Research Statistics - Science topic

Cohen, Cohen, West, & Aiken, Applied Multiple Regression/Correlation Analysis for the Behavioral Sciences.

You should take the time to invest in yourself. We all had to learn coding, if it was something like C or C++ I would understand your reservations but Rstudio and Python are not too abstract for someone to learn. I find that doing histograms with Python/Rstudio much easier than Excel. Learning these GUI stats packages to me are a game of dimishing return, whereas if you put in the time to learn Python/Rstudio you will be rewarded much more for your troubles especially in the long term!

Neil Patrick raguine Razon I guess Factorial ANOVA is the appropriate one here. ANCOVA is relevant only when one of your independent variables is non-categorical, but in this case, all your independent variables are categorical.

Best !!

AN

Dear friend.

This article can helpful for you:

(PDF) Perceived Stress and Coping Strategies (researchgate.net).

Very useful discussion. I remember, the saying "it is better to study one rat for 1000 hrs rather than studying 1000 rats for one hr if you need to study behaviors of rats". But that particular rat should be selected carefully; not to choose a 'mad' or 'disabled' rat. Now by studying this carefully selected rat (theoretical sampling) we can generalize certain behavioral patterns of rats, as well as can gain a deep understanding of the behavior of rats, in that time and space. Observing 1000 rats, randomly or even purposively (statistical sampling), for one hour would help to identify a single or couple of features of rats, in that particular context. So, the sampling technique depends on the nature (not the purpose) of the research.

Hallo!

Unfortunately, I faced another example of the use of quantitative study evaluation criteria to evaluate my qualitative study by an editor: "There is a concern that the sample size in your study is rather small – 33 individual interviews. It may not produce robust results ". By the way one of classic sigle case study research (Echeverri and Skålén 2011) did 33 interviews: 'The present article draws mostly on the 34 interviews conducted during the second and third rounds' (Echeverri and Skålén 2011, p. 13)

Echeverri P and Skålén P (2011) Co-creation and co-destruction: A practice-theory based study of interactive value formation. Marketing Theory, 11(3): 351-373.

What do you think of it?

I recommend WinADV, you can easily sort and analyse the ADV data and it's really useful for people working with ADV.

You may want to consider this free online course which starts with the basics and becomes quite advanced - it was vetted and approved externally from our Centre

It can be followed in MLwiN, Stata and R; each module is divided into two parts - conceptual and practical; we provide the data for the practical. There are quizzes that help you evaluate you understanding.

Hi Dear Jaime Eduardo Gutiérrez Ascón

Thank you for your answer. I think it is very useful for me.

Best Regards...

Tq Sir @Ariful Islam

Thanks for the response....sure.

That answer makes a huge difference.

In a thesis it is allowed to discuss one or more projects that are failures (in my opinion). The important part is what you learned from the failure. With 4 questions and 23 completed surveys you can analyze the data. The next part is to interpret the analysis. The discussion is then about what failed, why, and how to change things to increase the chances of success, and within this to also talk about what the data might say. Even if it turns out to be unpublishable, this project could still be an outstanding success.

My first problem is that I don't understand what you need to test normality and homoscedasticity if you know that the distribution is skewed.

If you transform a variable, the tests (Shapiro-Wilk or Levene) are about the distribution of the transformed variable. That does not make them more or less accurate.

Removing outliers is only sensible if these values are "bad values", that is, when they are extremely implausible (or even actually impossible, e.g. as result of a typo or some other mistake) o if there are extrenal reasons that invalidates them (e.g. it is known that the animal from which that outlying value was taken was unintendedly sick, or that a person being interviewed turned out to have misunderstood the question, was drunk or sponsored to give a manipulated answer etc.).

You current studies don't seem to involve experiments (it seems you take observational data / questionaires). I personally don't see much sense in significane testing in these cases. Particularily not, if the task is to explore different factors (their correlation with the outcome measure), and to identify or find those with the "biggest effects" (these could be the statistically non-significant ones!).

