Longitudinal Analysis - Science topic

did you calculate eventually? need guidance

This is an excellent site

includes SPSS syntax , that accompanies the book

Longitudinal Analysis: Modeling Within-Person Fluctuation and Change (by Lesa Hoffman)

Hello Pascal,

I'm a bit unclear about your query. Do you mean that a batch of cases has exactly the same score on each predictor? If so, omitting them would bias estimates of score variance on the variables (and standard errors) and most likely the regression coefficient estimates associated with them.

Aside from the fact that you might find these instances inconvenient, do you have a substantive basis, whether theoretical, or empirical, to warrant jettisoning such data? (For example, do you have after-the-fact information to suggest that such instances were not accurate portrayals of the true values for individuals?) If not, it sounds as if doing so would be ill-advised.

Good luck with your work.

@ Wesley Kung ,

Demographics could be considered as the factors, however, psychological factors and other factors are also could be predictors of mental health.

Kind Regards,

The timescale required for a longitudinal study to demonstrate effects depends on a number of factors.

These include the phenomenon under investigation, and the comparative data already available - there may be comparison data from studies of the effects of other approaches against comparable non-intervention controls and you are looking at a novel intervention. In this context, the timescale for effects may already have been already established.

Change may be being evaluated against a stable baseline where the phenomenon is not otherwise expected to change over time and there is data to substantiate this.

Some phenomena have an expected chronobiology - for example changes in menstrual-period related phenomena may take longer before it can be established whether change has taken place against a monthly pattern.

Epileptic seizure and migraine frequency can vary markedly over time making a stable baseline more difficult to establish.

Changes in seasonal affective disorder, asthma, and certain allergies show strong seasonality and annual patterns some of which are hemisphere, day length and latitude-dependent and might require several years of data.

In short, for a research study it depends on what is known about the H⁰ phenomenon you are studying and its variability.

Hope this helps.

Ken

You are looking for the functions merge and reshape. Lots of packages, like data.table, also have functions, but merge and reshape are the ones in R. Just look at their help files.

No, cause-and-effect conclusions cannot be made from subgroup analyses or meta-regression analyses of potential moderator variables. These analyses are done across studies, not within studies where cause-and-effect conclusions can be made (e.g., treated group vs. untreated group).

Hi,

Repeated measure analysis. Here are a few references:

Schober P, Vetter TR. Repeated Measures Designs and Analysis of Longitudinal Data: If at First You Do Not Succeed-Try, Try Again. Anesth Analg. 2018;127(2):569-575. doi:10.1213/ANE.0000000000003511

For atheoretical insight try: Get this book in print

M. Ataharul Islam, Rafiqul I Chowdhury.Analysis of Repeated Measures Data. Publisher-Springer Singapore, Year: 2017

ISBN: 978-981-10-3793-1, 978-981-10-3794-8

Why you needed to collect data at three different time periods? Your answer lies in there!

You need somthing more than standrad linear models to analyse longitudinal data. You dependet variable whill be time varying but gdsex wil be time invariant and applt to teh peson. The natural method of analsyis is the two level random effcets model will can hadle within individual and between individual variation.

This is a short practical example of what can be done

this is an overview of alternative models

this is a comprehensive site that considers a range of software :

https://www.pilesofvariance.com/

It looks good except that I would code "event" as 0 and 1 and "wave" as 0, 1, 2, 3, .... Doing so will facilitate your interpretation of the regression coefficients (in particular, the intercept) because then both independent variables have a meaningful zero point.

For general orientation you may be interested in

In this manual , see Chapter 8 where we use another multilevel model to get precision-weighted estimates of the group means

https://www.researchgate.net/publication/260771330_Developing_multilevel_models_for_analysing_contextuality_heterogeneity_and_change_using_MLwiN_3_Volume_1_updated_March_2019_Volume_2_is_also_on_RGate [accessed Aug 18 2021].

For a fuller discussion of the multilevel model as a measurement model) and a more convincing example of using precision-weighted estimate of the group mean can be found at

https://www.researchgate.net/publication/252146040_Do_multilevel_models_ever_give_different_results

Finally , as in all research call involving a judgement call, do it both ways and see if it makes a substantive difference.

