16
When is it safer to say nothing? Some considerations on biases in sampling In nursing research, once one has a solid design, one has still to think about a sampling strategy and implementa- tion. Too often, the paraphernalia of inferential statistical reasoning is inappropriately deployed when the achieved sample can in no way be claimed to represent the drawn sample. Given the traditionally low rates of response in most nursing research (usually well under 90 per cent, and often unknown), there is a danger that perfectionist coun- sels would lead to an end to serious research. In this paper, Laurence Moseley and Donna Mead argue that such a nihilistic position is not necessary and that, instead, researchers should tailor their inferential analyses to the demands of any particular study They argue that for many purposes, simple computations of both maximum and min- imum population estimates are both defensible and useful 9 > sampling g > non-response > drawing inferences > generalisation Introduction This is not a general paper about research design. In it we assume that all the usual safeguards are in place. For example, any questionnaire has been tested for comprehension; scales have been shown to be psychometrically 20 NURSEiRESEARCHER volume 12 number 1

Sampling for nurses

Embed Size (px)

Citation preview

When is it safer to say nothing?Some considerations onbiases in samplingIn nursing research, once one has a solid design, one has

still to think about a sampling strategy and implementa-

tion. Too often, the paraphernalia of inferential statistical

reasoning is inappropriately deployed when the achieved

sample can in no way be claimed to represent the drawn

sample. Given the traditionally low rates of response in

most nursing research (usually well under 90 per cent, and

often unknown), there is a danger that perfectionist coun-

sels would lead to an end to serious research. In this paper,

Laurence Moseley and Donna Mead argue that such a

nihilistic position is not necessary and that, instead,

researchers should tailor their inferential analyses to the

demands of any particular study They argue that for many

purposes, simple computations of both maximum and min-

imum population estimates are both defensible and useful

9 > samplingg > non-response

> drawing inferences> generalisation

IntroductionThis is not a general paper about research design. In it we assume that all

the usual safeguards are in place. For example, any questionnaire has been

tested for comprehension; scales have been shown to be psychometrically

20 NURSEiRESEARCHER volume 12 number 1

adequate: question ordering has been addressed: and so on. This is a large

assumption. However, if those issues have not been addressed, the study is

likely to produce useless, or even dangerous, results, and should not even

be attempted. That is a severe criterion, but it is the least that we can do to

produce research results which are (a) positively useful and (b) negatively at

least not misleading. If those methodological basics have not been met. any

processes which are intended to permit generalisation - such as calculations

of power, needed sample sizes, confidence intervals, or p values - serve

only as the window dressing of science, and should not lead readers into

accepting the study for more than it is.

In this paper, instead, we address the very nature of the units (people,

wards, trusts, etc) that we select to study. In particular, we shall be looking at

the ways in which sampling, recruitment and response rates may lead to false

conclusions (and what can be done to prevent that happening). The prob-

lems which we discuss are particularly clear cut when dealing with experi-

mental research, but in principle they can arise in any sort of research, from

the most simply descriptive to the most complex of multivariate designs. Of

course, when undertaking simple exploratory descriptive work, the problems

to which we refer do not arise, because, in any case, one will not have reli-

able, representative or generalisable results. Nonetheless, even with this sort

of work, it is still worthwhile at least considering whether the sampling

method on its own can lead to biased results. The most strongly designed

study in the world can produce uninterpretable results if the sampling is not

well done. If both the study design and the sampling are weak, the dangers

are insurmountable.

To summarise, to produce information that will make life better for

patients, we need both good measurement and good sampling. In this paper,

we deal with some elements of the latter

Why sample?We could try to study the whole universe of interest. There are three main

reasons why we do not, and instead undertake studies of samples:

n The first is that in many situations taking a census of the whole universe

of interest (all mental health patients in Wales, all nurses in the UK. or

volume 12 number 1 NURSEKESEAKCHER 21

whatever that universe of interest might be) is prohibitively expensive

or impracticable.

