If the question is model type, and are questions of distributions and variances/covariances, it sounds like you may want to consider some brand of generalized linear model (GLMM or GEE)...looking to the underlying processes of the dependent variables to help guide selection of the distribution (I frequently encounter lognormal in biochemical and hematological data) and confirm with examination of the residuals. Compare fit statistics for variance-covariance structures of varying complexity. Consider sandwich estimation to tweak the SEs of your hypothesis tests (correct for model misspecification), but this shouldn't be necesary if you construct a model that reasonably approximates the data (compare model to data with figures if feasible).

These models will be estimated with some brand of maximum likelihood (or residual ML), which will also give you the benefit of being "robust" to some brands of missing data (MAR)...that is, will estimate variance/covariances with whatever data is available to it...and implicitely assume the same would have been true if the missings had been there.

Secondly, I wasn't sure if there was some element of "what hypotheses" in the question. You note having 3 controls, an old treatment, and 2 new treatments. Were you also asking about "complex orthogonal comparisons"? Alpha maintanance (multiplicity)? Asserting the null is always tricky, but if your negative controls are reasonalby similar theoretically and empirically, then I would go that rout. ...adjusting for multiplicity with the Holm test (Holm-Tukey). Your first comparison might be a 2 df comparison of the controls...if n.s. (no alpha adjustment), then construct all subsequent comparisons to these 3 controls in aggregate (1/3 weight in orthogonal contrasts versus weight 1 in comparison group). That would limit you to 6 comparisons between the aggregate control group and the 3 treatment groups across which you would need to maintain alpha. The Holm-Tukey will further logically constrain those based on the means.

You can follow one way ANOVA as well as two way ANOVA also

Yeah, one way ANOVA is ok, but what about post test? Can I use Tukey's test?

And which values are to be included in result tables?

Ya we do use tukey's test, even u can go for Dunccan test also

TUKEY is more popular. Try using EXCEL or Graphpad Prism

it would be easy for you.

Thanks, im using graphpad prism

Thats good then then go for TUKEY only

its easy to handle

Indoxacarb Toxity Studies In Wistar Rats: Haematology

Sonphule A. M, Khan M. A, Moregaonkar S. D, Pathan M. M

if u have any conference paper like this and u want to publish

U can do it in

http://www.bma.org.in/IJFAS.aspx

If you're interested on seeing how your response variables change over time, you have to account for the non-independence of observations from time period to time period. in such a case, a mixed effects model or a repeated measures ANOVA would be more appropriate options.

Thanks Sushil, and i will chck abt that journal when i will publish articles based on my thesis work,

@ Dylan....i will try to do that but it looks complicated as the number of parameter under test are more...bdw thanks!

i suggest a two way ANOVA will be good for the data (ie. Treatment groups Vs the study interval)

here the comparison s between control, standrd and treatment groups, so only one ANOVA s applicable

I agree with Dylan--a mixed-model ANOVA is going to control for the within-subjects variation over time and the between-subjects variation in experimental condition. N-way ANOVAs alone don't allow you to test the within-subjects (time) effect.

But actually time dependance is not the major consideration in my research work, the major thirst is to be given on comparisn between normal control, toxic control, standard treatment and test treatment

Pathan, based on your explanation, I think one way ANOVA will be better for the analysis.

I think the same, bdw Thanks Mr. Machebee!

If you collected baseline data, then ANCOVA would do the job better.

thanks Edet

I supposed that for each day you have a control group so for comparison in each day a Anova or Kruskal wallis while for comparison over the time repeated measure. Don’t mix control groups. Good luck

It was once thought that classic statistical techniques perform well when the standard assumptions of normality and homoscedasticity are violated. Three major insights revealed why, under fairly general conditions,

violating these assumptions can result in poor power and a highly misleading understanding of data, even with a large sample size. The first insight has to do with the central limit theorem. Many books claim that with a sample size of forty, normality can be assumed. It is now understood why early studies missed practical

concerns when using the mean or any least squares estimator. When using the two-sample Student's t test, for example, sample sizes greater than one thousand can be inadequate. The second insight is that outliers and heavy-tailed distributions can have a major (negative) impact on the classic ANOVA and regression methods. The third insight has to do with a deeper understanding of how heteroscedasticity can invalidate any method that assumes homoscedasticity. Classic rank-based methods are now obsolete as well. Suggest looking at Wilcox (2012, Introduction to Robust Estimation and Hypothesis Testing, Academic Press) for more details. For a book that assumes little or no training, might try Wilcox (2012, Modern Statistics for the Social and Behavioral Sciences: A Practical Introduction, CRC Press).

thanks Manzymeh and wilcox for ur valuable coments

It seems you'll use a Mixed ANOVA desing.How many subjets do you have in each group?

Second, you have to check normality (K-S test) and homocedasticity (Levene test) for each group.

Third, when you use a Mixed ANOVA you test if there is differences between means. And the results only tell you if there is differences or not. Only this. But if you want to know "where" is the difference, you have to do a Post-Hoc test. And Post-Hoc test depends on Levene test and if your samples are dependent or related.

I hope this helps.

Good luck with your research!

