Hallo Jaydeep,

Because you have more than two concentrations Student's t test is not applicable. You have 6 concentrations (zero and 5 more). Therefore you will need to use a multisample test. Such a test is anova. But, whether you can use anova depends on what your data are - whether they are counts, percentages or some other measure. Anova is one variety of tests that belong to the generalized linear model family. One test from this family will probably be suitable for your data. Also possible would be a Kruskal Wallis rank test.

Whether you need to use post hoc tests like Bonferroni depends on your question or questions. Post hoc tests are only needed if you want to show which concentrations are different from which others. If you want only to show that the samples are not all the same, you only need the multisample test.

good luck with your work, Andrew Davis

Hallo Jaydeep,

Because you have more than two concentrations Student's t test is not applicable. You have 6 concentrations (zero and 5 more). Therefore you will need to use a multisample test. Such a test is anova. But, whether you can use anova depends on what your data are - whether they are counts, percentages or some other measure. Anova is one variety of tests that belong to the generalized linear model family. One test from this family will probably be suitable for your data. Also possible would be a Kruskal Wallis rank test.

Whether you need to use post hoc tests like Bonferroni depends on your question or questions. Post hoc tests are only needed if you want to show which concentrations are different from which others. If you want only to show that the samples are not all the same, you only need the multisample test.

good luck with your work, Andrew Davis

There is no "correct" and no "incorrect". There is only a "better" and a "worst".

You seem to look for a dose response. Have you plotted the data? How does it look like? It is probably possible to fit a dose-response model (EC50/IC50). This should be available in Prism (but I am not sure).

ANOVA and t-tests will be ok as long as the residuals are roughly homoscedastic and normal distributed. You can do the ANOVA and make a normal-quantile-quantile-plot and a residual-vs-group plot to check this. If you face problematic distributions, a transformation of your data might bring things in shape (often, a log transformation helps; note that the effects are then modeles as "fold-changes", not as "differences"). You can then simply do the "multiple-to-one"-post hoc tests (Dunnett). That's it. If Dunnett is not available, you can choose the Bonferroni correction (this should be available), better the (Bonferroni-)Holm (don't know if this is available in Prism).

However, I think the question you have is more specific and can be tested more specifically. The best suited analysis is to fit a dose-response curve, but if this is not possible or wanted, one may ask for the "smallest dose with a (significant) effect". To get the answer you have to perform ordered tests, starting from the highest concentration. If the comparison to the controls is significant, proceed to the next lower concentration. Continue until you get the first non-significant result. Then you identified the lowest dose with a significant effect. The advantage is that here an adjustment of the p-values is not required. However, I suggest to use the pooled variance for the pair-wise comparisons (you can calculate it by hand of with the help of an ANOVA).

I think that yoou will compare 5 concentration to the control one ; the test Dunnet is the appropriate test in your case;but if you will consider that you compare 6 concentrations where the zero is considered as a treatment , the anova procedure is applicable for testing differency

Dear Jaydeep,

If you want to find a relation between doses and quantitative amount of effects (dose-treatment relations) you have to use PROBIT analysis. I do not know what is the samples' Variables, the number of patients?? if do, you have to use Chi2 for instance.

But I assume you have a 5 groups with 5 different doses and you are up to compare a blood sample with a qualitative category (such as Glucose or Tropinin) in this case, the best way to analyse these information is one-way ANOVA and LSD post hoc test (which you can almost have a t-test inside your ANOVA)

however, you have to check your variable distributions' Normality with K-S test before performing these tests. if you have a non-normal dest., you should use non-parametric equivalent test "Kruskal-Wallis test" . if so, you have to set the software for finding the difference between the mean of your treatment indicator variable to analyze between each 2 group you want and use "Mann-Whitney-U test" (a non parametric equivalent for T-test).

Calculating Number-Needed-to-Treat (NNT) could be very helpful.