D The second is that the vagaries of non-response may well mean that we

miss many sub-groups in the universe of interest (for example, cardboard

box dwellers [in a population census], patients with cognitive problems [in

a patient satisfaction study], some ethnic minorities [in midwifery studies]),

and we might do better to target those sub-groups directly if we are inter-

ested in them. It is usually better to direct one's energies into contacting a

smaller sub-group (a sample) and obtaining a 100 per cent response, than

in trying to contact a larger sub-group but obtaining only a small response

rate - even if the absolute numbers obtained in the latter case are larger.

n However, the most important reason is both methodological and psycho-

logical. A major reason for drawing random samples is to overcome the

danger that we may be misled by our clinical or other experience. It is in

the nature of experience that it is gained in one setting, or a small num-

ber of settings. This will usually cover a particular professional discipline

or a small number of specialities and will, in general, give only a partial

picture of the whole of the reality about which one wishes to draw con-

clusions. Even if we have the most solid of research designs and have reli-

able data about our patients, the ward, or whatever, we cannot make a

reliable judgement about how representative our experience is. If our

experience suggests to us that a particular treatment works with a particu-

lar sort of patient, we cannot be confident that this is more generally true

unless we undertake some sort of controlled experiment. This tendency to

sampling bias in making clinical or experiential judgements is one of the

most solidly grounded findings in cognitive science. Little is known more

firmly than that human judgements are influenced for the worse by fac-

tors such as recency, salience, anchoring, hindsight bias, ignoring base

rates and other factors associated with complexity (Kahneman et al 1982).

If we do manage to draw and then obtain reliable data on a sample that is rep-

resentative of the universe of interest, there are well-known statistical techniques

that enable us to draw inferences from that representative sample to the universe

from which it is drawn. These are techniques that have been developed over

decades of work. They cover the dangers that may arise from random effects in

22 NURSERESEARCHER volume 12 number 1

sampling. Those, however, represent only one of the dangers that can lead to

misinterpreting the findings of research. They do not, for example, take account

of the dangers of interpretation that are posed by incomplete coverage, includ-

ing non-response. Overall, the most important funaion of random sampling is to

protect us from the errors of judgement that are likely to arise by trying to learn

merely from unstructured and inconsistently recorded experience.

Universe, drawn sample and achieved sampleThree terms are often used and we try to illustrate how they are different and

why each of them is important. They are: (A) universe (or population), (B)

drawn sample, and (C) achieved sample. The reason we labour somewhat

over these distinctions is that the latter two are often not distinguished from

each other. The diagram in Figure 1 is intended to illustrate the situation.

Legitimate and illegitimate influences

A. Universe/Population The group about which we wish to draw --~;',

general conclusions A

Legitimate inference

B. The complete drawn sample The representative sample on which we

plan to gather data i

Constrained inference

C. The achieved sample The sub group of the drawn sample about

whom we succeed in actually gathering data

(but what is usually done!)

What we have are the following: -

A. Universe/population The group about which we wish to draw

general conclusions

B. The complete drawn sample The representative sample on which we

plan to gather data

C. The achieved sample The group about whom we succeed in

actually gathering data

volume 12 number 1 NURSERESEARCHER 23

What we want to know about is the universe or population of interest. If

we have complete data about the drawn sample (B), drawing inferences from

it to the universe (A) is straightforward and should be influenced only by ran-

dom fluctuations, and there are well tried and tested statistical techniques for

estimating the degree of confidence which can reasonably be placed in those

inferences. It should be a mechanical procedure, with little or no scope for

judgement. Unfortunately, in practice, researchers do not have full data

about the drawn sample, partly because the record-keeping which would

permit such data to be gathered is often incomplete, inaccurate, out of date,

and not organised for easy searching, but mainly because response rates are

usually less than 100 per cent, so the achieved sample is only an incomplete

version of the drawn sample. Nonetheless, most reports attempt to draw

inferences about the universe (A) from the data obtained on the achieved

sample (C), rather than data on the drawn sample (B), This is in principle ille-

gitimate. In practice, the dangers which it introduces depend largely upon

how well the achieved sample represents the drawn sample, although it may

also depend upon the actual inferences which are to be drawn.

