Sie erhalten aus Ihren Daten nicht die Erkenntnisse, die Sie benötigen. Was fehlt in Ihrer statistischen Analyse?

I dont see anything about data collection, which is such a key part. These steps are fine if you have a well-designed study, with perfect data collection, but most studies aren't like that. So after defining your goals, you need a well thought out study design - ideally before the data is collected. There is scant mention of data quality. Ensuring this should ideally be built into the collection and curation stage. If this hasn't been done, the data needs to be thoroughly investigated, not just by data cleansing, but by finding problems and their root cause. This is so important.

Define the problem. Before you dive into your data, you need to have a clear idea of what you want to achieve with your analysis. What are the questions you want to answer, the hypotheses you want to test, or the problems you want to solve? Having specific and measurable goals will help you narrow down your data sources, select your variables, and design your analysis plan. It will also help you avoid getting distracted by irrelevant or misleading information that may not serve your purpose. Collect data Use appropriate analysis

While defining goals, it shall also be decided whether the intended research has to be biased according to researcher's objectives or the research is based on neutrality. Having that clarity, will help to decide about selection of data, data types and techniques of data collection. The input of data and data collection techniques will eventually determine the results or outcome of the research. Hence while defining the goals, the biasness or neutrality of the researcher has to be decided and clarified.

While defining the goal it is important not to do anything which will make the data biased. Because it important to dealing with unbiased mindset which will lead us to the important insights. As mentioned in this passage it is important to get rid of the unnecessary analysis which will waste your time. Also it is important to work with causality of the features which will make you to understand the importance of the features.

While if you getting insights from your data so it is compulsory that you will choose models or tests according to structure of your data whether it is qualitative or quantitative if you select a correct model or a statistical test definitely you will insight quickly from your model results while sometimes we need more things just like data transformation scaling if data is not in appropriate position or not follow the assumptions of your model but in short all insights and meanings depends upon the model selection

First of all, do remember that most non-parametric methods will CHANGE your null hypotheses! Kruskal-Wallis, Mann-Whitney etc. test stochastic superiority and this may have NOTHING in common with comparing medians or means unless strong distributional assumptions hold. So you will obtain a technically valid answer to a question you have NEVER asked, maybe even nonsensical. So always pay attention to whether the method you want to use will change your null hypotheses. Permutation and bootstrap testing + GEE estimation are safe. But robust tests, rank-based tests, quantile-based methods (like quantile regression) will change the H0. Use model-based testing for flexible parametric models relaxing various "classic" violated assumptions.

Choosing the right type of methods to use is very important and one should be very careful before concluding the results. Proper tests should be carried out for instance 1) if you need to compare means of more than 3 groups you can use ANOVA only if normality assumption is holds otherwise you can use shapiro wilk test for normality and if normality does not hold then use Kruskal willis test for comparing means 2) when modeling data ,if your response is numeric then you should consider checking if its normal before jumping into using linear regression ,you could consider poisson regression if you assume your response follows poisson or mean and variance almost the same , otherwise choose negative binomial for cases of overdispersion .

In most cases, reporting statistical significance as P values is not sufficient to demonstrate impact. Most journals now require some measure of effect size. Proper selection of the different statistics is important to consider, depending on if you are running parametric (normal distribution) or non-parametric, (non-normal distribution) data.

I see some very valuable insights from my colleagues. Please remember data cleaning before you start... Rubbish in Rubbish out!

Help your viewer understand what points you are trying to make, without overexaggerating your findings or otherwise distorting them. Some pitfalls that I often see are distorted graph axes that make effects seem more dramatic than they are in reality, or bimodal data represented as a single bar graph showing a mean. Different fields have different conventions to be aware of when visually representing data -- in mine, it is common to pick a visualization that shows the shape of the distribution (such as a violin plot) over a bar graph. Additionally, multivariate data can be challenging to represent in 2D. Many such visualizations are color coded -- be sure that your palette still works in grayscale, because not everyone sees color well!

First of all, never push for the results by juggling statistical methods until the results YOU want are obtained. This is just wrong on many levels and you should never think about it. Do always plan your analyses BEFORE you see the data, plan the alternative scenarios if certain assumptions don't hold and stick to the plan. If the situation surprises you and makes the plan useless, justify every step you take ad hoc. But first - do always design your experiment, do a research first, address all well-known pitfalls, consult it with statisticians, control conditions and make sampling as representative as possible. And always define the practical significance, i.e. the smallest meaningful magnitude of effect (difference, ratio, %).

This question reminds me of the saying - "if you torture the data long enough, it will confess to anything". It is good to have a hypothesis. But it is not wise to have rigid assumptions on what insights you want to get out of your data. Going into an 'insights fishing' expedition leads to statistical bad practices like p-hacking (data dredging). It is always better to let the data talk rather than you forcing the data to talk.

Share data with others informally to see if they see patterns you are missing & to get reminded to analyze your data in many other ways. Plotting raw data subsets can yield fresh insights. Find the gem you were not seeking. Be your own devil’s advocate & test those ideas with your data.

The pressure to obtain positive results can often shut down our curiosity and send us into a panic when we don't get the results we had hoped for. No matter the pressure you are under, resist the impulse to subject your data to any number of tortured post-hoc analyses until you get something that just barely does the limbo and squeezes under the threshold of significance - data dredging is not honest and will benefit nobody in the long run, least of all you. On the other hand, sometimes revisiting your data from a "failed" experiment and talking to others about it can show you new ways of looking at the question that you had previously failed to consider, and this can give valuable insights to inform a future project.

I feel, when you say you are not getting insights from your data with your the existing statistical information is because you don’t have the clarity of the inferences that could be made with the information that you have. Understanding of the data has an underlying argument of understanding the statistical parameters, which needs to be mastered first. For example, a confusion matrix - it not only gives you a number for your accuracy score but generates loads of information on your Type I, Type II, proportions of predicted over actual and can be further used to plot sensitivity vs specificity and many more. At times, Less is more, we can infer a lot from simple statistical measures like mean, median and mode.