Good Luck

if u ask about a program you can use SPSS , use one way ANOVA test to compare

Hi Jaydeep,

Plot the data, if you see a relationship, then try and quantify it, I always check a linear relationship first (click Analyze then choose linear regression , google linear regression graphpad for instructions). If you have a relationshhip then you can quote the slope and p value for the slope. If there is no obvious increaseing/decreasing/ increasing then decreasing relationship then do an ANOVA: iif p > 0.05 then you have a nulll result. If p

Thank you all for your helpful guidance. I shall go through my data shortly and let's see with what I shall come up!

Hi,

Don´t forget to check normality in your data set (Kolmogorov-Smirnov or Shapiro Wilk Tests) before testing your groups with an ANOVA. If your distributions are not normally distributed, try with Kruskal Wallis.

...or data transformation (log or square-root most probably).

But first see, what kind of data set do you have (as mentioned above).

@Patricio and @Frantisek. Remember that ANOVA actually has no assumptions about the normality of raw data. The only thing that is important is the normality of model residuals. Many introductory statistics resources state that normality tests are required before analysis, and this is incorrect. Many people waste a lot of time dealing with data transformation unnecessarily.

Furthermore, I urge caution with data transformations. Many data transformations do not work well at the boundary conditions of the distributions. To convince yourself this, simulate a Poisson distribution with a very low mean, apply a commonly used square-root transformation to the data, and then see if normality has been improved. Whoops. By far, the worst thing you can do for analysis is to think that commonly used data transformations for your field (e.g log-transformation for hormone) will work in every case. Be careful!

That said, you are exactly right that it is important to understand your distribution beforehand. Knowing whether your data are Poisson or binomially distributed is critical for understanding if you should using a GLM-based method that will allow you to specify a function that will link your raw data to your response data (a link function).

Patrick, I could given you just a single up-vote. So here is another :)

Thanks, Jochen. Hopefully, my response saved you the trouble of a lengthy response. Ha!

You would need a pos hoc test to determine which treatment or conc. of treatment is significantly different from others as ANOVA will only give you one P value. I would suggest SPSS or STATA for analysis as these are considered superior over Graphpad. If your sample size is greater than 30 then you cant use Kruskal Wallis as the Central limit therorem states that distribution of the mean is considered normal when sample size is greater than 30

if the treatments are replicated then first do one way ANOVA and run a contrast between the control and the tretments. contrasts could be used to compare group comparison of groups of treatments as well. for all possible mean comparisons use LSD test.

Dear Jadeep

You have worked at five different concentrations was compared with control group (untreated once), This is very useful for your research, your data was allowed to post hoc multiple comparison, Its very used to how? the a single concentration was varied with significantly other concentrations (P

The best way is one way ANOVA, but if you want to analyze beyond the drug effect only, you must apply general linear model or two-way ANOVA

Hi

To compare 6 treatments levels in doubles, Anova will propose and then if the result is statistically significantdo comparison test by one of different tests such as duncan, tukey ,LSD. If you like have a comparison between control with other concentration, I propose grouping test through ortoghonal.

Hi

I agree with J. Patrick. Use 1-way ANOVA and check for normality and homogeneity of model residuals. If the assumptions are satisfied use any ad hoc comparison test. If the assumptions are satisfied you have to transform your data and check for the assumptions. If still the assumptions not satisfied use non-parametric test (Kruskal-Wallis).

I concur with, Hamid-reza Amini, that Probit analysis is one of the best shot for analyzing dose/concentration dependent responses in binomially distributed responses. Looks like you are actually using several dose-responses and perhaps using the a cohort of test animals in the experimental set up. This being the case i would go for repeated measures logistic regressions using Generalized estimating equations in GENMOD rather than the normal probit. This would explain the modeling of the residual values better and also provide the confidence internal rather than fiducial limits.

I also concur with, Andrish Da'idhe, that student T-Test is out of the least probable way of analyzing your data. However, i don't think that 'ANOVA actually has no assumptions about the normality of raw data' as, Patric Kelley, has suggested for the reverse is true, and that is one of the reasons why we look for alternative way of analyzing non-normal data viz using Non-parametric tests.