When there is a non-response, one should always consider how the nature

of the task may reasonably influence the degree to which our judgement may

vary according to the congruence of the drawn and achieved samples. In some

cases, a 90 per cent response rate may make little or no difference to our

judgement; in others, it may completely vitiate the results. If it is not possible

to obtain data about the people, wards etc of interest, and, in particular, if you

do not obtain a 100 per cent response rate from them, two problems arise.

Firstly, one may not be able to describe even one's drawn sample adequately.

If the response rate is, for example, 50 per cent, so that the achieved sample

represents only half those whom you had wished to study, then there will be

mysteries even about what the drawn sample was like, and further work will

be needed to see what biases may be involved. Secondly, severe limits are

placed upon any conclusions which may in turn be drawn about the universe.

Non-response is so common that researchers often report the ways in

which their achieved sample differs from their drawn sample. In particular, it

is common to compare the two in terms of gross demographic variables,

especially age, sex, and social class. If, on these measures, the achieved and

24 NURSEIi?ESEARCHER volume 12 number 1

drawn sample are similar then the researcher will tend to conclude that the

recruitment and response parts of the study have introduced no biases. In

that case, they assume that the achieved sample equals the drawn sample,

and one can treat them as equivalent when calculating one's confidence

intervals, p values and the like. We wish to argue that such an assumption is

dangerous, and that such gross comparisons should be seen as the first, not

the last, step in checking on the representativeness of the achieved sample,

A study by Banerjee et al (1996) provides an illustration of the dangers. The

paper comes from a prestigious journal (the British Medicai Journal), is fairly

recent, cites some scientific-looking numbers (odds ratios) and draws strong

conclusions which, if wrong, could adversely affect many people from two hard-

pressed minorities (elderly and disabled people), Banerjee et al claimed very

bluntly: 'Depression is treatable in elderly people receiving home care.

Therapeutic nihilism based on an assumed poor response to treatment in these

socially isolated, disabled elderly people in the community is not supported.'

That is very strong and one would be tempted to accept such a conclusion.

However, they reached this conclusion on the basis of a study that started with

441 patients of whom 69 were successfully recruited. This gave a response rate

of 15 per cent. So little is known about the other 85 per cent that such a strong

conclusion should not be drawn, despite the apparent paraphernalia of science

surrounding the results. The authors also report that of the experimental subjects

58 per cent improved (compared to only 25 per cent in the control group), a

difference which they report as 33 per cent. Of course, it is nothing of the sort.

The difference is 33 percentage points (not 33 per cent). How many per cent

that represents depends on what it is a per cent of. If one takes the control group

as the baseline, then the experimental group are 33 x 100/25 or 132 per cent

improved. If one takes the experimental group as the baseline then the control

group are 33 x 100/58 or 57 per cent deteriorated. Clearly one needs a higher

response rate than 15 per cent before drawing such dramatic conclusions.

Censuses covering the whole universe of interestWith some specialised universes, it is possible to define and locate 100 per

cent of the population of interest. For example. Mead (1983) was interested

in assessing to what extent primary nursing was being practised in hospital

volume 12 number 1 NURSERESEARCHER 25

wards in Wales. She was able to locate, at a given point in time, all wards

and ward sisters in the principality. There were 91 Oof them, and, after a year

of driving up and down minor roads, arranging briefing sessions, using facil-

itators to ease her entry and so on, she managed to obtain completed ques-

tionnaires from 655 of them, giving a 72 per cent response rate. This was 72

per cent of the whole universe of interest, not of a sample, so it would have

been inappropriate to undertake inferential statistics. One might think that it

would be safe to draw conclusions about such a universe. However, the non-

response poses a problem. We have only the crudest of background infor-

mation about the non-responding wards, and it is unlikely that such infor-

mation is relevant to disambiguating our inferences.

Suppose that half of those who responded felt that their ward was doing

primary nursing. What we actually know is that 50 per cent of the 72 per

cent who responded felt that. Strictly speaking, therefore, all we can say for

certain is that 0.5 of 72 per cent felt that, that is, 36 per cent were in their

view doing primary nursing. We know nothing about the non-respondents.

If they were just like those who did respond, then, similarly, 0.5 of 28 felt

that they were doing primary nursing, that is, 14 per cent. Adding the two

together, we get 14 -i- 36 = 50 per cent. However, a lot hangs on the

assumption that non-respondents and respondents are similar.

