Should I prefer pointers or references in member data?

My own rule of thumb :

Use a reference member when you want the life of your object to be dependent on the life of other objects : it's an explicit way to say that you don't allow the object to be alive without a valid instance of another class - because of no assignment and the obligation to get the references initialization via the constructor. It's a good way to design your class without assuming anything about it's instance being member or not of another class. You only assume that their lives are directly linked to other instances. It allows you to change later how you use your class instance (with new, as a local instance, as a class member, generated by a memory pool in a manager, etc.)

Use pointer in other cases : When you want the member to be changed later, use a pointer or a const pointer to be sure to only read the pointed instance. If that type is supposed to be copyable, you cannot use references anyway. Sometimes you also need to initialize the member after a special function call ( init() for example) and then you simply have no choice but to use a pointer. BUT : use asserts in all your member function to quickly detect wrong pointer state!

In cases where you want the object lifetime to be dependent on an external object's lifetime, and you also need that type to be copyable, then use pointer members but reference argument in constructor That way you are indicating on construction that the lifetime of this object depends on the argument's lifetime BUT the implementation use pointers to still be copyable. As long as these members are only changed by copy, and your type don't have a default constructor, the type should fullfil both goals.

Avoid reference members, because they restrict what the implementation of a class can do (including, as you mention, preventing the implementation of an assignment operator) and provide no benefits to what the class can provide.

Example problems:

you are forced to initialise the reference in each constructor's initialiser list: there's no way to factor out this initialisation into another function (until C++0x, anyway edit: C++ now has delegating constructors)

the reference cannot be rebound or be null. This can be an advantage, but if the code ever needs changing to allow rebinding or for the member to be null, all uses of the member need to change

unlike pointer members, references can't easily be replaced by smart pointers or iterators as refactoring might require

Whenever a reference is used it looks like value type (. operator etc), but behaves like a pointer (can dangle) - so e.g. Google Style Guide discourages it

Objects rarely should allow assign and other stuff like comparison. If you consider some business model with objects like 'Department', 'Employee', 'Director', it is hard to imagine a case when one employee will be assigned to other.

So for business objects it is very good to describe one-to-one and one-to-many relationships as references and not pointers.

And probably it is OK to describe one-or-zero relationship as a pointer.

So no 'we can't assign' then factor. A lot of programmers just get used with pointers and that's why they will find any argument to avoid use of reference.

Having a pointer as a member will force you or member of your team to check the pointer again and again before use, with "just in case" comment. If a pointer can be zero then pointer probably is used as kind of flag, which is bad, as every object have to play its own role.

Use references when you can, and pointers when you have to.

In a few important cases, assignability is simply not needed. These are often lightweight algorithm wrappers that facilitate calculation without leaving the scope. Such objects are prime candidates for reference members since you can be sure that they always hold a valid reference and never need to be copied.

In such cases, make sure to make the assignment operator (and often also the copy constructor) non-usable (by inheriting from boost::noncopyable or declaring them private).

However, as user pts already commented, the same is not true for most other objects. Here, using reference members can be a huge problem and should generally be avoided.

As everyone seems to be handing out general rules, I'll offer two:

Never, ever use use references as class members. I have never done so in my own code (except to prove to myself that I was right in this rule) and cannot imagine a case where I would do so. The semantics are too confusing, and it's really not what references were designed for.

Always, always, use references when passing parameters to functions, except for the basic types, or when the algorithm requires a copy.

These rules are simple, and have stood me in good stead. I leave making rules on using smart pointers (but please, not auto_ptr) as class members to others.

Yes to: Is it true to say that an object containing a reference should not be assignable, since a reference should not be changed once initialised?

My rules of thumb for data members:

never use a reference, because it prevents assignment

if your class is responsible for deleting, use boost's scoped_ptr (which is safer than an auto_ptr)

otherwise, use a pointer or const pointer

I would generally only use a pointer in member data when the pointer could be null or could change. Are there any other reasons to prefer pointers over references for data members?

Yes. Readability of your code. A pointer makes it more obvious that the member is a reference (ironically :)), and not a contained object, because when you use it you have to de-reference it. I know some people think that is old fashioned, but I still think that it simply prevent confusion and mistakes.

I advise against reference data members becasue you never know who is going to derive from your class and what they might want to do. They might not want to make use of the referenced object, but being a reference you have forced them to provide a valid object. I've done this to myself enough to stop using reference data members.