Econometric Model and deciding the frequency of data collection I am looking to build an econometric model and I am wondering if using annual data vs monthly or quarterly data is going to produce a less accurate model. If the dependent variable is affected by seasonality, would it be innappropriate to use monthly or quarterly input data? Or is it still generally better to use a higher frequency of input data regardless? I would prefer to use monthly or quarterly inputs but I am worried this would skew my model. Any suggestions or insight would be greatly appreciated.
 A: More data (collected at higher frequencies) is not really more data as higher levels of auto-correlation exist. What I have recommended is a two-fold approach 1) At what frequency would you like to detect recent unusual activity or in other words what frequency do you wish to make forecasts i.e. end of day , end of week , end of month .,etc. Once you have determined that then consider higher frequencies from that point. So if you had said quarterly as the answer to the first question , you now have to assess/specify what forecast length you are concerned about. For example you might select/specify 1 quarter. Now you can use monthly data to predict the next quarter OR daily data to predict the next quarter or hourly data to predict the next quarter. These three different approaches can all be used and you can compute the accuracy of each and then select the winner from this 1 origin. Make sure that you repeat this experiment for a few origins not just 1.
A: It think that it will be worthwhile to investigate data availability before deciding upon a frequency. For example, you'll probably encounter a trade-off between having annual data dating back further in time than the available quarterly data. This would be the case for the Annual and Quarterly National Accounts in many countries. For example, in Ireland, you'll find Annual National Accounts data as far back as 1970, but the Quarterly National Accounts only begin in 1997Q1. 
This point crosses over with the purpose of the model. In the situation just outlined, if you were going to be considering episodes in, say, the 1980s then the quarterly data would pose obvious limitations. You could, of course, employ interpolation methods to derive lengthy quarterly series, but this can be no mean feat if you have a substantial amount of data to interpolate. So, the size of the model should be borne in mind, too.
Another option would be to build an econometric model that uses annual, quarterly, and monthly data. There a number of methods that can be used to do this kind of mixed-frequency modelling; some more sophisticated than others. For example, there are MIDAS regressions, Kalman filter models, and simple time aggregation methods (closely related to bridge equations).
In the broad scheme of things, if you decide to use higher frequency data (quarterly) which is not available as far back as low frequency data (annual) then you may be omitting important long term trends. This could be important in the context of, say, modelling the output gap or any other economic concept that involves an observed or unobserved trend. Moreover, if you choose to model at the lower frequency, you may lose important high frequency information. Further still, employing time aggregation methods may distort the overall dynamics of the model by inducing contemporaneous relationships that may not really exist.
You briefly mentioned seasonality, so let me touch upon that, too. If seasonally adjusted data is available from the statistical office (or wherever you source your data) then you may want to use that to simplify the modelling process. On the other hand, your preference may be to model the seasonality explicitly. Again, there are trade-offs to be aware of; how complex do you want your model to be and do you have a time-frame in which to build the model? You may also need to consider if the data you choose to model gets revised and how you'll update your model's dataset if you're to use it on a regular basis. Maintaining a good quality up-to-date dataset can be a challenge.
A: Generally, larger and larger number of observations leads to increase the accuracy and precision of your estimates. So, firstly, by assuming you will use all data you can find, using an high frequency data sample obviously increase your sample and, of course, the accuracy of your estimates.
Look here for a simple explanation of that, the minimum number of observation required and the math/statistics behind.
Secondly, increasing the number of observations reduces the sample noise because of the same reasons explained in that link I posted above.
Finally, the best thing to do is to use intraday data, even if it is pretty difficult to get, since usually they are available mainly in the financial industry.
A: The more the granular data better the results. Choosing lowest frequency data is better , in this case choosing monthly data is better as annual or quarterly data will mask seasonal trends as well as smooth out months where there are prominent holiday effects..Especially if the models are auto regressive in nature where the effect continues for some time. Choosing a higher granularity such as annual data can lead to "data aggregation bias" . There are some recovery methods.
A good way to test will be to run the model at every interval - quarterly, monthly and annual and test if there is a difference in coefficients ( of course after adjusting for temporal differences)