If you have an ANCOVA style model with pre-score as a covariate it isn't repeated measures. If you use a repeated measures model in the ANOVA context this is equivalent to an analysis of the difference scores in a non-repeated measures model.

So if you just have two time points it isn't always necessary to use a repeated measures model.You can see this with something as simple as a paired t test where it is equivalent to a one-sample test of the differences between pre- and post-. So there is no problem using an ANCOVA style model if you'd like to (it may also have greater statistical power than the standard approach) and arguably makes less strong assumptions.

Well, the cohort study is highly suggested in covid pendamic. however, the covid-19 is no more a cohort situation it is done for observation and per and post-test by training and other medication.

Hi Ying, please see our paper to read about how what we eventually did:

Contribution = the new knowledge you are bringing in

Implication = Benefits from the derived knowledge

Hello Alessia, I would use mixed models with "subjects" as random intercept and "gender" as a fixed effect: mixed models allow you to use eventual alternative distributions in case of your dependent variable is not normally distributed and they are less sensitive to missing data.

Anyway, the different time points are so unbalanced, in my opinion, and I would suggest considering just T1-2-3.

You can repeat the same analysis including T4 and T5 only on participants included at T4, so with a reduced sample.

This seems to be the only reliable solution, in my opinion, but I'm curious to know whether someone else has some best suggestions.

Good work and good analyses!

I do not have experience of this level of complexity in repeated measures , but it seems to me that you actually have multiple episodes - what you call events. So I am suggesting a three level structure with an identifier at level 3 for individuals; at level 2 for episodes within individuals and level 1 for repeated occasions within a specific episode; the minute by minute recording of repeated measurements. You could have variables measured at all three levels. The general approach is considered here

Fiona Steele (2008) Multilevel Models for Longitudinal Data, Journal of the Royal Statistical Society. Series A (Statistics in Society)Vol. 171, No. 1 (2008), pp. 5-19

https://www.jstor.org/stable/30130728?seq=1

Eoin Finegan , the IDRE site at UCLA also has multiple examples. See this page, for example:

Scroll down to Exercise example, model 2 using MIXED Command. You can also find examples from the well-known book by Singer & Willett on this page:

HTH.

Hi Andrey,

If you have balanced panel data, the proper test for this case is Dumitrescu and Hurlin (2012) for testing Granger causality in panel datasets. You can install this -function/ code by - ssc install xtgcause- . Then, you may try this command

xtgcause Y X, lags(#) and the other way around

xtgcause X Y, lags(#)

--------------------------------------------

H0: Y does not Granger-cause X. H1: Y does Granger-cause Xat least one panelvar (id).

----------------------------------------------------

For more information see these papers:

1. https://www.unine.ch/files/live/sites/irene/files/shared/documents/Publications/Working%20papers/2017/WP17-03_V2.pdf

I hope this helps.

I am assuming that you specifically want to compare a standard random intercept model (participants allowed their own intercepts only) to a standard random coefficient model (participants are allowed their own intercept AND slope). In other words we are not talking about other types of models (e.g. shared intercept, but random slope)

There are two basic approaches to choosing between these two models.

The first approach (purely empirical) is to compare the model fit of a random intercept model to that of a random coefficients model (i.e. random intercepts AND random slopes) using something like AIC or BIC. The advantage of these measures (and other IC statistics) is that they penalize for model complexity. The basic idea is that if the AIC (or BIC) is lower for a random coefficient model compared to a random intercept model then the gain in fit is worth the extra model complexity.

