And to add a sub-question...what do you think of importance of these measurements to amount of work relation. Is the amount of work worth regarding the small possibility of wrong measurement?

Nice Question; Based on the thing basically the quality of the water which alters periodically by several factors like temp, diffusion rate, wind speed, erosion, rainfall, runoff from various aspects. in order to know the exact result of that particular location we have to measure the things in timely manner, the statistical measures with number of investigation provides you the best result ever before.

there are several automated systems avail now a days for regular measurements. it may help in this regard if you are in developed country with well established laboratory facilities. manual continuous monitoring might be increase the risk of work load.

Yea, if you've got a lot of variation in your predictor variables, sub sampling is going to be your friend. The more error is in your measurements the less likely you are to find the underlying relationship between X and y so I'd say the extra work is worth it.

As for how many sub-samples you'll need I can't say. The more variant the data the more sub-sampling is required.

Thank you for your answers. Let me clear some details.

I'm having a dilemma about taking 5 different samples (subsamples) in one time, because I know for example pH or CaCO3 isn't changing (in stable habitats like forests, meadows, peatbogs...). And ecologically looking, how important is it if the pH is 7,1 or 7,5? And if I get in one plot pH 7,1 and another, nearby, pH 7,5 does statistical testing and proving statistical difference actually give a result which ecologically isn't of much relevance. I hope I managed to put my thoughts in words clearly.

The number of replicate samples you need to adequately represent something like pH is going to depend on two factors: how variable your pH measurements are (standard error or standard deviation) and how big of a difference in pH you think is ecologically relevant. In other words, you should target the number of replicates you collect to suit the natural and meaningful variability of the system.

In my experience, peatbogs can have highly variable pH across just centimeters, depending on hummock/hollow topography. So I would not describe pH as very consistent and would recommend many measurements to adequately represent conditions at my site. The exact number would depend on the standard error of the measurements as well as the size of the plot I am trying to represent. If I'm correlating pH with 1 m by 1 m quadrats for moss, maybe 5 is adequate, but if I'm working at 10 m by 10 m plots for black spruce, I would need many more to adequately characterize the larger area.

Try a power analysis to give you some justification for the number of replicates you need. You can read more about them here: http://en.wikipedia.org/wiki/Statistical_power

adding to what others have said already, and just stressing one point a bit more: sometimes it is exactly the variability in space AND time that is characteristic for a site. You cannot know whether there is variability until you have measured. Even if you measure 5 days in a row, this sample will only be representative for the specific climatic conditions prevailing during these 5 days - so your 1 to 5 measurements always are meaningful only when indicating the sampling date as well. On the other hand, in order to appreciate seasonal effects, 5 samples will surely not be sufficient. Or do you mean 5 samples per sampling event per sampling site? This would make sense if you fear that your sampling site is variable on a small scale, as Rebecca pointed out, and that your sample will already be likely a mixture of different micro zones. In that case I would plead for stratified sampling: 5 samples each on a narrow scale within x larger plots and compare the within-block variability with the between-block variability... and check whether sampling in a smaller scale is possible, e.g. micro-electrodes, optodes etc. If you think pH differences are not biologically meaningful, why are you thinking about sampling them at all?

I see what you're saying. I agree with the statment that importance of a measurment is depending on how much it affects our organisms. Also, the measurments are made for ecologically characterizing the research site.

To get to the point, we have several plots (5x5 meters) inside which we have measured certain parameters three times, not for time variability but for statistical issues (but yes, samples vere taken in time - every time we went to take our, biological, samples, which means once a month). However, at the time the plan was smaller than it turned out to be and we lack soil for other measurements. So the thing is how important is it to make the minimum number of measurments that statistical testing (e.g. Kruskal-Wallis and similar ask). For this purpose we should take soil samples again, in one time but in several spots in one plot. In measurments that we have the oscilations were minimal (e.g. plot 1 pH values 6,92; 7,06; 7,16 - other plots have similar oscilations).

Does this make any sense? Or to leave it at three measurments?

well the minimum number of samples for statistics depends on the statistical method, but also on the background variability within the data - that's why Rebecca linked to the STatistical power. So how exactly did you come up with the number 5 for the parallels? Kruskal-Wallis might be quite robust against small samples (I would have to verify that, I am not sure), but if you have high variability, you surely don't want to stick with low numbers of samples - the background variability would be too high, as the power calculations would tell you - you would only measure inconclusive house numbers. You can calculate statistical power individually for each variable and decide to measure pH only every second time if you deem it safe, but measure nitrate e.g. as often as possible, with as many parallels as possible. It does sound like a tricky set up. If you require such high amounts of soil for your analyses then maybe the study design needs to be revised. Maybe it is better to start not with the field situation where everything varies on top of each other, but start with some standardized plot with homogenous artificial sand?