If, for example, all the non-respondents were also doing primary nursing,

our combined population estimate would rise to 64 per cent (the 36 plus all

of the 28). If, on the other hand, none of the non-respondents were doing

primary nursing, our combined population estimate would remain stub-

bornly stuck at 36 per cent (36 -I- 0). So, the characteristics of the non-

respondents matter. Depending on the validity of the various assumptions,

the true figure could be 36 per cent, 50 per cent or 64 per cent.

This means that because of the non-response, precise estimates on the

topic of primary nursing cannot be made. That does not mean that we can

make no estimates, simply that we can make no precise estimates. We can at

the very least make estimates to represent the minima and maxima possible.

Thus, we can say that at least 36 per cent thought that they were doing pri-

mary nursing. Similarly, we can say that at most 64 per cent were doing pri-

mary nursing. That is a pretty broad band, but at least one is not forced to

26 NURSERESEARCHER volume 12 number 1

say absolutely nothing, and for some questions such minima and maxima

may be adequate as answers.

Thus, even with a good response rate (72 per cent), one cannot draw all the

conclusions one would wish to, although one can with ingenuity make some

useful population estimates. In practice, response rate figures vary widely. At

the high end. we know of one study which achieved a rate of 96.3 per cent

(Thomson et al 1999). However, some areas do much worse than this (for

example, studies in prison often achieve only in the low 30 per cent). In a pre-

vious study of prison officers that we had undertaken, the response rate was

only 34 per cent, and we considered abandoning the study (Mead and

Moseley 2004). However, when we looked at the UKCC (1999) report. Nursing

in Secure Environments, we discovered that response rates in the low 30s are par

for the course in prison studies. In some areas, such as cancer care, response

rates are often under 10 per cent (Mayor 2000). Clearly, with such low rates,

we cannot say all that we would like to say about the universe. However, with

some thought we can say something that is likely to be of interest or use. The

main problem is that our estimates are likely to be within very broad bounds.

What one should not do is to analyse the data gathered and then mindlessly

ask the computer to produce p values and confidence limits without giving any

consideration to what biases may have been introduced by the low response

rate. However, such analysis in fact happens commonly in nursing research.

Samples are usually neededBecause of the costs and practicalities involved in censuses, samples are usu-

ally taken. The first step is to ensure that you know what your universe of inter-

est is. Then a sampling frame has to be found and a sample drawn. Finally,

one has to do the work to try to obtain one's achieved sample. Every effort

should be made to ensure that the drawn sample is representative of the uni-

verse, and that the achieved sample is close to the drawn sample. Often,

researchers do not have the resources even to attempt to draw a sample which

is either demonstrably representative of the universe or at least is likely to be

biased only by random fluctuations. There are many variations. For example,

if it is likely that that responses will be different for different groups, and those

groups vary a lot in size, there is a good chance that you will under-represent

volume 12 number 1 NURSERESEARCHER 27

the smaller groups. In such cases, we often use stratified sampling, to ensure

that these smaller groups do not drop out of consideration. For examples, see

Riffenburgh (1999). Similarly, one may use cluster sampling, for example tak-

ing all the potential respondents in a county, and then trying to contact them.

Such methods are quite technical, and often leave some possible biasing fac-

tors uncontrolled. It is therefore quite common to use convenience sampling

to get hold of whatever you can, or snowballing/cascading to ask the first sam-

ple members drawn to suggest the second tranche and so on. Clearly, such

methods should not engender the degree of confidence in one's results that a

properly drawn random sample can offer. Should we therefore throw up our

hands and say: 'Well, we can't do research at all.'?