Personally, I think this approach alone is a little too 'black box'. I like to think about the problem in terms of the data (and research problem itself). For a random intercept model we are assuming that participants are allowed to have their own baseline values (and these are normally distributed around the average intercept). BTW remember this is AFTER we account for differences in participants due to treatment group and/or risk factors (i.e. we can assume the fixed effect predictors don't interfere with these decisions). The subtext for the random intercept model is that we assume (despite there different starting points) that participants respond to time in exactly the same way (in other words, all of the individual participants' regression lines are parallel). So the important question is: Do we think such an assumption is valid? If we assume that participants may respond to time differently (e.g. different patients may respond to therapy in a different way...remember though, beyond that explain by the fixed effect predictors in the model).To summarize, if you had to fit a regression line individually for each participant, would you constrain them to a share slope (random intercept) or not (random coefficient). In other words, have look at your spaghetti plot.

Hope this helps

Thanks again for your response Kelvyn. I only could ask for one measure after the program. Defitively 1 week is a short time. I have doubts between 1 and 3 months. The company wants to evaluate 1 month later for having the results as soon as possible. However, as we want to measure the real effect of the course in the attitudes and behavior of attendants I don´t know if one month is too little time. What do you think?

Qualificative methods and case study research are by far the strongest ways to study causality

Hello Robert, there are extensions of recursive partitioning and trees for longitudinal and clustered data. They essentially include a mixed model element into the algorithm. I have used the RE-EM algorithm in the past (see DOI: 10.1007/s10994-011-5258-3 and DOI: 10.1016/j.csda.2015.02.004). There are also binary partitioning for continuous longitudinal data (DOI: 10.1002/sim.1266) and mixed-effect random forest (DOI: 10.1080/00949655.2012.741599). Implementations can be found in R packages: REEMtree, longRPart2, MixRF.

Kelvyn Jones thank you so much for your help. I like the idea of building the model stepwise and kind of adding more complexity to it step by step. I'll do some more research to see which procedure fits our data best. Now I have some starting points for finding out, which is great. Thanks!

Should I combine the daily measurements into weekly measurements

NO you are loosing valuable information

Should I use the time variable (measurement oder week) as continous or as a factor?

A bot of both

In a repeated measures random effects multilevel model you have occasions at level 1 nested within individuals. The residual at the individual level will give 21 differences one fro each individual around and intercept and the 98 by 21 occasion residuals will give differences from the individual overall mean (intercept plus individual residual) for each and every occasion.

You can then put in variables to try and account fro this differences in the fixed which may be a function of continuous time ( linear or quadratic etc to get the overall trends), categorical time (a dummy for weekend or morning) and dummies distinguishing the three groups. You may also need interactions eg a linear rend by group interaction but 21 is not many for this type of modelling.

It is also usual to fit a random-slope model whereby the differences between individual may increase over time. A potential sequence of models is given in my answer to this question.

Let me see if I understand. You have a single variable as an outcome, and a bunch of sets of covariates, where each set is a single variable with multiple measurements. And then there is the SD of something. First, what is this SD of and is it there to address heteroskedascity? On the sets of covariates, is the issue how to deal with the collinearity within each of these sets? Are you expecting similar covariances within these sets (and are the times the same)? Are you after something like estimating the slope and mean of each set, and then using these? Or wanting some more model-ish thing? How many variables do you have, how big a sample, how many time points? Are you testing a specific hypothesis or just trying to find a model with good fit?

OK, have a look at this site

http://www.bristol.ac.uk/cmm/software/mlwin/mlwin-resources.html#multipleresp

and especially straightforward analysis

http://www.bristol.ac.uk/media-library/sites/cmm/migrated/documents/repeated-measures.pdf

and also this site

http://www.pilesofvariance.com/

اختبارات البارامترية

Modern legal researches assume the appeal not only to classical methodology, but also postmodern. Not only the received result, but also its interpretation by the author is important.

Thank you Gregory, I have 224 schools, grade 8-12, been rated since 2011. I want to exam if the authority performance review is worth it and relevant to improving the country ranking.

I am doing a meta-analysis some studies have OR while others have RR. Study design is case-crossover or time series. I use Comprehensive Meta-analysis Software version 2.0 for analysis. How can convert OR to RR?

