Comparing employee sales monthly data despite regional differences I'm not exactly sure how to title this thread, or what tags to use since advanced statistics isn't really my thing right now. Please change the title and tags as needed to improve this question.
Say I have 10 different employees who sell newspaper subscriptions, and each employee covers a different geographical region of the state. I have data on the total number of newspapers sold  for each employee at the end of each month for a 12 month period, January-December.
What I want to know, is how do I normalize the numbers and get a performance evaluation of the employee with complete understanding that some geographical regions naturally perform better than others based on various demographical information. One employee might be doing very well in an area that traditionally has lower volume due to a lower population density, whereas another employee might have sold a ton of newspaper subscriptions but they operate in a region that has a much higher population density than the aforementioned employee.
Where do I start? What information do I need to make this happen? How much historical data do I need?
 A: First, a population that large, 10, is not going to give you anything at all. Even with good predictors from the demographics there will be a great deal of noise that is difficult to associate meaningfully even with more descriptive and elaborate algorithm designs.
Second, the particular normalization you use is dependent on the model you decide to implement. Each has its use, but overall the practice is useful in itself for some issues in effects of numerical performance. 
Third, determining a performance evaluation method could depend on doing some exploratory data analysis to determine some simple correlation measure to eyeball. Otherwise you design your model, by fitting the data you can determine impact on the newspapers and then get a vector of the difference between that predicted subscription adjusted and the original subscriptions over time by employee, then a some norm or any function you want can make that a scalar rating.
First start, determine your ability to design the model. It seems that you are starting off: there are many user-friendly statistics applications that will take that data so you can then attach the demographics and run a discrete time autoregressive analysis or similar discrete time analysis methods.
The problem is that this could be very sensitive, the more distinct employees, regional independent demographic data and samples helps but your model may need to include a number of dummy variables and or weight vectors for the employees. There are many reasons outside of the demographics that subscriptions could differ yet be explained by random chance, the employee, the purchasers of the subscriptions, the weather, etc. 
Ultimately it should be about learning, if you get to the point where you can customize the code of the algorithms determined by the underlying numerical linear algebra. One of the methods for model design techniques BellKor developed I liked and used in my research is to use the structured data more with multi-faceted methods to expose the relationships more explicitly. NetFlix implemented some, but ultimately had switch ideologies. 
A: (This will be rather long, but you are asking for a whole model here).
We have a "dependent" variable (Newspaper sales = $Y$), and we believe that this variable is affected by various environmental factors (like population density in each area), but also it is affected by the sales-employees' "intrinsic" performance. We want to separate these so as to compare the intrinsic performance of the employees.
This can be approached as a case of panel-data regression analysis, where the residuals from the regression will not measure "estimation error" but the unknown factor (the employee performance). The cross-section dimension $n=1,...10$ of the panel is the employees, and the time-series dimension of the panel $t=1,...,T$ is the monthly sales series (the higher the $T$ the better of course).  
Note that the panel framework is necessary in your case, because the environmental factors affecting newspaper sales are likely to be "long-term" - meaning that they will exhibit zero variability in short periods like in a monthly setting: such factors could be, population density, participation in elections (the more you care about commons, the more likely is to want to be informed about commons), education level (usually a positive effect on reading newspapers), income level...  These do not really vary month-per-month so you could not specify a separate time-series regression for each employee (the regressor matrix would exhibit perfect colinearity, hence it would be singular and non-inveritble, hence we could not obtain least-squares(LS) estimates. I don't advise doing anything more complex than LS regression, since you state that you are not very familiar with these techniques).  
Now the panel data model is formulated by writing 
$$ y_{it} = \beta_0 \,+\,  \beta_1x_{1it} \,+\, \beta_2x_{2it} \,+\, ... + \beta_kx_{kit} + v_{it}\qquad  i=1,...,n\qquad  t=1,...,T$$
where the $k$ $x$'s are the factors that you believe influence newspaper sales. For example let's say $x_{1it}$ is the population density in the area of employee $i$. Given the previous discussion, most likely you will have $x_{1i1} = x_{1i2}=...=x_{1it}$, but different from $x_{1j1} = x_{1j2}=...=x_{1jt}$. The $\beta$s are the marginal effects these explanatory variables have on newspaper sales, they are unknown and they are the subject of estimation... Note that we assume that the marginal effect of each regressor is the same in all areas. Where is employee performance?
It is in $\beta_0$ and in $v_{it}$. The $\beta_0$ does not have an explanatory variable attached to it. So it is assumed equal (in levels not just as a marginal effect) along all cross-sections, meaning that it is the average sales per employee irrespective of the month or the area; it is the long-term average performance (measured in sales) of your salesforce viewed as a whole. Then, $v_{it}$ can be viewed as expressing how the sales in each area and in each month deviated relative to this pooled, long-term average, after we have also separated the effects of the $x$'s.
Then: you run this panel-data model applying the most basic panel-data estimator, the Pooled OLS estimator. You obtain estimations of the $\beta$s (which is not your prime focus of interest), but also you obtain time series of residuals: for each employee i you obtain $\hat v_{i1},\; \hat v_{i2},...,\hat v_{iT} $. Now if, say, $\hat v_{i2}>0$, it can be interpreted as "in month $2$ employee i performed above the long-term average of the salesforce". Then count how many times for each employee the $\hat v_{it}$ were positive (=number of times performance was above salesforce average) and how many times they were negative (=number of times performance was below salesforce average). This is a valid quantitative base to evaluate relative performance between employees (I wouldn't make it the sole criterion for final evaluation though). Just counting above-below average without taking into consideration "how far from average", has the advantage of not being influenced by idiosyncratic one-time large deviations: it gives you a sense of whether each employee performance is "long-term" above average, or "long-term" below average. Remember that this average comes from your own salesforce, it is not some average from all employees working in newspaper sales in the industry - but which is only fair, since you want to compare the performance of your own employees.
Three final notes:
First, a short-term phenomenon affecting newspaper sales may be the season (winter against summer, etc). So it may be the case that the variability observed in each month, even after we separate the effect of the other regressors and of the pooled salesforce average,  comes partly from which month is it, not just from employee intrinsic performance (i.e this seasonality effect is included in the errors and hence in the residuals). But if it so happens that you can assume that seasonality affects proportionately the same each area (say, newspaper sales go down by same percentage in summer in all areas), then you can ignore this aspect. Otherwise you should specify as regressors dummy variables for each month (minus one month), and things would get a bit more complicated then.
Second, a note on the number of regressors, i.e. how large $k$ will be. You will be estimating $k+1$ unknown coefficients, or $k+1+11$ if you include the dummy variables for the months. You sample consists of $nT$ observations. Tyr to keep your degrees of freedom as large as possible, without omitting factors that you strongly believe influence newspaper sales. There is no formula to determine "how many degrees of freedom".
Lastly, the dependent variable may enter the picture in levels (i.e. monthly sales volume of newspapers), or in its natural logarithm. This last specification has usually some desirable effects on the technical aspects of the estimation procedure (like stabilizing the variance of the error term) and it is to be preferred. In such a case, the $\beta$s will represent percentage  effects, and so will the residuals. But this doesn't affect the approach where you count how many times the residuals are positive, and how many times they are negative.
