In general it is not advisable. However, there are some situations where it can be ok.

One is when you don't have many responses for some categories. For example, imagine you have a dataset with 100 "nevers", 70 "rares", 3 "sometimeses", and 0 "alwayses". You wouldn't really be able to make any conclusions about an ordinal effect, so it would make more sense to split the data into "never" vs. "not-never". I'm not aware of technical articles on this issue (I'm sure some exist, and you can probably find them via Google, I just can't think of them off the top of my head), but as for examples of papers that do it, one of my own papers does: http://www.mypolyuweb.hk/~sjpolit/pubs/Politzer-AhlesZhang_2014_JCL.pdf the variable called 'Repeated' is actually a dichotomization of an ordinal variable, as explained in the paper.

Also, sometimes dichotomizing the data can help to visualize or explore it, i.e., you might be able to make a plot that's easier to understand. But generally in that case it's recommended to just use the dichotomization to help understand the data, not to actually use dichotomized data for your statistics.

In general it is not advisable. However, there are some situations where it can be ok.

One is when you don't have many responses for some categories. For example, imagine you have a dataset with 100 "nevers", 70 "rares", 3 "sometimeses", and 0 "alwayses". You wouldn't really be able to make any conclusions about an ordinal effect, so it would make more sense to split the data into "never" vs. "not-never". I'm not aware of technical articles on this issue (I'm sure some exist, and you can probably find them via Google, I just can't think of them off the top of my head), but as for examples of papers that do it, one of my own papers does: http://www.mypolyuweb.hk/~sjpolit/pubs/Politzer-AhlesZhang_2014_JCL.pdf the variable called 'Repeated' is actually a dichotomization of an ordinal variable, as explained in the paper.

Also, sometimes dichotomizing the data can help to visualize or explore it, i.e., you might be able to make a plot that's easier to understand. But generally in that case it's recommended to just use the dichotomization to help understand the data, not to actually use dichotomized data for your statistics.

When your the researcher you're the HNIC, if you can justify it than by all means go for it. I would take a binary logistic regression over a taxing polytomous regression anyday. It is often easier to justify a dichotomous decision as people have a difficult time saying they always or never do something offering extremes (of which I have a hardtime doing even if I feel that way. Also, while Stephen brings up a goodpoint of missing responses. if that was the case you would just group them into two groups anyways (there is your justification by the way which hopefully you can see the dramatic/comic irony here and no one need be insulted as I make an ass of myself on  a daily basis). It's all about cut-offs completely continuous variables can be made ordinal or even dichotomous if necessary (this is discussed mildly in Hosmer & Lemeshow's book of applied regression).

Regarding issues dealing with the use of hyperbolic rhetoric when never and always are options it would dichotomize the question myself and then come up with a scaling for the importane of the question by using some form of Thurstone scaling. I hope my comments were helpful.

All the Best,

Logan Netzer

Tanstafl, Logan.  There ain't no such thing as a free lunch.  Those choices have consequences, sometimes serious ones.  Dichotomizing or otherwise turning continuous or ordinal variables into categorical ones is a very problematic practice except under very specific circumstances.  There are many published papers citing why.

Your point about people struggling with rating scales is a good one, and many researchers would be better served by asking a series of dichotomous questions.  People find it much easier to answer a question like 'do you think person A or person B is more attractive, racist, or creative'  versus using a rating scale.  However, the question at hand involved ordinal data so dichotomizing after the fact is questionable unless only two responses predominated.

Converting the Likert scale to binary variable is totally valid. But the point is that the conclusion and inference we draw will be a bit different. The models used to explain the data will be different. Nothing wrong I see. 

Collapsing cell values incurs loss of information, which may render the results less valid than it originally is. But, then, one needs to be practical, doing what is possible and makes sense.

Why the researcher chose to use a Likert scale is the question that comes to my mind in the first place. My common sense guess to this question is that he/she predicted some variability that would not be captured by two categories. Being the collected data difficult to interpret might lead to the conclusion that the prediction failed, and changing the codification strategy could be interpreted as a backward prediction, something that I do not follow. Finding a strategy to account for the uninterpretable variability is the challenge I would engage in.