When using rank-based methods (supposedly what you call "non-parametric") you loose the quantitative relationship between the variables, something I consider not really helpful (at least in 99% of the cases, to make up a statistic, too :)). Be careful when talking about "non-parametric alternatives (to parametric tests)". These "alternatives" are about different hypotheses. For instance, MWU and t-test do not test the same hypothesis (except in those cases where the distribution is in fact normal - where you would not think about any "alternative" at all). The sad truth is that most people re-iterating these textbook phrases just don't know what hypothesis the MWU-test (or similar non-parametric tests) is (are) actually testing. As an analogy of this absurdity I could ask you if the apple or the pear I have is heavier (testing the null then the weight of both fruits is equal). Now you say: "ah, sorry, man, I don't have a good weighing maschine for these things son I wouldn't trust it. But, hey! I can easily spot the color, and let me tell you: the apple is significantly more reddish than the pear (p < 0.01). You have that nice p-value. That should answer your question!"

Ok, it' not that stupid, but the direction is right, I think.

Take-home message: Don't let tests dictate what you should do. Think thoroughly what you want, what data you need to provide the relevant information, and how you can extract, visualize and communicate that information. Tests may not at all be helpful. Look for plausibility rather than for (statistical) significance.

Question was asked 2 months before, amd i saw it now :(

Hello Terun. You listed SPSS as one of the topics for your question. I suspect that either file 6 or file 7 from the page linked below will give you what you want.

HTH.

Hi Heidi,

Given your paramaters of within-US, I believe with a bit of digging you can come up with the descriptive data you're seeking by searching the US Bureau of Labor Stats (BLS), URL --> https://www.bls.gov/opub/ted/employment.htm?view_full

I can suggest to keyword search (variable name scope) 'work experience' OR 'TENURE'.

Hope this may of some useful help to you, and best of luck on your study!

With kind regards and cordially yours,

Matthew J. Kerry

Dear Abdullah

The attached documents are very helpful in calculating and interpreting the measures.

Regards,

Zuhair

Hi Yi,

Use some type of ANOVA if you're looking at differences. Use some type of regression if you're looking at relationships between independent and dependent variables.

Here's are book links that may be helpful:

Have a great day!

--Adrian

/*(For NFHS-3)

For Underweight (-SD), Stunting (-SD), Wasting (-SD)*/

use IAKR52FL

gen haz06=hw70

replace haz06=. if hw70>=9996

gen waz06=hw71

replace waz06=. if hw71>=9996

gen whz06=hw72

replace whz06=. if hw72>=9996

(For -2SD)

gen below2_haz = ( haz06 < -200)

replace below2_haz=. if haz06==.

gen below2_waz = ( waz06 < -200)

replace below2_waz=. if waz06==.

gen below2_whz = ( whz06 < -200)

replace below2_whz=. if whz06==.

(For -3SD)

gen below3_haz = ( haz06 < -300)

replace below3_haz=. if haz06==.

gen below3_waz = ( waz06 < -300)

replace below3_waz=. if waz06==.

gen below3_whz = ( whz06 < -300)

replace below3_whz=. if whz06==.

label variable haz06 "Length/height-for-age Z-score (stunting)"

label variable waz06 "Weight-for-age Z-score (Underweight)"

label variable whz06 "Weight-for-length/height Z-score (Wasting)"

label variable below2_haz "Stunting (-2SD)"

label variable below3_haz "Stunting (-3SD)"

label variable below2_waz "Underweight (-2SD)"

label variable below3_waz "Underweight (-3SD)"

label variable below2_whz "Wasting (-2SD)"

label variable below3_whz "Wasting (-3SD)"

save IAKR52FL, replace

Hello Khalil Khan How did you solve the problem. Please share the solution.

Assignment of Reliability for an industrial product that has several parts. These parts act in series.

Hello Naike,

When the goal is estimation of a parameter (like a population proportion) rather than a comparison of conditions/treatments/methods, then publication bias (due to nonsignificant results) really isn't pertinent, as you correctly suggest. You can, of course, generate funnel plots and use checks such as Egger's to see whether the distribution of sample estimates follows what one would ordinarily anticipate (less variation with higher N, roughly symmetric dispersion about the middle values). Past that, though, the exercise is not very informative.

Proportions, though, are often transformed to behave more like continuous (e.g., unconstrained) variates; the arcsine transform (2arcsin(sqrt(p)) and logit transform [ln(p/(1-p))] are common methods (see http://strata.uga.edu/8370/rtips/proportions.html).

Good luck with your work!