I do recommend " Applied Longitudinal Analysis, 2nd Edition"

by Garrett Fitzmaurice, Nan Laird & James Ware

Yes mixed or multilevel models are quite forgiving about imbalance and missingness. That is they do nor require MCAR abut only that the response is MAR. However if the missingness is due to the response variable (ie NMAR) then the estimates will be biased.

here is a straightforward annotated example

This is an outstanding site for beginners (with lots of different software)

Md Washim Akram Emil Hristov Yankov Alain Goriely

Thankyou very much!!

Yes. You can set your DVs to match Your time frame periods and apply a ANOVA.

You are about to discover the concept of incidence density! This is the rate of events per unit time. In your case, the rate of events per 100 or 1000 person-years.

You can tackle the problem using Poisson regression, with length of observation set as the exposure time variable.

Alternatively, you can treat the data as time-to-event data with repeated events. This has the advantage that the probability of chest infections probably rises with age. You can use age as the time variable in the analysis, with subjects entering at the age they were first seen and exiting at the age of last follow up. In order to avoid immortal time bias, you need to declare them to be at risk from some point, say from age 18.

This is pretty straightforward in Stata, if you have it.

I did something like this some time ago. I just did multiple linear regression of rate of FEV1 loss ( FEV1 at time x - initial FEV1) / time between measurements) (dependent variable) vs age, height, dust., smoking,alcohol etc (indep variables). you could easily add sex.

see attached paper

You may want to look at this Proect on Research Gate

you will see that we are very specptiacl amout autmatic fail safe proecures, but this shows what can be done

I think my best advice here is to try to plot out or describe the data in a way that you would if you were trying to convey the results to an audience. This will guide you as to what analysis you want to do. What would you want to point out to your audience if you were giving a talk presenting your data?

As you mention, you might just report average date or average practice time to some performance benchmark, perhaps several benchmarks. Reporting these plus some measure of dispersion ---- range, IQR, or standard deviation ---- might be the best way to convey what you want to convey.

Or maybe you want to plot the curve of performance vs. practice time. As David Morgan points out, with repeated measures, the statistics here can be a little complicated.

I think to answer your question, you probably want to find a way to present the data from all subjects together, not treat them individually. This gives some sense of the variability among your sample. Although, with 12 subjects, you can pull out a couple cases that exemplify some features you wish to highlight.

Hi Rowan,

Since you are likely to have more measurement points later and you plan to adjust for multiple covariates, the mixed model (or generalized linear mixed model, to be precise, because of the categorization) framework seems to me to be more useful in the long run to get familiarity with the model. Mixed model also permits you to test various contrasts rather easily considering that you have more than one outcome.

About your research question: you could potentially have a number of research question in mind with this setting, but I assume it is the usual: is there change over time points in the effect of ecstasy use on the outcome (when adjusted for the covariates). Is this it (i.e. are you interested in the time-ecstasy use interaction)?

Another approach might be to compute the differnces and presumably to use the linear model for analysis, but since you have categorized variables this depends on how the categorization was done (did you use same cut points in each time points etc.) and if you are happy to report the rank-based results. Also, in this type of analysis it is difficult to accommodate the other potential outcome variables in the same model.  

But a few questions about your variables: Is ecstasy use a dichotomous variable?What do you mean by not normally distributed? Are the outliers, skewed distribution or what. I would advice against categorizing - you will loose information from the varibles and also power from you test statistics. Categorization is also an easy target for reviewer criticism. Instead, skewed distributions can be easily treated as log-normally distributed or you could treat them as censored varibles to account for floor or ceiling effects (the latter is easier in the framework of what I suggest in structural equation modeling approach below). If you are adamant about the distributional assumptions, there are also non-parametric longitudinal models that you could consider. I have constructed some scripts to run such models on R.

If you would still like to proceed with the categorical data, then the number of categories in the variables is important, because it directly relates to the kind of model that is useful to consider. So why would you consider multinomial logistic model the best for your data and not e.g. the ordinal logistic model? I would imagine this results from having more than two categories in the categorized variables and you may expect that the predictors have different coefficients for the outcome category pairings. If you have just two categories then binary logistic model would suffice.