We believe that the answer to that is 'No'. Some conclusions can be

drawn, but if our sampling methods are weak, we should be very cautious

about how strongly we draw those conclusions, and in particular should sug-

gest how one might confirm or disconfirm our tentative conclusions (specu-

lations, hunches etc) by means of further work. One can go further. If the

measurement element of a study is badly done it cannot be used in a cumu-

lation of evidence. We should not bundle up ten badly done studies and

somehow claim that they become more persuasive because of their sheer

volume. For the sampling element, however, cumulation can be useful. If

many people have undertaken solid separate studies in. say. a variety of spe-

cialities, then it may be possible to treat them as a single sample from which

conclusions may be drawn about the specialities combined. For this to be

trustworthy, the measurement elements of the studies will need to be strong,

and in particular, the studies should use identical measures, or at least meas-

ures that can plausibly be interpreted as indicating the values of the same

variable. If this fits into a programme of replication, then the cumulation of

evidence can do much to increase our confidence in the overall picture. This,

of course, underlies the practice of meta-analysis.

Defining the universeOften, it is initially unclear what the universe is about which you wish to

draw conclusions. You may wish to study 'all hospital wards', or 'all hospital

wards, excluding mental illness'. It might be 'all patients with stays of longer

28 NURSERESEARCHER volume 12 number 1

than three days in medical and surgical wards', or 'all elderly patients with

no regular contact with their CP'. We raise this point, for two reasons. Firstly,

how you define your sample is likely to affect where and how you look for

your data, and also how much preliminary effort you may need to expend

even to draw up a sampling frame. Take the question: 'How would you

obtain a sample of elderly people?' Probably the only place where you

would find a reasonably comprehensive list of such people would be at the

Department of Work and Pensions. Even if the problems of confidentiality,

access, and recruitment could be overcome, there may be important groups

who do not appear on the appropriate records at that department: for exam-

ple, people who had been working only in the black economy.

Secondly, you might think of using the electoral registers for a randomly

selected sample of households. Once again, not everyone is on that register

as entry to it is voluntary. Indeed, during the political battle over the poll tax,

one of the arguments against such a tax was that it would lead people not to

enter their names on the register.

Similar considerations apply for many groups you may wish to study. We

have experienced problems when trying to obtain a sampling frame, and

ultimately a sample, of:

D elderly people

D people suffering from severe and enduring mental illness

D reformed criminals

n nurses who are not currently working in their profession

n patients suffering from particular conditions.

Approaches to samplingOne of the difficulties which arises is that we tend to think of obtaining a list,

applying a random number algorithm to it, and then selecting names and

addresses. To do that, we need to start off with a list- a sampling frame. That

list is often difficult or impossible to obtain. However, that is not the only

way of approaching the problem of representative selection. At least two

other principles can be used: sampling by time, and sampling by place. One

common form of sampling by time is to take consecutive patients: for exam-

ple, the first, second, third etc patient who turns up to a fracture clinic. You

volume 12 number 1 NURSERESEAKCHER 29

might also wish to use some other fraction, and take the first, eighth, or

whatever patient. Even then, possible biases, such as periodicity, have to be

considered. If the length of your research interview and the length of the

consultations are related, you might miss a number of patients who have

particularly long (or short) consultations, and thus get a sample that is not

genuinely representative. For example, not undertaking any data gathering

in the early hours of Saturday or Sunday morning may well omit some

important sections of the accident and emergency department clientele. The

general point is that you need to be on the lookout for such possibilities.

A second possible form of sampling is by place - a method that we have

used in a study of the elderly. You select at random a number of addresses

from the electoral register (or by map grid references) and then approach the

residents at those addresses to try to locate any elderly people living there.

The randomness of the address selection safeguards against many biases. You

still, however, have to be careful about other sources of bias. If you take, say,

the first elderly person who comes to the door, you might miss those elder-

ly people who are disabled, bedfast or otherwise less active. If you ask if

there is an elderly person resident in the household whom you wish to inter-

view, there is a chance that your immediate contact will, for example, omit

to mention their confused, or otherwise 'difficult' relatives.

For some particularly difficult cases, you may need to do the sampling in

two sweeps - one to draw up the sampling frame, the other to draw the actu-

al sample. That is a method that we have used to find people who suffer from

severe and enduring mental illnesses (working initially through community

mental health teams, even before drawing up the real sample). It could also

be used to find retired nurses. If you have no initial sampling frame, no sin-

gle method is perfect. However, with some ingenuity, you can normally find

a method which minimises known or imaginable sources of bias, and which

enables you to make some estimate of the likely effect of any remaining ones.