Very good point, Adrian and Peter, regarding one statistical program checking another.  (I was not aware of the details on Canada which you provided, Adrian. Thank you.)  But in the US, agencies generally are not allowed to share data, so if data from one agency might inform another regarding issues such as regressor/auxiliary data, or even enough information to be better able to stratify when sampling, it may be difficult to impossible to allow  cooperation.  If say a manufacturers' energy  consumption survey would benefit/be more efficient if the survey methodology were planned and estimators based on supplementary information, such as number of employees or expenditures, which may establish a size measure for each member of the population, that kind of thing may not be possible.  US agencies use memoranda of understandings, MOUs, to try to work together when allowed, but even then it can be tedious/inefficient.  

There are many textbooks that could help such as

Kellar, Sacey & Kelvin, Elizabeth. (2013). Munro’s Statistical methods for health care research. 6th. ed. Philadelphia: Lippincott.

I use this book to teach applied statistics for master students. 

If you specify your research quesitions, then it will be come easier to guide you through the best statistical analysis.

Hi Gerardo,

A Handbook of Statistical Analyses

Using R is my favorite..

If you have the R software (which is open source) installed, the following steps can help perform seqMK test:

1. Write the command: install.packages('pheno')

2. Write the command: library(pheno)

3. Write the command: seqMK(c(data values separated by comma))

Hope that helps!!

Without knowing more about your data, it is possible to have a nonsignificant overall test but show significance in main effects or simple main effects.

Check power and effect sizes.

Marcie Zinn

Charbel: Didn't the Holy Spirit mention to you that software piracy was illegal, and that it deprived people of their livelihoods? Please don't post links to pirate software, and please pay for what you use.

I notice that you have a chapter on ethics and marketing. This is kinda ironic.

There is a growing body of research linking customer satisfaction to outcomes like customer retention, word-of-mouth, sales, stock price and so forth. We provide an overview in our MSI paper, which may be downloaded at the link provided. Hope it helps.

Dear Sougui,

To be clear: there is no problem in using both techniques in the same research. However, there is an adequate context suggested by literature for each one.

Now, using two different data processing techniques in the same research, it is necessary to bear in mind that in addition to the double effort of processing, analysis and interpretation of data; all theoretical support, should also be built to support both interpretations and conclusions made about the data analysis with CB or PLS.

In other words, you need: 1) a theory to support interpretations with CB-SEM (The goal is theory testing, theory confirmation, or the comparison of alternative theories), and 2) a theory to support interpretations with PLS-SEM (The goal is predicting key target constructs or identifying key "driver" constructs).

Kind Regards

It all depends how much information you have - how variable are the X's and how many level 1 and level 2 units you have. That is the issue is statistical power to estimate such a sizeable model. To give you some idea Raudenbush and Bryk (1992) indicate  with 60 students in each of 160 schools it is possible to estimate 4 random variables at school level (so random intercepts and 3 lots of random slopes).

I would proceed pragmatically

1 include all the level 1 variables in a random intercepts model, then add

2 each level 2 variable (one at a time) then sequentially in terms of the biggest reduction in the deviance

3 then allow random slopes at level 2 (one at a time) then sequentially in terms of the biggest reduction in the deviance ; to see what can be fitted and to give you an idea of what needs to be accounted for at level 2

4 Finally include each cross-level interaction  (one at a time) then sequentially in terms of the biggest reduction in the deviance 

This is deliberately exploratory and you have to be open about it so that people will understand what you have done. If you have lots of data you can have a calibration dataset and a hold -out sample to ';test' the final model.

You can see a published example of this type of approach here

Dear Nicki,

First if the same underlying structure is measured with different tools in different studies for example quality of life, it is possible to combine across studies using the SMD (see for example http://handbook.cochrane.org/  section 9.2.3.2 'The standardized mean difference')

Second if this is not advisable because of different populations etc. I would suggest a simple descriptive analyses. To give a structure you could use the PICO format (population, intervention, comparator and outcome) when formulating your research question. Then for the reporting make a table describing the population, the intervention and the comparator in each study. Group the reporting of the outcomes so that for example if your question is about an intervention on how to help people quit drinking and one of your outcomes is alcohol consumption, then for each study you report the outcomes you think covers the overall concept of alcohol consumption for example no. of heavy drinking days, drinks per week etc. Your next outcome might be abstinence here you could report no. of persons abstinent or time to first drink or first binge.

I hope this help.

Cheers,

Britta

To my mind you can never observed the 'population' even if you have a census as this is a theoretical concept that is not achievable in practice.

I have set out the argument and consequences of this in much more detail in a previous posting

as I say there " the observed count should be considered to be the outcome of a stochastic process which could produce different results under the same circumstances. It is this underlying process that is of interest and the actual observed values give only an imprecise estimate of this"

In brief I think you are always dealing with a sample.

