A regression model with three observations generally is overfitted and NOT reliable (i.e. as long as the effect is not orders of magnitude bigger than measurement error). OTOH with the additional time points you buy additional terms in the model in repeated measures analysis, so it won't get better. If you really only have three individuals in your sample, why are you going to do statistical testing? If your effect is really robust (similarly to a classical physics experiment) then you won't need statistical testing, otherwise it seems better to just describe your data.
can you give us a little more information to understand your question better please. Can you describe the experiment design and how you got the data? If you are trying to avoid repeated measures bias in analyses, it may be possible to use the difference (or standardised difference) to allow
You need to check what type of regression you are using. In some regression tools it is assumed that the data is equally spaced and this can lead to spurious regression calculations. Hence in ordinary linear regression it is an assumption in the development of the method - usually totally ignored.
I get round this by placing the data into equally spaced intervals on the X-axis and averaging the data in each interval.
I think there are more sophisticated regression methods that allow you to use data which is not equally spaced or which contains clusters.
Maybe talk to a Statistician regarding the functions within SPSS
A regression model with three observations generally is overfitted and NOT reliable (i.e. as long as the effect is not orders of magnitude bigger than measurement error). OTOH with the additional time points you buy additional terms in the model in repeated measures analysis, so it won't get better. If you really only have three individuals in your sample, why are you going to do statistical testing? If your effect is really robust (similarly to a classical physics experiment) then you won't need statistical testing, otherwise it seems better to just describe your data.
As already mentioned, a proper answer would require more information about your data and about the model you're fitting.
If you have equally many replicates for all experiments, you'll get (in least squares regression) exactly the same model both when using the mean values and when using the replicates as such. If the number of replicates vary, you should uses weighted regression with means. In general, it is better to use the replicates, since carrying out e.g. the important lack-of-fit test is more straightforward.
If you have three respondents that you are following over time, you could do time series forecasting, individually, or as a total, if that is of any use to you.
If you want a cross sectional model, where you have three data points, ie, (x,y) pairs of data, for a given single period, you are limited. If theoretically, from your subject matter, you expect something simple, say a linear regression through the origin, and scatterplots show that to be a reasonable assumption, you could try such a model and consider the variance of the prediction error. In SAS PROC REG, for example, the square root of the estimated variance of the prediction error is STDI for each 'prediction.' (A cross sectional 'prediction' is not the same as a time series 'forecast.')
But it really all depends on what you are trying to do. Generally, however, you cannot simply mix data from various time periods. Make sure that whatever you do is logical to your situation.
Cant use repeated measures in regression! Use paired t-test, repeated measures ANOVA, or if the assumptions are not met, use a Friedman test (no assumptions).