Descriptive Statistics show an increase but regressions shows no significance - Why? Probably a basic question for many of you:
If I look at my descriptive statistics (looking at mean values as standard deviations are too high) on the impact of different investor types on startup accounting data pre- and post investment (amount of sales, amount of total assets, employee headcount, etc.) than I see a strong increase for all factors after investor provided startups money - no surprise. However, using different multiple regressions we only find significance on the respective dependent variables for some investor types but not all.
Shouldn't I find a significant impact for all investors than?
What's the explanation and why do I even look at the descriptives if the numbers are irrelevant even though some investor types show the highest increase in e.g. sales but the regression analysis show no significant p-value?
Is a univariate regression necessary and advisable?
 A: 
(looking at mean values as standard deviations are too high)

That should be your clue. The means across types are too noisy. A regression, which more formally gives you the difference in means for a variable, echoes this finding. It may be that some investor types as defined are not meaningfully different from each other. You can bin together types to show your intended variation. Making strong claims as a result of that binning comes close to p-hacking, so I'd recommend at least sharing how/why you constructed your types in your paper/project.
A: The aims of descriptive statistics and tests are different. Descriptive statistics, as the name says, describes what can be seen in the data, e.g., that one mean is higher than another.
A test addresses the question whether what the data show could be compatible with random variation (or more generally a null hypothesis model). The idea is that what you observe is not the whole truth, but rather there is an underlying true process, and what you observe from it has some random variation. Now it is conceivable that in fact the underlying data generating processes of two means are the same, still for reasons of random variation one mean comes out larger than the other. A test addresses the question whether, assuming that there is no underlying true difference, means as different as you observe could happen. More precisely, the p-value gives you the probability that as large or an even larger difference could happen under the null hypothesis model (in univariate regression for example a model with the slope coefficient zero or no effect of the x-variable for the t-test). If the p-value is large, it means that an observed mean difference like the one observed by you can very realistically even occur if in fact there is no difference according to the underlying process. In fact, as mentioned already by @wahid, in case you have large standard deviations, this means that your observed means are very imprecise for estimating the true underlying means, so that the data can't indicate very precisely what is going on, and it may well be compatible with just random variation.
(In the question reference to means and regression is mixed so I'm mixing it up here, too.)
"why do I even look at the descriptives if the numbers are irrelevant even though some investor types show the highest increase in e.g. sales but the regression analysis show no significant p-value?" - Obviously it's your own decision what you look at, however this means that (a) the observed means are different, but (b) you cannot be sure that the difference is meaningful, it could be due to random variation. (Note that the test cannot tell you that indeed the null hypothesis is true, so you cannot know it's just random variation either.) Both of these may be of interest; (b) does not make wrong the descriptive statement that observed means are different.
"Is a univariate regression necessary and advisable?" Without knowing the nature of your data and what exactly you want to find out, this is hard to say. For sure if you have more than one explanatory variable, multivariate regression is superior to running several univariate regressions. Then, the regression model interprets your data as random sample from an underlying data generating process about which you want to find out. This is not always appropriate. Particularly there may be dependence structure among your data that invalidates the regression, and/or your data may not be drawn at random.