I know that a lot of people will disagree with me!

Hi, You may try a heatmap grid.

Thank you for the responses.  It is greatly appreciated.

Perhaps you would like to take a look at the book of Zuur et al. on Mixed Models... This is an excellent intro with lots of examples on "zero-inflated" data.

Thank you for help! I think it is good resolve, decision.

Best regards, Magda

Personally, I particularly enjoy the work of Podsakoff on this matter, who presents specific ways to try to cope with CMB. Also, using a single survey-based data source might be the best method when 'both the predictor and criterion variables are capturing an individual’s perceptions, beliefs, judgments, or feelings’ (Podsakoff, MacKenzie, and Podsakoff 2012, 549).

Nevertheless, in one of our recent studies that was accepted for publication in Public Money & Management we coped with CMB as follows: The independent variables (formality of and participation during strategic planning) were asked to single informants (one per municipality) whereas the dependent variable (strategic-decision quality of the strategic plan) was asked to multiple informants (at least 2 per municipality). This allowed us to, to some extent, cope with CMB while still using data from a single survey-based source.

Food for thought!

Sincerely,

Bert

Hello,

For me, possibly the most challenging is/was/will be to identify a niche within a vast amount of knowledge to be able to introduce meaningful research questions that would be novel and could enhance understanding of mechanisms underlying psychological disturbances (Yes, I am a scientist and a psychologist).

Sounds like a philosophical statement. But it matches your question - very broad and philosophical too.

Regards, Witold

There is possible to get a very accurate approach trasforming to taylor series ;).

Dear Leigh, 

that is not the same as repeating or confirming the other study. This a highly acceptable and good way to assure that the results are valid. 

As explained by you  it seems as though you are doing a mixed method study .You either do a qualitative analysis ,look at emerging themes and then use these themes to do a quantitative analysis which would involve the process of sampling . Alternatively you can collect data quantitatively and do and analysis to see the correlations and then by interviews  validate the results of the analysis. It is advisable to determine methodology before sampling

In multiple regression analysis, suppose there are two independent variables, will get two partial coefficients and in interpretation of these coefficients, one will give the change in dependent variable for unit change in that independent variable and other variable is kept constant.  

Reviewers are concerned about fishing for results and then presenting as if you have not.

I think that complex interactions are entirely plausible if you explain what is going on - voting behavior is now affected by gender by age by class and not to model this in full is missing out on  important relationships.

This paper

has some arguments at the beginning for the importance of interactions - it also presents a model that automatically guards against over-interpretation as differential effects are shrunk back to no effect if they are based on unreliable estimates.

But we will probably have difficulties getting accepted too!

Please check your mail

You first calculate the partial derivatives of the two fromulas for each measured variable. Then you calculate the sum of the products of the errors (Dr, DR, and dh) with the squared corresponding partial derivative.

Example for the length of the mantle:

dm/dR = (R-r)/root(w)

dm/dr = -(R-r)/root(w)

dm/dh = h/root(w)

where w = (R-r)²+h². The squared derivatives are

(dm/dR)² = (R-r)²/w

(dm/dr)² = (R-r)²/w

(dm/dh)² = h²/w

The propagated error is

Dm = DR*(dm/dR)² + Dr*(dm/dr)² + Dh*(dm/dh)²

Dm = DR*((R-r)²/w) + Dr*((R-r)²/w) + Dh*(h²/w)

Dm = (DR+Dr)*((R-r)²/w) + Dh*(h²/w)

I hope I did not make some error somewhere.

You will surely be able to do the same for the simpler formula of the surface.

Dear Leonard,

Thanks for respond. Actually, I have the observed data from 1961 until 2010 and the projection should be until 2099.

may be Cordex is another option?

Dear Jeff and Iryna, thank you very much for your suggestions!

Best

JP

It sounds like you asked different questions in each of your samples (so the difference is more than just 4-point versus 5-point scales). If so, there is no possibility of doing correlations between the two sets of questions because no one answered both sets.

If you are thinking in terms of writing an article, I would organize into what is sometimes known as a Study One and Study Two format, where you would present each set of results separately, and then compare them in the Discussion section.

