Why do some people use -999 or -9999 to replace missing values? I have a dataset. There are lots of missing values. For some columns, the missing value was replaced with -999, but other columns, the missing value was marked as 'NA'. 
Why would we use -999 to replace the missing value? 
 A: This is a holdout from earlier times, when computer software stored numerical vectors as numerical vectors. No real number has the semantics "I'm missing". So when early statistical software had to differentiate between "true" numbers and missing values, they put in something that was "obviously" not a valid number, like -999 or -9999.
Of course, that -999 or -9999 stood for a missing value is not "obvious" at all. Quite often, it can certainly be a valid value. Unless you explicitly check for such values, you can have all kinds of "interesting" errors in your analyses.
Nowadays, numerical vectors that can contain missing values are internally represented as "enriched" numerical vectors, i.e., numerical vectors with additional information as to which values are missing. This of course is much better, because then missing values will be treated as such and not mistakenly treated as valid.
Unfortunately, some software still uses such a convention, perhaps for compatibility. And some users have soaked up this convention through informal osmosis and enter -999 instead of NA even if their software supports cleanly entering missing values.
Moral: don't encode missing values as -999.
A: Such values are for databases.  Most databases long ago, and many today, allocated a fixed number of digits for integer-valued data.  A number like -999 is the smallest that can be stored in four characters, -9999 in five characters, and so on.
(It should go without saying that--by definition--a numeric field cannot store alphanumeric characters such as "NA".  Some numeric code has to be used to represent missing or invalid data.)
Why use the most negative number that can be stored to signify a missing value?  Because if you mistakenly treat it as a valid number, you want the results to be dramatically incorrect.  The further your codes for missing values get from being realistic, the safer you are, because hugely wrong input usually screws up the output.  (Robust statistical methods are notable exceptions!)
How could such a mistake happen?  This occurs all the time when data are exchanged between systems.  A system that assumes -9999 represents a missing value will blithely output that value when you write the data out in most formats, such as CSV.  The system that reads that CSV file might not "know" (or not be "told") to treat such values as missing.
Another reason is that good statistical data and computing platforms recognize many different kinds of missing values: NaNs, truly missing values, overflows, underflows, non-responses, etc, etc.  By devoting the most negative possible values (such as -9999, -9998, -9997, etc) to these, you make it easy to query out all missing values from any table or array.
Yet another is that such values usually show up in graphical displays as extreme outliers.  Of all the values you could choose to stand out in a graphic, the most negative possible one stands the greatest chance of being far from your data.

There are useful implications and generalizations:


*

*A good value to use for missing data in floating-point fields is the most negative valid number, equal approximately to $-10^{303}$ for double-precision floats.  (Imagine the effect that would have on any average!)  On the same principle, many old programs, which used single-precision floats, used somewhat arbitrary large numbers such as 1E+30 for missing values.

*Adopt  a standard rule of this type to make it easy to invent NoData codes in new circumstances (when you are designing your own database software).

*Design your software and systems to fail dramatically if they fail at all.  The worst bugs are those that are intermittent, random, or tiny, because they can go undetected and be difficult to hunt down.  
A: Are there computed variables in the dataset? Or is this an analytic dataset that comes form merged / sorted data? Some software uses very large negative values to denote missing data. But other software creates missing values with NA or .. When they are discrepant, usually some post processing has led to disagreement. 
A: Of course, in SPSS, the missing value(s) 999 or whatever IS tagged as a special missing code and handled separately from other values.  It may be tabulated separately or excluded entirely.  A distinction is made from the result of things like zero division or log(0).
A: You can use anything to encode missing values. Some software, like R, use special values to encode missing data, but there are also software packages, e.g. SPSS, that do not have any special codes for missing data. In the second case you need to make arbitrary choice for such values. You can choose anything, but generally it is a good idea to choose some value that visibly differs from your data (e.g. your data are percentages in 0-100 range, so you choose 999 for encoding missing data, or your data is human age and you use negative values for missing observations). The idea behind it is that by doing so you should be able to notice if something went wrong and the numbers do not add up. 
The problem with such encoding is however that you actually can not notice the special encoding and end up with rubbish results. 