It depends on the type of your data. For example you could be working with nuemric values BUT you could also work with categories or groups according to values-findings eg, hyperlipidemic non hyperlipidemic etc. You would need non parametric tests (MW, KS, Wxn sp etc) with interval, nominal or ordinal data in which normality check is possible but rather unrealistic. You could also see multiple non parametric regression. Otherwise all the aforementioned about ANOVA and multiple regression are acceptable. Good luck!

*numeric

The particular test to choose is entirely dependent upon the metric used. So, is this nominal data, where you have either a 1 or 0, or is it some other metric, like interval data, where you measured the concentration, etc...

If it is nominal data, then you need to use statistical test for that type of data. If it is interval data, then ANOVA and the like become possible.

thanks friends fur valuable suggestions! im going for ANOVA followed by tukey's test.

thanks all!

If the question is model type, and are questions of distributions and variances/covariances, it sounds like you may want to consider some brand of generalized linear model (GLMM or GEE)...looking to the underlying processes of the dependent variables to help guide selection of the distribution (I frequently encounter lognormal in biochemical and hematological data) and confirm with examination of the residuals. Compare fit statistics for variance-covariance structures of varying complexity. Consider sandwich estimation to tweak the SEs of your hypothesis tests (correct for model misspecification), but this shouldn't be necesary if you construct a model that reasonably approximates the data (compare model to data with figures if feasible).

These models will be estimated with some brand of maximum likelihood (or residual ML), which will also give you the benefit of being "robust" to some brands of missing data (MAR)...that is, will estimate variance/covariances with whatever data is available to it...and implicitely assume the same would have been true if the missings had been there.

Secondly, I wasn't sure if there was some element of "what hypotheses" in the question. You note having 3 controls, an old treatment, and 2 new treatments. Were you also asking about "complex orthogonal comparisons"? Alpha maintanance (multiplicity)? Asserting the null is always tricky, but if your negative controls are reasonalby similar theoretically and empirically, then I would go that rout. ...adjusting for multiplicity with the Holm test (Holm-Tukey). Your first comparison might be a 2 df comparison of the controls...if n.s. (no alpha adjustment), then construct all subsequent comparisons to these 3 controls in aggregate (1/3 weight in orthogonal contrasts versus weight 1 in comparison group). That would limit you to 6 comparisons between the aggregate control group and the 3 treatment groups across which you would need to maintain alpha. The Holm-Tukey will further logically constrain those based on the means.

Wilcox

This blog post has the R code for the Friedman Test with Post Hoc comparisons. It allows you to add a level for the block design and handles repeated measurements for the non-parametric Friedman test (which does not require all the assumptions Rand Wilcox pointed out below). http://www.r-statistics.com/2010/02/post-hoc-analysis-for-friedmans-test-r-code/

Thanks friends for the help and info!

@Budzeir Thanx 4 sharing link

Excellent discussion, Jeffrey! I'd not addressed the repeated measures aspect of the data and I quite agree...particularly regarding the most likley correlation structure. A first-order auto-regressive variance-covariance structure would be a good first pass...though I suppose it would only be 2 more degrees of freedom spent to use a completely unstructured structure. The "easy" way to make a best choice is to compare models that vary only by the structure based on thier BICs (for example). Lowest wins.

My prior comment regarding the so-called complex orthogonal comparisons can (and should) be extended to difference scores. That is, for example, change from baseline in each treatment condition versus the sum of the 1/3 weighted difference scores of the control conditions. There would be 8 coefficients in each such a comparison, all in the interation effect since the main effects would cancel (a +1, a -1, three +1/3, and three -1/3 if that helps clarify). I would also consider not including my initial comparisons between control changes under my alpha umbrella since you're carefully asserting the null and to adjust would be considered in that case. I suspect there are cases for and against adjusting the subsequent hypothesis tests for this initial comparison, but I likely would not let that alpha spent "count" since beta was really the capital of most interest. Reasonable people could disagree.

It may also be that only one of your controls is important to "beat" and would be expected to be superior to the other two. I think the message is "spend your alpha wisely!" ...and don't let convention push you around if it defies the particulars of your situation.

Depends on your data... As I read you are going to do an anoma (wich has got many restriction in use) try also a Kruskall - Walis that hasn't the biaisis of anova....

The statistical advice given is excellent, but I would approach the data a little differently. The possibility of Type II errir can be costly. I would start by using the EXPLORE function of SPSS or something similar to identify the shape of data in each group, and to identify outliers, whifh are can drive the potential significance of differences in both directions. Are the 3 control group different kinds of controls, or just 3 different samples using the same control conditions, in which it would make sense to identify differences among those first and then combine them if they seem to perform like samples from a single population.

Often we devote all our attention to Type I error, while missing differences based on things like gender, age, etc., which can mask real effects, and therefore benefits that may work with sub-populations. In medical studies, especially with relatively small samples, missing real effects for some groups of individuals is expensive. Getting 'up close and personal' with data, remembering that individuals differ in many ways besides just the specifics of how controls were selected. If a drug works for some people and not others, I want to know. If a few individuals with extreme reactions, like getting worse instead of better or no change, I want to know why, or as much about that possibility as I can.. It is also well to remember that subjects in studies like this fairly often identify what group they are in, no matter how carefully we adhere to "double bllind" procedures.