RMSEA is does not work with ULS estimator but you still can evaluate your model with GFI (Goodness of fit). This is because RMSEA and GFI is considered under absolute fit. Actually, there are three categories namely absolute fit (RMSEA, GFI), incremental fit (IFI, CFI, TLI, AGFI, NFI) and parsiminous fit (chisq/df). Thus, GFI is adequate for evaluate your model.

You can calculate the significance of the difference between two independent proportions.  If you need to do many tests look into a correction of the criterion p value to protect against Type I errors (e.g., Bonferroni).

Here is a worked example from the internet:

Hi,

If your data is stationary at level, you can not use JJ Cointegration. You should use simple regression analysis. or You may use stock prices data for JJ test.

regards

If the bimodal distribution is a mix of two normal distributions, then the squared bimodal is a mix of two gamma distributions.

X_bimodal = i * X_a + (1-i)*X_b

with i ~ Bernoulli(p), X_a ~ N(µ_a,s²_a), X_b ~ N(µ_b,s²_b)

(X_bimodal)² = i * (X_a)² + (1-i)*(X_b)²

=>

(X)² ~ Gamma( shape=µ²*rate + 1, rate=exp(-s)/2 )

Hi Bonnie,

the two best books on longitudinal data in a mixed model (using Stata) context are:

Singer, J. D., & Willett, J. B. (2003). Applied longitudinal data analysis: Modeling change and event occurrence. Oxford university press.

Rabe-Hesketh, Sophia, and Anders Skrondal. Multilevel and longitudinal modeling using Stata. Stata press, 2012.

You will also find that the help file in Stata is extremely helpful.

I hope this helps

Ariel

Maria,

first, do not dichotomize any of your variables. There are so many problems resulting from dichtomization, that you could not trust your results - either a linear relationship or a nonlinear relationship.

This article is state-of-the-art and will explain the problems meticulously: MacCallum, R. C., Zhang, S., Preacher, K. J., & Rucker, D. D. (2002). On the practice of dichotomization of quantitative variables. Psychological Methods, 7, 19 40.

Second, 13 independent variables are too many for multiple regression. If these variables are correlated, you would have a multicollinearity problem. This means, that parameter estimates may be quite small and standard errors may be quite large. Therefore it would be better to reduce the number of independent variables. You could either perform an exploratory factor analysis and use the factor scores as independent variables (or a PCA, if you do not expect latent constructs to be the origin of the correlations), or you could reduce the number of variables based on theory.

Third, check your independent variables whether they are censored above or below  or terribly skewed - this would introduce artificial nonlinearity. However, your dependent variable should be nonnormally distributed if a nonlinear effect exists in the population. This article (which also explains problems of moderated and quadratic regression) may be helpful: Dimitruk, P., Schermelleh-Engel, K., Kelava, A. & Moosbrugger, H. (2007). Challenges in nonlinear structural equation modeling. Methodology, 3, 100-114.

Then, include quadratic terms (products of single variables) into the regression equation as suggested by David and - in the next step - moderator terms (products of two independent variables).

HTH

Karin

Muhamad,

There is a nice Wikipedia article about stochastic drift: Go to http://en.wikipedia.org/wiki/Stochastic_drift and scroll down to the section titled "Stochastic drift in economics and finance".

The first model, yt = f(t) + et, (notation: the t in yt and et should be a subscript) is the deterministic trend model. Your time series (in your case, climate data and productivity data) is then a function of some deterministic function of time, plus a stationary error. "Deterministic" means that it has no random component. One example of such a function is a linear time trend, another example is a quadratic time trend. If you plot your time series data, in the case of a linear deterministic trend, you would see that the Y values keep increasing (or decreasing, for a negative trend) at a constant rate (in other words, if you draw a straingt line that trends up, the yt observations would randomly jump up and down from that line).  For the linear trend, you need to create a column of data with the values t = 1, 2, 3, 4... representing the time index, and fit a regression model where your target variable (climate or productivity variable) is the Y variable, and the time index t is the X variable. Save the residuals. These will be the "detrended" version of your Y variable. For the quadratic trend, you would have to also create a variable-squared, with values t2 = 1, 4, 9, 16, 25, ... Again fit a regression model with t and t2 as the two predictors and save the residuals to get the de-trended version of your Y.