Any of the cases described above can always be analysed as a mixed model (continuous variables) or generalized linear mixed model (other outcome types). You can then include the two outcomes (e.g depression and QoL) in the same mixed model, if you have them as outcomes and specify no relationship between them except the within- and between-outcome covariances. Then, you just need to be careful in speficying the variance-covariance matrices for the outcomes. Mixed model is also better than alternatives models (except for its 'easier' cousin, generalized estimation equation (GEE) approach), because it permits specifying arbitrary covariance matrices, so it is not dependent of the assumptions of constant covariance and variance among the time-points like the repeated measures ANOVA. Since the covariances are part of the mixed model, it is fine if QoL and depression covary. However, if you have more than one or two variables as outcomes over time points, the covariance specifications may become a tedious job to do.

If you wish to model more complex structures among the outcomes, then I would suggest that you get familiar with structural equation modelling which permits more flexible models between outcomes. You can also test more specific hypotheses regarding the relationships between the outcomes. For example, software like Mplus are very useful for models like this. It is also easiest to learn to use in comparison to the others (see statmodel.com for further information).

Finally, about QoL. Which QoL measure is it? How to link the different measures depends heavily on the kinds of items presented to the subjects, but also on how you think the subjects perceive the items when they take the items. I guess you are in the best position to know this since you know the measure and persons who responded to it. Perhaps you could build some kind of an ordinal index from the QoL measures that could connect measures at 15-17 to 18+? It is difficult to answer without  more detailed knowledge about the Instruments.

Regards,

Timo

Thanks Ajit. I saw the IDRE website before but it doesn't totally apply to my situation here. It has been confirmed that stata cannot run multilevel model using BRR weight and I think neither does R or SAS can do that.  

Without seeing the article, my guess is that they had repeated measures within patients (or some other unit), so they used GEE to model to correlated nature of the data.  I suspect that their use of the term, "random effects", refers to the repeated measurements wiithin patients or units.  GEE does not model random effects, rather considers the clusters or units as nuisance parameters, used only to account for the lack of independence among observations.  Their approach to the data analysis, however, may be satisfactory -- only their description confusing.

Paul

This online course which is free covers logit modelling

You can follow it in MLwin Stata and R 

these are the current modules

Modules

Using quantitative data in research (watch video introduction)

Introduction to quantitative data analysis (watch video introduction)

Multiple regression

Multilevel structures and classifications (watch video introduction)

Introduction to multilevel modelling

Regression models for binary responses

Multilevel models for binary responses

Multilevel modelling in practice: Research questions, data preparation and analysis

Single-level and multilevel models for ordinal responses

Single-level and multilevel models for nominal responses

Three-level multilevel models

Cross-classified multilevel models

Multiple membership multilevel models

Missing Data

Multilevel Modelling of Repeated Measures Data

Cox regression algorithm will be applied

The 'MEDIATION' package written by Hicks & Tingley may be more up-to-date than 'medeff'. Type "findit mediation" or search for it under the help or use the link below.

Hi David, I think you need the years before the minimum wage too because with DID, you are not only interested in groups with and without but also their differences with respect to time.

My only challenge is whether you distinctively have two groups -- those who benefited from the minimum wage and those who did not.

Hello Fabrice,

Sorry for my late response. I was pleasantly surprised because of your really immediate feedback.

Thanks a lot for your support. You understood it very well :)

With you help I found other equations (which looks a little bit easier to me to understand).

I hope they are more or less equal (but I really don’t know).

But in these equations I need always μ (I think the mean of the population). But this is unknown. On the other hand my N (population size) is very large (but unknown, too). For this reason I can remove (N-n)/N.

My sample size (n) is about 60.

Nevertheless I don’t understand the equations of the link you gave me fortunately. Might it be possible to explain it with help of the example? (But please don’t feel forced!)

Many thanks again!

This might be helpful on the Lord's paradox angle.

Thank you so much dear Mr. Weaver.

Thank you so much, for your generosity sir...