The general message is that it is not usually simple or straightforward to

obtain a sample of representative respondents. Just taking some people who

are handy and willing to participate (convenience sampling) is usually inad-

equate. Of course, if the universe in which you are interested is only your

local hospital, and if you think that omitting unwilling participants is unim-

30 NURSERESEARCHER volume 12 number 1

portant, that method may be marginally acceptable. For obtaining more gen-

eralisable results, it is not.

Size of the sample needed to draw conclusions with confidenceWhen estimating how many people you will need in your drawn sample (and

preferably in your achieved sample) at least two numbers are usually

involved: (a) the total number of people in the universe (and particularly the

size of your intended sample), and (b) the size of phenomenon or effect

which you wish to measure. All sample measurements are subject to some

random error. Suppose that you think that a particular way of organising nurs-

ing would be useful to implement, if it reduced the average length of stay by

20 per cent. You undertake a sample study to check that this actually happens.

If in your achieved sample (and that achieved sample is close to your drawn

sample), the reduction actually is 20 per cent, the search is not complete. Your

observed 20 per cent is in the achieved/drawn sample. In the universe of

interest, it might be higher or lower, with a calculable probability. You will

need to undertake, even before starting the study, an estimate of the confi-

dence limits within which the universe figure will lie. How wide these confi-

dence limits are will depend mainly on how variable the phenomenon is (do

most people stay about five days, or do some stay one day and others 38

days), and on the size of the sample. With a large sample (say in the hun-

dreds) a figure of 20 per cent in your sample may imply that, with a known

degree of confidence, in the universe the real figure will lie within the range

of 18 to 22 per cent. Such a narrow range is likely to be useful for making pol-

icy or clinical recommendations. However, if your sample is very small (say in

the tens), then the range may not be from 18 to 22 per cent, but from 5 to

35 per cent. Such a wide range is likely to take you outside your comfort zone

when it comes to making recommendations. We would stress that for some

purposes, even a wide a range as 5 to 35 per cent can be useful. For a start,

it means that the new way of organising care is unlikely to have a negative or

zero effect. In addition, it means that the effect is unlikely to be any larger

than 35 per cent. With that information we could decide whether or not it is

worthwhile making any investment in the new way of organising care. So, in

this case, size does matter

volume 12 number 1 NURSERESEARCHER 31

Determining a sampling frameFor any of this to work, we need to know how common is the condition or

other phenomenon that we wish to study. For this, estimates are needed.

These estimates may come from routinely published data, other research

studies, local audits, or even from clinical judgement (or, as it is known tech-

nically, guesswork).

Two concepts are relevant: incidence and prevalence. They are important

not only in determining a sampling frame for research, but also for health-

care planning and administration more generally. If you are trying to deter-

mine the level and type of staffing needed, the type and amount of equip-

ment to purchase, and ultimately the level of funding which should be

required (even If you do not manage to obtain all that is needed), then it

matters whether you are planning for 10 expected patients, 100, or 1,000.

However, as with all statistical data, one has continually to ask the question:

'What are these numbers going to be useful for?' In the current case, they

are going to be useful not for planning purposes, but to give us an idea of

how large the universe for research is.

Size of the universe: incidenceIncidence is the rate at which new cases occur. It is often represented as the

number of patients per unit of population (100, 1,000, 10,0000) newly

diagnosed with a condition during a given period of time. So, a high inci-

dence will indicate that an extra demand will be placed on healthcare staff.

For example, every new case will require at least a consultation and a diag-

nosis, perhaps with laboratory tests being undertaken. However, whether

that places a serious demand on the system will depend on how long the

condition is likely to persist. This is dealt with under 'prevalence'.