The second model, yt = yt-1 + c + et, is a random walk with drift (notation: t in yt and et and t-1 in yt-1 are subscripts). This model is first-order autoregressive. Each observation at time t is a function of the observation at time t-1. If you bring these two components to the left-hand side of the equation, you get yt - yt-1 = c + et. The left-hand side is the first difference. The right-hand side is a constant and a stationary error et. If you plot your time series data, in the case of the random walk with drift, You would observe a long-term increase (or decrease for negative drift), but, in the short run, you would see the data to go up, up, up, also randomly fluctuate up and down a little, but keep going up, up, up, for a few periods, maybe 6 or 10 periods, then start going down, down, down, again randomly fluctuate up and down a little, then down, down, for a few more periods, then again up, up, up, etc. So, within 1 or 2 observations you would see random up and down fluctuation, within 6-10 observations you would see a consistent upward or downward local trend, and within 50-100 observations you would see a systematic drift, either upward or downward. Based on this model, you can de-ternd your data by simply computing the first differences! Create a new column of data where zt = yt - yt-1. Then zt (day-to-day differences, or month-to-month differences, depending on your periodicity) would be a random error with mean equal to c (i.e., random fluctuation from a horizontal line placed at level c).

So, which one of these models should you use? The bst way is to (1) observe the time series plot for the patterns I explained above, and (2) produce  autocorrelation function and partial autocorrelation function plots. In Minitab, select Stat > Time Series > Autocorrelation, and Stat > Time Series > Partial autocorrelation. In SPSS, select Analyze > Forecasting > Autocorrelations, which will produce both ACF and PACF plots. If all bars in BOTH plots are insignificant (i.e., within the threshold lines), then your data should be modeled using the deterministic trend model. If the PACF plot has one significant bar at lag 1, with all the other bars insignificant, and the ACF plot starts significant at lag 1 and decreases very slowly, staying significant for a few lags, then the stochastic  trend model (random walk with drift, or AR(1) ) is appropriate. If you see a significant bar at lag 1 on the ACF plot, with all other bars insignificant, and significant bars at lags 1, 2, 3, etc. that go down slowly on the PACF plot, then you have a different model, called MA(1), or moving average.

 In SAS, use proc ARIMA with the statement identify var = y(1), which produces the first differences, and see if the first differences pass the autocorrelation test (producing high p-values on the chi-square test for zero autocorrelation) and also produce ACF and PACF plots that only have non-significant bars at lags greater than 0. 

Good luck!

I agree with Helena that you should start with using Exploratory Factor Analysis on your multi-item scales to determine if they are Uni-dimensionsal (i.e., they each measure one and only one thing). If that works, then use Cronbach's alpha to assess the reliability of the scales.

There have been so many questions on RG about using these two methods that I have compiled a set of resources on this topic:

Hi Rajesh,

I am not sure that histograms can provide you with anything other than visual displays of your data. More importantly, the binning process based on kernel densities are usually automated, so that you cannot be certain that the visual display will always look the same, even after altering the binning procedure. Again, histograms are good visual depictions of the data, but they do not provide you with the sort of details you are looking for. You may need to examine the distribution of the data using standard statistical measures such as percentiles, etc.

Ariel

In the following article, there are multi methods used to test mediation. hope it help

Dardas, L., Ahmad, M. (2015). For Fathers Raising Children with Autism, Do Coping Strategies Mediate or Moderate the Relationship between Parenting Stress and Quality of Life? Research in Developmental Disabilities. 36, 620-629. http://dx.doi.org/10.1016/j.ridd.2014.10.047

I eventually answered this question myself: 

Smithson, M. & Shou, Y. (2014). Randomly stopped sums: Models and psychological applications. Frontiers in Psychology: Quantitative Psychology and Measurement, 1-11, doi: 10.3389/fpsyg.2014.01279.

This paper is an example in serology, but I think that the principle is just the same in your case (observed and expected values):

Hope that will be useful

Best regards

Have I understood correctly? You are being asked to present the "pooled mean" of each of the two groups, and the "pooled SD" of each of the two groups?

Why do you need to report that? I guess you can do it in the same way that you can get pooled estimates of anything, take the mean from each study, and the standard error (either from the SD and n, or from the confidence interval) and conduct meta-analysis on that. The "pooled SD" could possibly be derived from the within-study variance.

So I think it's easy enough to do, but I'm just not convinced you really need to do it. The mean and SD could be more prone to differences between the study populations chosen, and it's the difference between the means for each study that you need to look at and pool.

I'd be interested in other people's views. Maybe I'm missing something.

Thank you. I am actually interested in differences among seasons but the data are highly correlated across seasons and I am not sure about using season as a fixed effect. 