I support, Shaheena Anwar that STATA, SAS and R-statistics are considered more superior than Graphpad. Their programming algorithmic signatures are 'tighter' in terms of consistency of the outputs results???

@David Mburu. Thanks for your response. Just to clarify--and most others will agree here---when speaking of the normality assumption, I was strictly referring to raw data. (I chose my wording carefully!...or so I thought. Darn down votes!) There is no assumption for ANOVA or regression models about the normality of raw data; there is only the assumption of the model residuals. It's all about Y conditioned on X, or Y|X, and not just the fixed X (the raw data). One can see this in Generalized Linear Models, and their mixed effect extensions such as GLMMs or GAMMs, as they allow specification of a function--a transformation--that links the predictors to the distribution of Y.

For some reason, checking the normality of raw data has been injected into introductory statistics texts as being the first step to any such analysis. Others may have different perspectives, but as I've mentioned, this has wasted many practitioners' time attempting to normalize data, which is difficult for small datasets anyway or unnecessary all together.

That said, the normality of residuals is linked to the distribution of raw data. So, it may well be that a data transformation is necessary to improve normality of residuals. The advent of nonparametric tests simply arose to solve those cases where assumptions of these models (i.e. normality of residuals) were left unmet.

I'm loathe to cite Wikipedia as the source of all knowledge, but here goes: "The normal-model based ANOVA analysis assumes the independence, normality and homogeneity of the variances of the residuals." Wow, Wikipedia, you helped us again.

I would maintain that for Bhat's scenario, i would think that repeated measures logistic regressions using Generalized estimating equations in GENMOD rather than the normal probit is the best option.

If you're using a combination of drugs then the Isobologram or any of the other synergistic models can help you assess the drug activity and interactions in relation to another substance.

But in your study i suggest plotting the mean responses of the groups in a bar chart format with error bars to give you a visual on the trends of responses across groups.

...you may proceed to do a one-way anova with a post-hoc test e.g. Duncans or Turkey's test to give you the exact groups that significantly (p

Azubuike, Ducan and Tukey do not at all require an Anova.

Repeated measure logistic regression would work for me for it can give use the confidence limits and the lethal doses at which the drug can give a proportional clearance of the target microbes

Probit analysis using repeated measure logistic regression would work for me for it can give use the confidence limits and the lethal doses at which the drug can give a proportional clearance of the target microbes

Jochen Wilhem, in SPSS post hoc tests are incorporated in the one-way ANOVA option, and that was the perspective i was speaking from.

Normally when you run a one-way ANOVA it only tells you from the p-value that some differences exist among the groups but it does not tell you between which groups. The post hoc tests when included highlight which groups actually differ.

Azubuike, thanks you for this explanation. I know well. It may lead to the strange result that the F test is "significant" but neither of the pair-wise comparisons. Also it may happen that pair-wise comparisons are "significant" but not the F test. But the worst thing is, that people start to be convinced that the significant F test is required in order to identify the "significant" pair-wise differences, and in general, that "post-hoc testing" is intimately related to Anova (it is not! these are very different things, answering very different questions). This is a common misconception. The F test does not "protect" the FWER (two exceptions: Fischer's LSD for 3 groups and Scheffe's test). In fact, disregarding the pair-wise comparisons when the F test is not significant is actually a loss of power.

See, for instance, Review of Educational Research 66(3), 269-306.

Thanks for the clarification Jochen, i'll read up some texts to get some more clarifications. Actually a number of popular texts do not portray the relationship between ANOVA and post-hoc tests in this light!

http://www.biochemia-medica.com/content/comparing-groups-statistical-differences-how-choose-right-statistical-test

Article Concentration-effect and dose-response relations in clinical...

ANOVA and Logistic Regression

Depending on what your outcome measurement is, the analysis can be different.