As a researcher you can literally do whatever you wish, as long as the choice is based upon some legitimate research methodology, preferably supported by literature. Your proposed dummy coding will work in a variety of statistical tests, but you must take into account the scale you are using for interpretation of results. You also use the word "binary" but you have more than two (bi) choices. What you are proposing is an ordinal scale. Perhaps you could dichotomize by say never/rarely = 0 sometimes/always = 1. 

See:

http://www.icbl.hw.ac.uk/ltdi/cookbook/info_likert_scale/printable.pdf

http://www.theanalysisfactor.com/likert-scale-items-as-predictor-variables-in-regression/

As others have pointed out, any time that you simplify the coding of a variable into fewer categories, you lose some of the variance in the data. Since parametric statistics are all based on the concept of variance as an interesting property, one that van be usefully partitioned into variance components causally (at least implicitly) linked to other variables in the system under investigation. Partitioning of variance leads directly to partitioning of the explanations for that variance. What you want to avoid is reducing the variance in your data so much that you run out of interesting variance to partition.

That said, it does depend considerably on how finely the concept being measured can be refined and categorized, categorically or ordinally. If there are really only two alternative states, then no matter how many fine divisions are introduced into the measurement, you'll only be able to measure which of the two states the participant is in. The more you know about the values that the data are likely to take, the better measures you can craft. It's sad to see people spending so much time crafting fine-ground measures for concepts that are conceptually cloudy. This is likely to introduce more noise than substance into the data. While you want to measure all the variance that is legitimately present in the data (in the interests of good analysis), random noise introduced by over-precise measurement not only doesn't help; it can actually severely mislead the investigators.

Having read your message a second time, it appears to me that you are interested in the extent or frequency of using the medical service and not the effect (outcome) of the service. If this is correct, perhaps you can ask for the frequency of use (e.g., no. of visits to the clinic/doctor). The actual record will give you better (more valid) data than Likert-ratings (which are more suited for attitudes -- how the participants feel about the treatment, for instance.

Dear Sanaullah, I think that collapsing the variable into a  dichotomous variable is reasonable. You may well lose some information, but if it's information you don't want then that's fine. If the aim of your study is to predict service use, rather than frequency of service use, then proceed and explain what you have done in the paper/report. Most people will accept that it's a pragmatic approach, accept there are some unwanted statistical outcomes, and then be more interested in what your analysis actually means.

Soh's re-reading prompted my own; so I'll likewise amend my answer. In practical terms, much depends on the distribution of your variable. I'm assuming that you're inquiring about data already collected using your four-point scale, rather than a future data collection. (If this question is for an unfinished questionnaire, then I agree with Soh that you should seek behavioral data rather than opinion data). With data in hand, you'd need to assess the distribution of the variable. The problem with four-point ordinal scales is that they often lead to highly skewed and/or bimodal distributions - e.g., "satisfaction" items where 90% of respondents report "satisfied" or "highly satisfied" results. Trying to use such data in quantitative models that assume the underlying variables to be more or less normally distributed can result in wildly wrong results. With such variables, dichotomous recoding into high/low groups with sizes as close to equal as possible is your best strategy; the variance you're losing is pretty much guaranteed to be mostly noise, as I noted earlier.

If you really have a variable with something that looks like a normal distribution, then you can use it in your model. But please remember that if you're using some form of regression-based modeling, it was designed for variables that have a wide range of values as well as normal distribution, so you're pushing things a lot to use a variable with only four values even if they are well distributed.

Mark is quite correct that you need to think carefully about just what phenomenon you really have the ability to make predictions about. The overall question is as much or more conceptual as it is statistical.

....sorry, I have to comment again. We frequently collapse data such as this and it's fine. For example, surveys often ask about education level in some detail and this is commonly summarized in a smaller number of categories, indeed sometimes  down to binary level (e.g. completed 13 years of education, yes/no). If the restructuring of the data is logical and transparent, and the subsequent statistical treatment is appropriate then things are OK.   

In Sanaullah's case there is a simple transformation from a coarse measure of frequency to a pretty good measure of occurrence. If he does that and calculates frequencies, uses chi-square or logistic regression the world will keep turning, he'll get sensible results, and the research is likely to add to our knowledge.