are welcome!

You are welcome!

You are welcome.

Glad it was helpful!

The "How to" part is software specific and typically you just pick a rotation method from a dropdown menu.

You can do it with paper and pencil. I have colleagues how use Excel for coding and others use specialized software e.g. Atlas.ti

You are welcome!

It depends how SEM is used. In psychology where panels are short, data are often analyzed in wide format where each repeated obsertation is its own variable. See e.g. doi.org/10.1177/1094428117715509 In this case there is just one observation for individual and non-independence due to repeated observations is not an issue. If you model your data in long format where each repeated observation is a new case in your data, then non-independence is a problem if any of the latent variables (including error terms) correlate within individual. If this is the case, then ML is still consistent but test statistics can be misleading. But this can be solved by using cluster robust standard errors and tests statistics.

@@mronkko Thank you very much. Your last sentenc is very impressive. That is "ML is still consistent but test statistics can be misleading. But this can be solved by using cluster robust standard errors and tests statistics". I want to know how to prove the consistency in this case? Or could you tell me the relative reference? Additionally, what software of SEM could provide cluster robust standard errors? Thanks again.

@@xuyang2776 I believe you can find the proof of consistency of ML in e.g. Lehman and Casella book on statistical estimation, but I do not think knowing the proof is useful for an applied researcher. Lavaan, Stata and Mplus, which are the softare that I use, all provide cluster robust standard errors.

@@mronkko Thanks again. What I really want to know is the proof for the consistency even in the case that all the variables are auto correlation in a SEM model. Though I know in a regression model, the OLS estimater is still unbiase and consistent when the random iterm are auto-correlashion, SEM is not same. Additionally, if ML+cluster robust standard erros could resolve the dependency problem in a SEM, why dynamic factor analysis and dynamic SEM are developed?

@@xuyang2776 Dynamic factor analysis is designed for situations where you are interested in estimating effects of latent variables over time. Cluster robust SEs are work in situations where you are not interested in such dynamics, but want to estimate a cross-sectional model with data that are autocorrelated.

This seems like an AI generated post given the placeholders [Your Name]. I think the best way to produce example datasets is to simulate them yourself. The second best would be to look at work by Paul Bliese. He has published a lot on multilevel models including example datasets.

You are welcome.

A random effect is a latent variable. A categorical latent variable is often referred to as a latent class. What you are looking for is a latent class model where the latent groups are allowed to differ on their path coefficients. Stata does latent classes, see www.stata.com/features/overview/latent-class-analysis/ SPSS does not.

System GMM requires mean stationarity. See doi.org/10.1080/00036846.2018.1540854

You are welcome

You are welcome!

You are welcome!

This is a really good question. You need to think about what the two groups are. You have everyone who did not serve, but only a subset of those that did. Therefore you cannot generalize to the full population or those that did serve. Here is a GPT4 generater explanation: The correct answer is c) "The average earnings effect of serving in the army for those who did not serve." This might initially seem counterintuitive, but it's based on a fundamental concept in econometrics known as the "Local Average Treatment Effect" (LATE). In the scenario described, there is a random assignment of military service exemptions through a lottery. This randomization ensures that, on average, the groups of exempted (treatment group) and non-exempted (control group) applicants are similar in all respects except for the treatment-here, serving in the military. This similarity is crucial because it mimics the conditions of a randomized controlled trial, allowing us to infer causality from the comparison between the two groups. Now, let's break down why each answer choice is what it is: a) "The average earnings effect of serving in the army" suggests we are looking at the impact on all individuals, regardless of their inclination or circumstances that led them to apply for the exemption. This is not what the comparison would reveal since the analysis only includes individuals who applied for the exemption and were subjected to the lottery system. b) "The average earnings effect of serving in the army for those who served" might seem like a reasonable answer, but it's not the focus of this comparison. This is because the analysis isn't solely focused on those who served; it also includes those who were exempted. The aim is to understand the impact of not serving (being exempted) versus serving. c) "The average earnings effect of serving in the army for those who did not serve" is correct because the analysis effectively compares individuals who wanted to be exempted (and thus, by extension, did not want to serve). Those who win the lottery (and are exempted) serve as the treatment group, and those who lose (and thus serve) are the control group. The comparison then reveals the effect of not serving on the group that applied for exemptions but had to serve due to losing the lottery. In essence, this setup allows us to estimate the impact of military service on those who would have preferred not to serve but were compelled to do so because they did not win the exemption lottery. It's a subtle but important distinction in understanding the causal effects in this scenario.

First differencing does eliminate beta_0, but there is an error in my slides. My video on first differencing has the correct equation ruclips.net/video/hQWSh_j3Oy0/видео.html

I do not understand the question. What call are you referring to?

You could use t for testing a single constraint. But in practice we use F because it can handle both single constraint and multiple constraints. A single parameter F distribution is simply a square of the corresponding t distribution.

@@mronkko thanks!! Then why in some cases we should use a wald test instead of a t test like in this video?

@@junbeombahk3668 I am not using a t test in the video. The test is a single parameter Wald test. See this for an explanation of the tests ruclips.net/video/AbhwpFX2Xdw/видео.html

No, I meant that we measure the characteristic of each individual after the treatment. We cannot measure the effects of treatment. We can only measure the characteristics of interest and then use these measured characteristics to calculate an estimate for the average treatment effect.