Have a look at the work done by Ross & Rawls (Arbitrage Pricing Theory) and some of the stuff taught by Damodaran.  Fernandes each year asks all accounting professors for their veiw of the risk premium (but that's hardly logical to ask people who never have any money to invest).  Logic dictates that there is a positive relationship, but the issue goes back to how do you measure risk - a single Beta is hardly a good measure and probably does not account for more than about 30% of the answer.  I think you need to look at some of the work on behavioural finance to see that, unlike the assumptions behind CAPM, investors are not really rational - and without that assumption (amongst others) there can be no linear relationship.

Dear Philipp,

thanks for correctly completing my question ;)

And even more for your answer. I totally agree, but it feels good to be more confident now.

All the best,

Johannes

Hi Barbara

You can use the Do command within a macro to do this.  The syntax below should do what you are looking for.

Andy

*************************************.

*** Part 1 - Define the macro ***.

*************************************.

* You can change the parameters of the regression command as required (e.g. plots, residuals etc).

* Highlight and run Part 1 to define the macro.

DEFINE ManyDVs  (DVs = !ENCLOSE("[","]").

!DO !I !IN (!DVs).

REGRESSION

  /MISSING LISTWISE /STATISTICS COEFF OUTS R ANOVA

  /CRITERIA=PIN(.05) POUT(.10) /NOORIGIN 

  /DEPENDENT  !I  

  /METHOD=BACKWARD CerebralWM_FD CortexVol

                                      Mean_GM_temporal_lobe_volume sex anniscolarita 

                                      EstimatedTotalIntraCranialVol age.

!DOEND .

!ENDDEFINE .

**********************************.

*** Part 2 - Call the macro ***.

**********************************.

* Replace the variable names in the square brackets with your own variables.

* Highlight and run Part 2 to produce the regression analyses.

ManyDVs DVs = [mmsecorretto **add other DVs here**].

There are different views about the effects of complex sample designs in regression estimation. If, as in the link you provided, you are estimating characteristics of a single population such as means or proportions you should certainly take account of the survey design. For the relationship between variables it rarely makes much difference. Sampling rates will have similar effect on both explanatory and response variables and will have little effect on the regression. You could try using the usual sampling weights and compare results.

You can collapse levels according to their ontology (applied area). For example, There are different variants of disease. some of them are rare. You can join rare variants into one group (level of categorial variable).

Of couse, David's book recommendation is very valuable.

try to use the average of the three as 2 other variable to be 5

I agree with both Cauane and Georgio above.

You are dealing with a multi-level analysis of panel data with 4 repeated measures.

When it comes to the wonderful stats package, R, it's a fair bet that someone has faced a similar problem and shared their solutions. See the link attached: Multilevel analysis: panel data and multiple levels.

Thank you again for your answers. Unfortunately, I am still struggling with my data. I do not want to bother you anymore with my questions, still trying to tell you what I intend to do now.

Because we measured self-concept as a latent variable  by several indicators I think I need to do a two-level CFA controlling for measurement invariance between the two levels.

If this works I realized that I probably have to run a contextual effects model. The reason for this is, that I have on both individual and class level the same predictor (testscores) and want to see what impact class means of test scores do have on self-concept beyound individual test-scores (see Marsh et al., 2009).

At the moment I think I will then just try to show

(1) that there are contextual effects  of class achievement at t1 on self-concept at t1 for both groups of students (those with and those without transition, in seperate models).

(2) that there is a contextual effect of class achivement at t1 on self-concept at t2 only for those students wihtout transition to secondary school after t1.

(3) hat there are contextual effects of class achievement at t2 on self-concept at t2 for both groups of students (those with and those without transition, in seperate models).

More sophisticated analyzes seem to exceed my competencies at the moment.

Marsh, H. W., Lüdtke, O., Robitzsch, A., Trautwein, U., Asparouhov, T., & Muthén, B. (2009). Doubly-latent models of school contextual effects: Integrating multilevel and structural equation approaches to control measurement and sampling error. Multivariate Behavioral Research, 44, 764-802. 

Good Question !

You can address partially the longitudinal aspects with  questions/issues that refer to the past and future - as well as the present.  