Ah, in that case, the integrated values are probably all you need.  Fitted Gaussians will not really have individual significance.  You can also get the average bond length from the integrations - it is the first moment of the distribution.  Simply divide integral of rT(r) by integral of T(r).

Propensity score (PS) matching (or other PS techniques such weighting, or subclassification) are part of a greater family of causal inferential techniques that try to emulate the randomized controlled trial (RCT) by ensuring that participants and controls are comparable on baseline characteristics. The difference is that RCTs randomize participants and thus groups are assumed to be comparable on both observed and unobserved characteristics. When using PS techniques, participants can only be matched to controls on observed characteristics and we assume that there is no residual confounding.

I am attaching links to several of my papers on the topic. They provide very specific examples of how these approaches are applied in different contexts.

I hope this helps

Ariel

Kumar

It is perfectly possible to have correlated random effects. The term Random Variable simply means 'allowed to vary' - that is variables can take on different values (according to a distribution) and does not necessarily mean 'independent'.

While you cannot get THE mean and SD, you can convert to a z-score (mean = 0 and SD = 1).  From there, you can convert it to any scale you wish.

I am not quite sure if you want to test the effect of weeks on animal weight. If you only want to test the difference between treatment a repeated measures ANOVA would be more appropriate, including time as the error term. And make sure to do pretests for normality of data and especially homoscedastity!

For further information this site could help you a lot i guess:

With the results you can argue about the treatment effect on animal weight, if that is your research interest.

Hope that helped a bit? Feel free to ask further questions!

Probably an example helps:

The fact that one celebrates his 70s birthday is perfectly correlated with not dying in a car accident at the age of 20.

Obviousely, neither the 70s birthday-party causes to have no accident, nor does the absence of a deadly accident at an age of 20 cause the birthday party.

This example shows that

a) one needs a reasonable causative MODEL to think about a possible cuasation at all,

b) the time of events and the direction of the possible causation must be considered,

c) other possible confounders must be (logically or -better- experimentally) excluded.

Simple correlational studies often lack all three of these requirements. Often there is no concrete causative model proposed, further is is impossible to see in what time-order the events happen, and there are usually more than a million possible confounders.

A clear conclusion about causation can only be derived from a planned experimental study where the suspected cause is controlled by the experimenter, everything else is kept constant, and the experiment is designed to read the response in time after the suspected cause could operate (according to a well described model of how this may happen - the causative model).

Giulia, 

As you're probably gathering, there are at least three ways to think about this question.

1) What's acceptable/published in your field. In the large-scale survey world, some people think of sample sizes of less than 1000 as "small," but much (maybe most) of the academic research out there is done with smaller samples. 

2) Sampling method aside (and whether you have good population coverage, response rate, etc), sample size specifically affects variance, standard errors, p-values, and confidence intervals (as all those things depend on n). The only way to get a concrete answer to your question is to plan your analysis and decide on what effect size you're trying to obtain, and under what  criteria (e.g., alpha level). Once you do that, you can do a power analysis to see what sample size you should have to get that effect size under those assumptions.

3) Another similar approach that will give you a specific number is to plan the sample size around the confidence intervals that you want (i.e., how wide do you want them to be). Like power analysis, you put in your assumed point estimate and standard error of the estimate, and you can see how different n will affect your CI width (or, logically, you could define a CI width and get a different n). This is like power analysis but doesn't actually choose the sample size based on power of a statistical test, just confidence intervals. 

There are a lot of other things to consider, like others have mentioned, such as whether your sample is truly random, the target population you are measuring, the sampling frame you are using and its properties, but these really apply to population surveys only. However, on a practical note, if you're doing any kind of design (even a convenience sample) where you contact people or send them something to respond to, make sure you take into account your response rate (i.e,. you won't get 100% of the people you contact or request to participate), so your calculated n based on a power analysis needs to be inflated by that factor. 

Here's a good set of references from AAPOR on sampling, response rates, and survey best practices http://www.aapor.org/AAPORKentico/Education-Resources/For-Researchers.aspx 

     

Hello Michael

The topic sounds interesting. I am an Assistant Professor at Statistics Department, University of Benghazi, Libya and am interested in helping you if I could.

Best Regards,

Dr. Adel Al Sharkasi

Tel: 00218924392190

If you use SPSS, you need to do only KMO (test for sampling adequacy) & Bartlett Test for sphericity (to ensure moderate inter-correlations). Normality test can be over looked