Incidence affects the level of likely demand, but it is most important in the

distribution of demand. A sudden influenza epidemic can place a substantial

demand on primary, secondary and community healthcare resources, for

example. In a sense, incidence tells you how you should arrange your serv-

ice, but on its own it does not tell you how much investment your service

may need. One time period for measuring incidence is the patient's total life

span. It is possible to make an estimate of how many people will develop a

32 NURSERESE&RCHIR volume 12 number 1

given condition at some point in their life. Strangely, we have seen this

referred to as 'lifetime prevalence rate' (Dale et al 1983). So, for example,

we have seen estimates of the proportion of the population who are likely to

suffer from leg ulcers ranging from 1 to 3 per cent (Callam etal 1985). This

sort of figure, given the long time span over which it is measured, is not par-

ticularly useful for service planners. It might, though, be useful to those who

allocate funds for research. If one has to choose between funding two inves-

tigations, one might take into account (among other things) just how many

people are likely to benefit from the results of each investigation. Other

things being equal, one would be tempted to fund the one which aims at the

condition with the higher lifetime incidence.

In terms of developing a sampling frame, one would use incidence strictly

only if you wished to study new (or recurring) cases.

Size of the universe: prevalencePrevalence is the number of people who suffer from a given condition at a

point in time. Some of the people currently suffering may have been recent-

ly diagnosed and so will count in the current period's incidence figures.

Many, however, will not. They will be chronic cases who have suffered from

the condition for many months or years. If they had been counted in any

incidence figures, they may well have made their contribution a decade ago,

and are therefore not counted in the current incidence figures. That is why

the second concept, prevalence, is needed.

By contrast to incidence, the prevalence gives very direct information

about the likely continuing demand that will be placed upon services. It is a

relevant statistic for clinical researchers, as they need to recruit patients meet-

ing certain inclusion and exclusion criteria for their studies.

Conclusion

Too often, one reads research reports in which the authors used a 'conven-

ience' sample. It is rarely, if ever, possible to draw generalisations from such

samples. Even when the sampling has been carried out rigorously, one still

reads of an achieved sample which is, say, only 60 per cent, 50 per cent,

5 per cent or even an unknown fraction of the drawn sample, yet authors

volume 12 number 1 NURSEKESEARCMER 33

nonetheless go ahead and calculate confidence limits, odds ratios, or p val-

ues. This is in principle wrong, and journals should do all that they can to

prevent such material being published. There is a danger that such a counsel

of perfection would rule out of consideration almost all of the research which

is published. We therefore would recommend that journals should consider

the implications of imperfect response rates, and should invite authors to

consider what conclusions (including minimum and maximum estimates)

may reasonably be drawn given the response rates which were achieved in

their particular study,

Laurence G Moseley MA, MBCS, CUP, Professor of Health Services Research,

School of Care Sciences, University of Glamorgan, UK; Donna M Mead RGN,

MSc, PhD, RNT, Professor of Nursing & Head of School, School of Care

Sciences, University of Glamorgan, UK,

Banerjee S ef al (1996) Randomised controlled

trial of effect of intervention by

psychogeriatric team on depression in frail

elderly people at home. British Medical

Journal. 313, 7054, 1058-1061.

Callam MJ eta/(1985) Chronic ulceration of

the leg: Extent of the problem and provision

for care. British MedicalJournal. 290, 6485,

1855-1856.

Dale IJ ef al (1983) Chronic ulcers of the leg: a

study of prevalence in a Scottish community.

Heatth Bulletin (Edinburgh). 4 1 , 6, 310-314.

Kahneman D ef al (1982) Judgement under

Uncertainty: Heuristics and biases. Part IV.

Cambridge, Cambridge University Press.

Mayor S (2000) Lung cancer trial has problems

in recruitment. British Medical Journal. 321,

7255, 195.

Mead D (1993). The Development of

Primary Nursing in NHS Care Giving Institutions

in Wales. Unpublished PhD thesis. University of

Wales, Swansea.

Mead D, Moseley L (2004) Awareness of the

health needs of prisoners. NT Research.

9, 3, 194-207.

Riffenburgh RH (1999) Statistics in Medicine.

New York, Academic Press.

Thomson H ef a/(1999) Randomised controlled

trial of effect of Baby Check on use of health

services in first 6 months of life. British Medical

Journal. 318, 7200, 1740-1744.

UKCC (1999) Nursing in Secure Environments.

London, United Kingdom Central Council for

Nursing, Midwifery and Health Visiting.

For related articles and author guidelines

visit our online archive at

vvvvw.nurseresearcher.co.uk

34 NURSEKESEARCHER volume 12 number 1