An example of this:  Perceived Chances for Promotion Among Women Associate Professors in Computing (below)

there is quite a lot going on here so I would advise reading an extensive account of these types of models

I have found this book useful for researchers without a lot of statistical background

Longitudinal Analysis: Modeling Within-Person Fluctuation and Change (by Lesa Hoffman)

This website gives all the code for a number of software implementations including Stata

My reading of you results suggest that that there are large unexplained differences between individuals (labeled ID)  that diminish over time (as shown by the negative covariance) and that these are large compared to the within person variation  ( that is labeled residual here); you have a very simple structure  at the lowest occasion level that is homogeneity and no autocorrelation within a person.

This gives a relatively short example that will help you understand what you have done

You have been modeling complex level 2 variation between individuals as a quadratic function of duration.

Thank you very much.

Dear Marwa, The correct approach to the analysis of your data - and any other, for that matter - is to state your Null Hypothesis first. In other words, what is your goal? Presumably, you mean to show that two events are the same or not. You must make this clear. The way you stated your question reads "Here are data, now help me find a test". Not appropriate. What you want to say is "I want to show that ... is true (e.g., equal means, equal medians, equal variances, etc) under the given experimental conditions". Furthermore, with stating your hypothesis, you make a choice regarding the significance cut-off. Usually, at 5% (or 0.05). You may not realize it, but by stating your hypothesis, you have defined "your" test statistic - in advance, mind you, and without having gathered data. Then, you go gather data, apply the test statistic and state whether under the conditions given, you accept or reject the hypothesis. In summary, your problem is that you do not state what you wanted to discover from the data. Please do so - essential for you, and essential for those who are willing to help you with your question. Best regards, Hans

for Michaela,

7 eviews program use more practical use. If the data exceeds 3 AR lag, should be re-examined using cointegration test. if cointegration test is not passed, then you test mneggunakan VECM, or VAR if it passes cointegrasi. or if you are using a mobile data ARMA

How many countries do you have ? And how many regions are there? I presume that the outcome is discrete . And I presume in this initial null model that there are no predictors included at any level.

If you have few counties ( and/or  an overly complex model) any software will struggle

Ideas behind a three level model are considered here:

and aspects of multilevel discrete model are fully considered here:

You may want to look at this paper that considers discrete outcomes with few countries and how well MCMC estimation does in comparison to likelihood approach

Political Analysis 21 (3), 2013 [PDF] [publisher link] Winner of the 2014 Jon Rasbash prize.How many countries for multilevel modeling? A Monte Carlo study comparing Bayesian and frequentist approaches. American Journal of Political Science 57 (3), 2013. [PDF] [publisher link]

MLwin has the capacity to use MCMC estimation

see

Finally and this may be your real problem - If you are  including a set of dummies at level 3 - then you are fitting a fixed effects model and then there can be no variance left unexplained at level 3 it will always be zero

See

You may want to look at the appendix  of this study to see an analysis of change due to an intervention - the outcome at 3 months later is regressed on 3 months earleir so that you are modelling change  and there is a dummy variable for those who did/did not receive the intervention.    

You do not say how intervention is determined - classically this would be randomly to the individual so that you are holding off potential confounders. But if this is not the case this book usefully discusses quasi experiments

 Dunning, Thad (2012). Natural Experiments in the Social Sciences: A Design-Based Approach. Cambridge University Press. 

So that you may want to look at a regression dis- continuity design 

Hi,

According to what you mentioned there is unexpected change in policy, this is not a justification to change your proposed study design. The longitudinal design which I assume you already started data collection accordingly will entail the presence of consistency in your measurement over time. 

However, if you still did not start data collection, you may go back to your funding agent or IRB committee to seek their permission to change your measures based on the new circumstances. 

It would be helpful if you describe the situation in some more detail.Are the 188  in a single herd or multiple herds - what is the nature of their contact?

If there is a complex structure you may need to be aware of these models as well as the two-level repeated measures of occasion (at level1) nested in cows (at level 2).

There is an example in the following  of salmonella in chickens in different flocks (nut no time)

These models are not for the beginner; consulting with a biostatistician would be helpful.